Human Factors in AI-Assisted Interpreter Training: A Design Science Approach to Human-Centred Educational Technology

1. Introduction

The increasing integration of AI into education is creating new opportunities to support learning through personalised content, automated feedback, intelligent assistance, and adaptive learning environments. Recent advances in generative AI, large language models (LLMs), automatic speech recognition (ASR), neural machine translation (NMT), and speech synthesis have substantially expanded the range of capabilities that can be incorporated into educational technologies [1,2,3]. However, the educational value of these technologies does not depend solely on their technical capabilities or performance. Their contribution to learning also depends on how they are integrated into human–AI interaction, how they affect learners’ decision-making and autonomy, and whether their use supports rather than interferes with the cognitive and pedagogical processes underlying learning [4,5,6].

This issue is particularly relevant in educational contexts involving cognitively demanding activities and real-time decision-making. Interpreter training represents a paradigmatic example of such a context. Interpreting requires the continuous coordination of listening comprehension, memory, terminology retrieval, linguistic reformulation, speech production, and monitoring under severe time constraints [7]. These demands are especially pronounced in simultaneous interpreting, where multiple processes must be coordinated while the interpreter continuously processes incoming speech. Consequently, AI-based assistance may provide valuable support, but it may also introduce additional attentional demands, uncertainty regarding automatically generated information, or interaction costs that compete with the primary task. The educational suitability of an AI capability therefore cannot be determined solely by whether the technology is technically capable of performing a particular task.

Computer-Assisted Interpreting (CAI) provides an established technological context in which these issues can be examined. Early CAI systems focused mainly on terminology management and the preparation of documentary resources. More recent developments have incorporated automatic transcription, terminology extraction, semantic search, speech recognition, and real-time contextual assistance [8,9,10]. The increasing availability of these capabilities has considerably broadened the potential role of technology in interpreting, extending support from preparation activities to different stages of the interpreting process itself. At the same time, this evolution has changed the nature of the interaction between interpreters and technological systems: whereas traditional CAI tools largely responded to user-initiated queries, AI-based systems can increasingly analyse information automatically, generate suggestions, and provide assistance with limited explicit user intervention.

Most existing research in this area has focused on the performance of individual technologies or on specific functionalities incorporated into CAI systems. Although this work has contributed substantially to technological development, less attention has been devoted to the broader question of how multiple AI capabilities should be integrated into educational environments without compromising the human and pedagogical conditions of learning. In particular, there is a need to understand how such environments can preserve learner agency, support meaningful human control, manage trust in AI-generated outputs, protect cognitive resources, and adapt technological assistance to the characteristics of the activity and the stage of the learning process.

This challenge is closely related to the principles of HCAI, which emphasise the development of AI systems that augment human capabilities rather than replace them [6,11]. From this perspective, AI should support human decision-making while maintaining appropriate levels of human control, transparency, and oversight. In educational settings, however, this perspective requires an additional consideration: the purpose of the system is not merely to improve task performance, but to support the progressive development of the competencies that learners are expected to acquire [2,12]. Consequently, an AI capability that improves immediate task efficiency may not necessarily be beneficial from an educational perspective if it removes opportunities for learners to practise, make decisions, monitor their performance, or develop the underlying competence.

Technology acceptance provides a complementary perspective on this problem. The Technology Acceptance Model (TAM) identifies perceived usefulness and perceived ease of use as central determinants of users’ acceptance of technology [13]. These constructs are relevant when considering the introduction of AI into interpreting, as users’ perceptions of usefulness and usability can influence whether technologies are adopted in professional practice. Nevertheless, the present study does not assume that technology acceptance is equivalent to educational suitability. AI-based systems introduce additional dimensions, including trust in generated outputs, uncertainty, human oversight, contextual appropriateness, and the cognitive cost of interacting with the system. Moreover, technology adoption takes place within professional and organisational contexts in which training, access, established workflows, and institutional conditions may affect actual use. The distinction between perceived usefulness and contextual suitability is therefore particularly important when translating evidence from professional practice into the design of educational environments.

This observation points to a methodological gap at the intersection of AI in education, CAI, and HCAI. Although the literature provides evidence concerning the capabilities of AI, the characteristics of interpreting, and the importance of human-centred interaction, there remains limited guidance on how these different forms of knowledge can be systematically translated into concrete design decisions for AI-assisted interpreter training. In particular, general principles such as human control, transparency, or cognitive protection do not by themselves specify how an educational interface should behave, which functionalities should be provided, when assistance should be activated, or how learners should interact with AI-generated outputs. A systematic connection is therefore required between evidence concerning users and their context, the human factors derived from that evidence, concrete design requirements, and the resulting implementation.

To address this gap, the present study adopts a DSR methodology [14,15]. DSR is particularly appropriate because it treats the development and evaluation of an artefact not merely as an engineering activity, but as a means of generating and communicating design knowledge. Within this approach, empirical and theoretical evidence can be translated into design requirements, implemented in an artefact, and subsequently examined through evaluation. The artefact thus functions as a research instrument through which the relationships between human factors, design decisions, and human–AI interaction can be made explicit.

The study applies this approach to the design of an AI-assisted educational environment for interpreter training. Rather than treating the resulting system as the primary contribution, the artefact is used as a concrete context in which human factors can be operationalised, design decisions can be examined, and principles can be refined through evaluation. The research therefore focuses on the design process and the knowledge generated through it, rather than on the technological novelty of any individual AI functionality.

The study integrates several sources of evidence. First, the literature on AI in education, CAI, HCAI, human–AI interaction, technology acceptance, and interpreting cognition is used to establish the conceptual basis for identifying relevant human and contextual factors. Second, a technology acceptance study involving 42 professional interpreters provides empirical evidence concerning the use, perceived usefulness, ease of use, adoption barriers, and contextual conditions associated with assistive technologies in interpreting practice. The findings of this study are not translated directly into functionalities. Instead, they are interpreted through an intermediate layer of human and contextual factors and subsequently transformed into design requirements. Third, these requirements are incorporated into an AI-assisted educational artefact through an iterative DSR process. Finally, a heuristic evaluation is used to examine the resulting interaction design and to identify aspects requiring further refinement.

The study consequently addresses the following research questions:

• RQ1. What limitations do current AI-based CAI environments present for interpreter training?
• RQ2. What human factors should guide the design of AI-assisted educational environments for interpreter training?
• RQ3. How can these human factors be operationalised as concrete design requirements within an AI-assisted educational environment?
• RQ4. What human-centred design principles can be derived from the development and evaluation of an AI-assisted educational environment for interpreter training?

These questions establish a progression from problem identification to the generation of design knowledge. RQ1 examines the limitations and contextual challenges associated with current AI-based CAI approaches. RQ2 identifies the human and contextual factors that should inform design. RQ3 addresses their translation into explicit design requirements, while RQ4 examines the principles that can be derived from the complete process of design, implementation, and evaluation.

The main contribution of the study is therefore not a particular AI technology or software artefact, but a human-centred design approach that connects empirical evidence, human factors, design requirements, implementation, and evaluation. The study identifies five interrelated design principles: Meaningful Human Oversight, which preserves users’ ability to review, modify, accept, or reject AI-generated outputs; Protection of Cognitive Resources, which seeks to prevent technological interaction from competing with the primary task; Contextualised Assistance, which adapts AI support to the task, workflow stage, and circumstances of use; AI as a Learning Support, which positions technology as a means of developing rather than replacing learner competence; and Transparency and Calibrated Trust, which promotes trust that is calibrated to the actual capabilities and limitations of the AI system. These principles constitute the main conceptual contribution of the study and provide an intermediate structure through which human and contextual factors can be translated into concrete design requirements.

The significance of this contribution extends beyond interpreter training. Although the principles are grounded in the specific characteristics of interpreting and its cognitively demanding learning activities, the underlying design problem is shared by other educational contexts in which AI interacts closely with human decision-making and competence development. The study therefore proposes a shift from asking primarily what AI can do towards asking what role AI should play within a given learning activity, under which conditions, and with what degree of human control. This perspective provides the basis for examining AI-assisted educational environments not simply in terms of technological capability or user acceptance, but in terms of their alignment with human needs, cognitive demands, and educational objectives.

2. Materials and Methods

2.1. From Computer-Assisted Interpreting to Artificial Intelligence-Assisted Educational Environments

CAI encompasses a range of digital tools designed to support interpreters throughout the different stages of the interpreting process. Unlike Computer-Assisted Translation (CAT) tools, which have been developed primarily to support the production of written texts, CAI systems must address the specific demands of oral interpreting, an activity characterised by real-time decision-making, the simultaneous coordination of multiple cognitive processes, and severe time constraints that considerably limit interaction with external resources [8]. These characteristics mean that the design of interpreting tools cannot simply be approached as an adaptation of technologies developed for written translation; rather, it requires careful consideration of interpreters’ specific needs and the context in which interpreting takes place.

The evolution of CAI tools also reflects a significant shift in the role played by technology within the interpreting process. Early systems focused primarily on the creation and management of glossaries, functioning as repositories of terminological information that interpreters could consult when preparing for an assignment. Subsequently, the incorporation of NLP techniques enabled some of these tasks to be automated through terminology extraction, the organisation of documentary resources, and the generation of context-specific materials. More recently, the integration of technologies such as ASR, NMT, and generative AI has led to a new generation of tools capable of providing contextual assistance during the interpreting process itself [8,16].

This evolution represents a qualitative change in the relationship between interpreters and technology. Whereas traditional systems responded primarily to user-initiated queries, more recent tools are capable of automatically analysing speech, anticipating information needs, and generating suggestions during interpreting. Consequently, technology is no longer limited to a passive role but has become an active support agent within the interpreter’s workflow. This shift substantially expands the possibilities for technological assistance, while also increasing the complexity of human–system interaction.

However, the incorporation of new AI-based capabilities does not necessarily lead to an improved user experience or learning process. In interpreting, the usefulness of a given functionality depends not only on its technical accuracy but also on factors such as when information is presented, the cognitive effort required to process it, and the extent to which it interferes with an activity that already requires the simultaneous coordination of multiple mental processes. Similarly, conventional metrics used to evaluate AI technologies, such as ASR accuracy or machine translation quality, are insufficient to determine the actual value of a CAI tool unless interaction-related factors, cognitive load, and the integration of technological assistance into the interpreter’s workflow are also taken into account.

This issue becomes even more relevant in educational settings. Whereas a tool used in professional practice may be evaluated primarily in terms of its ability to increase productivity or reduce preparation time, interpreter training also requires consideration of its impact on learning. Fully automating certain tasks may improve system efficiency, but it may also reduce learners’ opportunities to develop essential competencies through practice, reflection, and decision-making. Consequently, the goal of an AI-assisted educational environment should not be to replace learners’ cognitive activity, but rather to provide forms of assistance that facilitate learning while preserving their active participation throughout the training process.

In this respect, the evolution of CAI has shifted the focus from the development of individual functionalities towards the design of interaction between people and intelligent systems. The challenge is no longer merely to determine which technologies can be incorporated into a tool, but also to establish when these technologies should intervene, what degree of autonomy they should be granted, how information should be presented, and how much control users should retain over AI-supported decisions. This shift in perspective situates the design of CAI tools within a broader problem of human–AI interaction and highlights the need for human-centred approaches to the development of AI-assisted educational environments. This issue provides the starting point for the following section, which examines the HCAI paradigm.

2.2. Human-Centred Artificial Intelligence in Educational Settings

The rapid advancement of AI has traditionally been associated with the development of systems capable of performing tasks with increasing levels of accuracy, autonomy, and efficiency. From this perspective, technological progress has been primarily measured in terms of algorithmic performance, with emphasis placed on aspects such as prediction accuracy, processing speed, and the automation of processes that previously required human intervention. However, the growing adoption of intelligent systems in educational and professional contexts has demonstrated that increased technical capabilities do not, in themselves, guarantee an improved user experience or better outcomes. The effectiveness of these systems also depends on how people understand their capabilities, interact with them, interpret their recommendations, and retain control over decisions supported by AI.

In response to this need, the HCAI paradigm has emerged, aiming to guide the design of intelligent systems towards augmenting human capabilities rather than replacing them [6]. In contrast to traditional approaches, in which higher levels of automation are often associated with reduced user involvement, HCAI argues that automation and human control can be optimised simultaneously through appropriate interaction design. From this perspective, intelligent systems should not be conceived as autonomous agents that replace human activity, but rather as tools capable of enhancing knowledge, supporting decision-making, and assisting with complex tasks without removing users’ responsibility for the final outcome.

This shift in perspective is particularly relevant in education. Whereas the success of a system in other contexts may be evaluated primarily according to its ability to automate processes or increase productivity, the ultimate goal in education is to foster learning and the progressive development of competencies. Accordingly, an AI-based tool should not merely provide answers or automatically perform tasks, but should contribute to learners’ active participation in the construction of knowledge. Automation can reduce the effort associated with certain low-value-added activities, but it should not replace the cognitive processes that underpin meaningful learning.

In recent years, this perspective has gained increasing relevance in research on AI in education. Several studies have highlighted that the effective integration of intelligent systems requires consideration of factors such as user agency, control over the interaction, trust in generated recommendations, system transparency, and the involvement of students and teachers in design processes. From this perspective, the quality of an AI-based educational tool depends not only on the accuracy of the models it employs, but also on its ability to be coherently integrated into teaching and learning processes while respecting the needs, capabilities, and objectives of its users.

One of the main contributions of the HCAI paradigm is its recognition that AI-based systems exhibit characteristics that distinguish them from conventional software. Unlike traditional systems, whose behaviour is typically defined through explicit and relatively predictable rules, machine learning models generate probabilistic outputs that may vary depending on the available data and involve different degrees of uncertainty. This characteristic requires a reconsideration of human–AI interaction design, including mechanisms that enable users to understand the system’s capabilities and limitations, monitor its outputs, and intervene when these outputs are inappropriate. In this regard, the guidelines proposed by Amershi et al. [11] provide a fundamental reference for the design of AI-based interfaces, offering recommendations aimed at facilitating communication of system capabilities, error management, adaptation to users, and the preservation of meaningful human control throughout the interaction.

Overall, HCAI conceptualises AI in terms of collaboration between people and intelligent systems, whereby automation is understood as a means of augmenting human capabilities rather than an end in itself. This perspective shifts attention from the development of algorithms towards the design of interaction experiences that support the effective, safe, and responsible use of AI. The central challenge is therefore no longer to determine how much a system can automate, but rather to establish how responsibility should be distributed between humans and AI in ways that enhance learning and decision-making.

However, translating this conceptual framework into the design of educational environments requires a more precise understanding of the interaction dimensions that shape the acceptance and effective use of these technologies. Factors such as user agency, control, trust, transparency, cognitive load, and the integration of AI assistance into the user’s workflow mediate the relationship between a system’s technical capabilities and its impact on learning. Analysing these factors makes it possible to translate the general principles of HCAI into concrete design criteria for the development of AI-assisted educational environments, an issue addressed in the following section.

2.3. Human Factors in AI-Assisted Educational Environments: a Conceptual Model

The HCAI paradigm establishes that the success of an intelligent system depends not only on its technical capabilities but also on how people interact with it. However, translating this principle into the design of educational environments requires identifying which dimensions of interaction actually shape the adoption, acceptance, and impact of these technologies on learning. The literature indicates that these dimensions extend beyond algorithmic performance and encompass factors related to user behaviour, the characteristics of the interaction, and the effects of AI-based assistance on the learning process [6,12].

To integrate these contributions, the present study organises the main human factors identified in the literature into three complementary dimensions: user-related factors, interaction-related factors, and learning-related factors. This organisation provides a conceptual framework for linking AI capabilities to concrete design criteria and, ultimately, for guiding the development of AI-assisted educational environments.

2.3.1. User-Related Factors

The first dimension encompasses factors that determine the role played by individuals when interacting with an intelligent system. Among these, agency, human control, and trust are particularly relevant. These three closely related concepts shape how users engage with and act upon AI-generated recommendations.

From an HCAI perspective, user agency refers to users’ ability to understand the role and purpose of the system, decide when to use its recommendations, and retain responsibility for the decisions they make. This principle is particularly relevant in educational environments, where the objective is not merely to obtain a correct outcome, but also to foster competency development through learners’ active participation in the learning process [6].

Agency is closely connected to the concept of meaningful human control. Automating a task does not necessarily imply eliminating user involvement. On the contrary, intelligent systems should provide mechanisms that enable users to monitor, modify, or reject system outputs when necessary. This balanced distribution of responsibility represents a fundamental principle of HCAI and is particularly important in systems based on probabilistic models, whose outputs may involve different degrees of uncertainty.

In this context, trust represents another essential factor in human–AI interaction. The literature distinguishes between insufficient trust, which may lead to technology rejection, and excessive trust, which can result in overreliance or automation bias. Accordingly, the goal of system design should not be to maximise user trust, but rather to foster calibrated trust, whereby users’ level of trust is appropriately aligned with the system’s actual capabilities and limitations. Achieving this balance is particularly important in educational environments, where learners may not yet possess the knowledge and expertise required to critically assess AI-generated outputs.

2.3.2. Interaction-Related Factors

The second dimension encompasses factors that determine the quality of interaction between individuals and intelligent systems. Among these, usability, transparency, and ease of use are particularly relevant, as they influence both users’ understanding of the system and the extent to which the technology can be effectively integrated into the tasks it is intended to support.

Usability is one of the most established concepts in human–computer interaction and is defined as the extent to which a system can be used by specified users to achieve specified goals with effectiveness, efficiency, and satisfaction in a specified context of use [17]. In AI-assisted educational systems, high usability not only facilitates the use of the tool but also reduces the effort required to interact with it, allowing users’ cognitive resources to remain focused on the learning activity.

Ease of use has traditionally been examined through the TAM proposed by Davis [13]. This model posits that the intention to use a technology is primarily influenced by its perceived usefulness and the perceived ease with which users can employ it. Over the past decades, TAM has become one of the most widely adopted frameworks for explaining the acceptance and adoption of educational technologies and digital systems.

However, recent research suggests that technology acceptance models may be insufficient when interaction involves AI. Factors such as transparency, explainability, the degree of user control, and the uncertainty inherent in probabilistic models play an important role that is not explicitly represented in the original model [4,6]. Accordingly, the acceptance of an AI-based system should be understood as the result of an interaction between the classical factors described by TAM and additional dimensions specific to human–AI interaction.

Transparency represents one of these distinctive factors. Rather than requiring users to understand the internal functioning of algorithms, transparency aims to enable them to understand what functions the AI performs, what its limitations are, and in which situations its outputs should be reviewed or validated. This understanding supports more critical use of the technology and contributes to the development of appropriately calibrated trust.

2.3.3. Learning-Related Factors

The third dimension encompasses factors that determine the impact of AI on the learning process. In this context, the value of a tool depends not only on the functionalities it incorporates, but also on its ability to support competency development without unnecessarily increasing users’ cognitive demands.

Perceived usefulness is one of the main determinants of technology acceptance and refers to the extent to which a person believes that using a particular system will improve their performance on a given task [13]. However, in educational contexts, usefulness should be interpreted from a broader perspective, considering not only the efficiency with which an activity can be performed but also the extent to which the technology contributes to learning and the development of transferable competencies.

Alongside perceived usefulness, cognitive load represents one of the most relevant factors in the design of AI-assisted educational systems. Any interaction with a digital tool requires attentional resources that compete with those needed to perform the primary task. Consequently, the incorporation of new functionalities is beneficial only when the value of the information provided outweighs the cognitive effort required to process it. This issue is particularly important in activities characterised by high cognitive demands, such as interpreting, where attention must be distributed simultaneously across multiple processes involving comprehension, memory, and speech production [7].

Finally, the integration of AI into the learning workflow is a key element in ensuring its effectiveness. Intelligent assistance should be provided in a contextual, timely, and selective manner, avoiding unnecessary interruptions to learners’ activities or replacing cognitive processes that are essential for competency development. From this perspective, the objective of design is not to maximise the degree of automation, but rather to achieve an appropriate balance between the support provided by technology and users’ active participation in the learning process.

Taken together, these three dimensions demonstrate that the value of an AI-assisted educational environment does not depend exclusively on the performance achieved by its underlying algorithms, but rather on the interaction between the system’s technical capabilities and the human factors that shape its use. This conceptual framework moves beyond a perspective focused solely on technology acceptance and provides a theoretical basis for translating the human factors identified in the literature into design principles applicable to the development of AI-based educational tools.

Building on this conceptual organisation, the following section presents the design principles derived from these factors, which provide the foundation for the development of the artefact proposed in this research.

2.4. Human-Centred Design Principles for AI-Assisted Educational Environments

The human and contextual factors identified in the preceding sections provide the basis for translating the general principles of Human-Centred Artificial Intelligence (HCAI) into concrete design guidance for AI-assisted educational environments. Rather than treating these factors as isolated considerations, the present study organises them into a set of design principles that establish an intermediate level between theoretical frameworks, empirical evidence, and implementation decisions. This perspective is consistent both with the HCAI approach, which places human agency and control at the centre of intelligent system design [6], and with established guidelines for human–AI interaction [11].

Five design principles were derived from the literature reviewed and subsequently refined through the empirical findings obtained in this study: (1) meaningful human oversight, (2) transparency and calibrated trust, (3) protection of cognitive resources, (4) contextualised assistance, and (5) AI as a learning support. These principles are not intended as universal prescriptions, but as context-sensitive guidelines for designing AI-assisted educational environments in which technological capabilities remain aligned with human agency, cognitive demands, and learning objectives. This formulation responds to the need, identified in the literature on AI in education, to move from a perspective focused exclusively on technological adoption towards one that considers the meaningful integration of AI into teaching and learning processes [2,3,4].

2.4.1. Principle 1: Meaningful Human Oversight

AI-assisted educational environments should provide assistance without removing users’ ability to understand, evaluate, and control the system’s outputs. Meaningful human oversight requires users to retain responsibility for decisions and to be able to review, modify, accept, or reject AI-generated information when appropriate. This approach is consistent with the HCAI conception proposed by Shneiderman [6] and with guidelines for human–AI interaction, which emphasise the importance of maintaining human control, making system capabilities understandable, and facilitating error management and user intervention [11].

This principle is particularly relevant in interpreting, where AI systems may generate probabilistic outputs whose accuracy cannot always be guaranteed. AI models may produce plausible but incorrect outputs, meaning that users need to be able to monitor system outputs and determine when it is appropriate to accept them [5,11]. Rather than replacing human judgement, AI should therefore support users while preserving their ability to intervene in the process. The appropriate distribution of responsibility between the human and the system should depend on the task, the level of automation, and the potential consequences of erroneous outputs.

2.4.2. Principle 2: Transparency and Calibrated Trust

AI systems should make their actions, outputs, and limitations sufficiently understandable to support an appropriate level of user trust. The objective is not to maximise trust, but to promote calibrated trust, whereby users’ confidence in the system is aligned with its actual capabilities and limitations [5].

Transparency is particularly important when AI-generated outputs are uncertain or when users may lack sufficient knowledge to assess them critically. Research on human–AI interaction highlights the need to communicate system capabilities and limitations appropriately and to provide mechanisms that enable users to interpret and monitor system outputs [11]. Similarly, recent research on trust in AI systems highlights the risks associated with both excessive trust, which may lead to inappropriate reliance on automated outputs, and insufficient trust, which may prevent users from taking advantage of available capabilities [5]. Educational environments should therefore facilitate a critical relationship with AI by making relevant system states and outputs visible and by supporting users in understanding when and how AI assistance should be relied upon.

2.4.3. Principle 3: Protection of Cognitive Resources

AI assistance should reduce rather than increase the cognitive demands associated with the primary learning activity. This principle is particularly important in interpreter training because learners must allocate limited cognitive resources to processes such as comprehension, analysis, reformulation, monitoring, and production. Gile’s Effort Models conceptualise interpreting as an activity in which different processes compete for limited cognitive resources [7]. Research on computer-assisted interpreting technologies has likewise highlighted the need to consider the temporal and cognitive constraints specific to interpreting when introducing new technologies into the activity [8].

Accordingly, AI-based functionalities should avoid unnecessary interactions, excessive information density, and the simultaneous presentation of information that is not relevant to the task being performed. Assistance should be designed so that technological interaction does not compete with the cognitive processes that learners are expected to develop. The timing and amount of information presented should also be compatible with the demands of the activity. From this perspective, more information or a higher degree of automation does not necessarily imply a better educational solution; the value of a functionality should be considered in relation to the benefit it provides, the cognitive cost of accessing it, and the objective of the activity [7,8].

2.4.4. Principle 4: Contextualised Assistance

The usefulness of AI assistance depends on its appropriateness to the specific task and context in which it is provided. AI capabilities should therefore be linked to particular activities, stages of the workflow, and user needs rather than being presented as permanently available and context-independent functionalities. This approach is consistent with the literature on educational technologies, which emphasises the need to integrate AI capabilities with learning processes and with the specific needs of students and educators [2,3,4].

In interpreting, this contextualisation is particularly important because different stages of the interpreting process present different requirements. Research on computer-assisted interpreting shows an evolution from tools primarily focused on preparation and terminology management towards systems capable of providing assistance during task performance [8,9]. However, the temporal and cognitive characteristics of interpreting mean that the usefulness of a given technological capability may vary depending on when it is used [8].

This principle therefore implies that the same AI capability may be appropriate at one stage of the interpreting process but inappropriate or distracting at another. Educational environments should consequently adapt the availability, timing, and form of assistance according to the characteristics of each learning activity. Contextualisation also requires consideration of existing professional practices and workflows so that new AI capabilities can be integrated into structures that are recognisable to users [2,8,9].

2.4.5. Principle 5: AI as a Learning Support

AI should contribute not only to the completion of tasks but also to the development of the competencies that learners need to acquire. In educational environments, automation should therefore be distinguished from the automation of learning itself. The literature on AI in education has highlighted that the educational value of these technologies depends on their integration into learning processes and their ability to support competency development rather than merely providing answers or automating activities [1,2,3,4].

While AI can facilitate access to learning materials, provide feedback, support preparation, or assist with repetitive activities, it should not unnecessarily replace the cognitive effort through which learners develop the target competence. In this regard, automation that maximises task efficiency may be pedagogically counterproductive if it eliminates precisely the cognitive activities that learners need to perform to develop the relevant competence.

This principle also implies that learners should have opportunities to develop the technological literacy required to use AI critically and effectively. The educational integration of these technologies therefore requires learners to understand not only how to use AI tools but also their capabilities, limitations, and potential implications for learning [1,2,3,4]. The environment should consequently support progressive familiarisation with AI capabilities and encourage users to understand their appropriate use, limitations, and potential effects on the learning process.

2.4.6. Adapting the Principles to the Learning Cycle

The five principles should not be interpreted as independent rules that apply uniformly across the entire educational environment. Their implementation should be adapted to the characteristics of each stage of the interpreting learning cycle and to the specific cognitive and interactional demands of the task. This need for adaptation is consistent with the literature on AI in education, which highlights the importance of integrating intelligent technologies with the characteristics of educational activities and the needs of their participants [2,3,4].

For example, preparation activities may benefit from richer forms of AI assistance, such as terminology management, content generation, or training materials, whereas activities requiring intensive real-time processing may require more selective and less intrusive assistance [8,9]. Similarly, the degree of human oversight, the amount of information presented, and the timing of AI interventions may vary according to the learner’s experience and the objectives of the activity.

The principles therefore provide a common design framework while allowing their concrete implementation to vary across functionalities and learning contexts. This contextual adaptation is essential to avoid treating AI assistance as a uniform technological capability and instead consider how different forms of automation affect learner agency, cognitive resources, trust, and opportunities for competence development [5,6,7,8,9]. In this sense, the five principles operate as an intermediate level between general HCAI frameworks and concrete implementation decisions, enabling the human and contextual factors identified in this study to be translated into specific design requirements for AI-assisted educational environments.

2.5. Research Design and Methodological Rationale

The study followed a sequential and iterative methodological design combining an empirical study of technology use and acceptance in interpreting, a development process grounded in DSR, and a heuristic evaluation of the resulting artefact. This combination responds to the nature of the research problem, which extends beyond determining whether professionals perceive specific technologies as useful. Rather, the study seeks to understand which human and contextual factors shape the adoption of AI technologies and how these factors can be translated into concrete design decisions for an AI-assisted educational environment.

From this perspective, technology acceptance is not treated as an outcome independent of the design process. Instead, the findings obtained from the initial study with professional interpreters were used as evidence to identify needs, constraints, and human factors relevant to design. These findings were subsequently translated into design requirements and guiding principles, which were operationalised during the iterative development of the artefact. Finally, a heuristic evaluation was conducted to identify interaction problems and assess the adequacy of the resulting design decisions, providing additional evidence for artefact refinement.

The overall research process can therefore be represented as the following transformation: technology-use and acceptance study, human factors, design requirements, design principles, iterative artefact development, heuristic evaluation, design refinement.

This structure distinguishes three complementary methodological functions. First, the professional study provides evidence concerning the context of use and the conditions affecting technology adoption. Second, DSR provides the framework through which this evidence is transformed into design knowledge and embodied in an artefact. Third, heuristic evaluation enables the interaction with the artefact to be examined and the identified problems to be used to refine the design.

Accordingly, the study does not seek to demonstrate that the resulting artefact improves interpreting performance or learning outcomes. Rather, its objective is more specific: to investigate how human factors associated with AI adoption can be systematically incorporated into the design of an AI-assisted educational environment and how these design decisions can subsequently be evaluated and refined during artefact development.

2.6. Study 1: Professional Technology Use and Acceptance

2.6.1. Aim and Participants

The first empirical phase aimed to characterise the professional use of different technologies applied to interpreting and to identify factors that may facilitate or hinder their adoption. To this end, a questionnaire based on the TAM was administered to professional interpreters.

The study involved 42 professional interpreters from diverse professional, linguistic, and geographical contexts. The sample included participants with different levels of professional experience and different interpreting modalities, including face-to-face and remote simultaneous interpreting, consecutive interpreting, bilateral interpreting, and, to a lesser extent, telephone interpreting. This heterogeneity enabled the study to capture perceptions arising from different professional contexts.

The questionnaire also collected information on participants’ professional status and the technologies used in their professional activities. In particular, the study examined automatic speech recognition and transcription tools, AI-based content-generation systems, terminology-management tools, and AI ecosystems or systems designed to support interpreters.

The purpose of this phase was not to evaluate the artefact subsequently developed, but rather to obtain contextual evidence concerning professional practices, technology acceptance, and adoption barriers that could inform the design process.

2.6.2. TAM-Based Instrument

The questionnaire was designed with reference to the Technology Acceptance Model, focusing on three dimensions: Perceived Usefulness (PU), Perceived Ease of Use (PEOU), and Behavioural Intention (BI).

In addition to these dimensions, the questionnaire collected information about the frequency with which different categories of technologies were used and elicited participants’ views regarding factors facilitating or hindering their adoption.

This combination was relevant to the objectives of the study. Whereas the TAM dimensions provide a framework for characterising users’ perceptions of technologies, information concerning actual use and participants’ responses provides a contextual account of these perceptions within the specific conditions of professional interpreting. Thus, the study did not merely examine whether a technology was perceived as useful, but also explored why a potentially useful technology might nevertheless have limited adoption in professional practice.

2.6.3. Analysis of Technology Use and Acceptance

The questionnaire results revealed substantial differences across technology categories. Terminology-management tools showed the highest levels of use and appeared to be routinely integrated into professional workflows. By contrast, automatic speech recognition and transcription tools were used considerably less frequently and were primarily associated with preparation activities.

AI-based tools for generating content from glossaries occupied an intermediate position. Participants mentioned, among others, general-purpose tools such as ChatGPT, Copilot, Gemini, and NotebookLM. However, AI ecosystems and support systems specifically intended for use in the interpreting booth showed comparatively low levels of adoption.

This distribution provides an initial indication that the availability of a technological capability does not necessarily imply its integration into professional practice. Familiarity with a technology, its compatibility with established tasks, and its incorporation into existing workflows may play an important role alongside perceived usefulness.

The TAM-oriented analysis also indicated generally positive perceptions of the potential usefulness of these technologies. However, this positive perception did not necessarily correspond to high levels of actual use. For emerging AI-based technologies, participants expressed greater uncertainty concerning ease of use, reliability, workflow integration, and compatibility with the temporal constraints of interpreting.

These findings are particularly relevant to simultaneous interpreting. Participants indicated that the activity requires high levels of concentration and that interacting with technological tools during the task may introduce additional cognitive demands. Similarly, the need to verify or correct automatically generated outputs may reduce some of the expected benefits of automation.

Finally, participants identified several barriers to technology adoption, including limited access to specific training, organisational policies, confidentiality concerns, uncertainty regarding reliability, and the perception that some technologies could replace rather than support professional interpreting. At the same time, participants generally expressed a positive attitude towards training in these technologies and towards their use as supporting tools.

2.7. From Empirical Evidence to Design Requirements

A central methodological feature of the study was to avoid incorporating the findings from the technology-acceptance study into the artefact merely as general conclusions. Instead, the findings were subjected to a process of translating empirical evidence into design requirements.

This process was organised at three levels.

At the first level, empirical findings concerning technology use, acceptance, and perceived adoption barriers were identified. At the second level, these findings were interpreted in terms of human and contextual factors relevant to interaction with AI-based systems. At the third level, these factors were translated into design requirements, which were subsequently related to the design principles guiding the artefact.

For example, concerns regarding the reliability of AI-generated outputs were interpreted as indicating the need to preserve users’ ability to inspect and modify generated information. Similarly, references to cognitive load and the attentional demands of simultaneous interpreting were translated into a requirement to minimise unnecessary interaction and prevent technological assistance from competing with the user’s primary task.

This process established an explicit relationship between the empirical findings and the decisions made during artefact development.

Table 1 constitutes a central methodological element of the study because it makes explicit the relationship between empirical evidence and design decisions. The analysis does not assume that each empirical finding directly determines a single artefact feature. Rather, findings are interpreted as evidence informing design requirements, which are subsequently translated into design principles and implementation decisions.

2.8. Design Science Research Process

Once the design requirements had been identified, artefact development followed a Design Science Research approach. The selection of this paradigm reflects the fact that the central objective of the research was not only to describe an existing phenomenon, but also to construct and evaluate an artefact addressing an identified problem within a real-world context.

The DSR process was conducted iteratively and involved complementary profiles from research, software development, and interpreting. Before the formal evaluation, team meetings, review sessions, and internal testing activities were conducted to identify problems and inform subsequent development.

The development process comprised three design iterations. These iterations were not treated merely as independent software-development stages, but as successive cycles of construction, review, and refinement of design knowledge. During each cycle, design decisions were revisited and feedback from researchers, developers, and interpreters was incorporated into the subsequent version of the artefact.

Importantly, the formal evaluation was not conducted on an initial, partially integrated version of the system. Prior to the formal evaluation, development and review activities were undertaken until a functional version of the artefact was available, with the main functionalities integrated into the system.

This distinction makes it possible to separate the iterative development process from the subsequent formal evaluation. The former provided formative knowledge for design, whereas the heuristic evaluation constituted a specific phase for systematically examining the resulting artefact.

2.9. Iterative Artefact Development

The artefact developed in the study was an AI-assisted educational environment designed to support different stages of interpreter preparation and training. Its design incorporated a set of integrated functionalities, including assignment management, glossary generation and management, note-taking, transcription, video-based training, memorisation, and text-to-speech-based training.

The modular architecture enabled these functionalities to be distributed according to the different needs arising throughout the interpreting preparation and training cycle. This decision was consistent with the empirical findings indicating that technologies are not used uniformly across all stages of professional activity.

The design strategy therefore did not seek to maximise the number of automated functions. Instead, it prioritised AI capabilities that could be incorporated into recognisable user activities while preserving users’ ability to supervise and control the resulting outputs.

This strategy enabled the artefact to evolve not only in terms of technological implementation but also in terms of its alignment with the human-factor requirements identified during the first phase of the research.

2.10. Study 2: Heuristic Evaluation of the Artefact

The second empirical phase aimed to identify interaction and usability problems in the integrated version of the artefact. A heuristic evaluation was therefore conducted with four evaluators representing complementary profiles: two interpreters, one researcher, and one software developer.

The inclusion of these profiles was intended to provide complementary perspectives on the artefact. The interpreters contributed domain knowledge and understanding of the professional context; the researcher provided a methodological perspective and a focus on the research objectives; and the developer contributed technical knowledge concerning interaction and system behaviour.

The evaluation was structured around a set of usability heuristics and employed a 0–4 severity scale, allowing identified problems to be classified according to their relative importance.

The evaluation produced 34 findings distributed across the different modules of the artefact. The findings were subsequently analysed according to both the affected functionality and the heuristic involved, as well as their severity.

The identified problems included issues related to error prevention, recognition rather than recall, user control and freedom, and consistency. Their distribution provided evidence concerning which components required further attention during the refinement process.

The evaluation was not used to claim that the artefact was "usable" in absolute terms or to demonstrate educational effectiveness. Its purpose was more specific: to provide evidence of concrete interaction problems that could be used to reassess and refine the design decisions adopted during the DSR process.

2.11. Evidence Integration and Design Traceability

The integration of the two empirical phases constitutes a central element of the methodological design. The TAM-based study and the heuristic evaluation addressed different questions and served different functions.

The professional study provided evidence concerning the professional context, technology-use patterns, perceived usefulness and ease of use, and factors potentially constraining the adoption of AI technologies. Its findings were primarily used to establish design requirements.

The heuristic evaluation, by contrast, examined the outcome of those design decisions in the artefact and identified specific interaction problems. Its findings were used to guide subsequent refinement.

The relationship between the two phases can therefore be expressed as follows: Contextual evidence, human factors, design requirements, design principles, artefact implementation, heuristic findings, design refinement.

This traceability mechanism makes it possible to distinguish between three forms of knowledge generated throughout the research.

First, the study generated contextual knowledge concerning the conditions under which AI technologies are used and accepted in professional interpreting. Second, it generated design knowledge, expressed through requirements and principles for incorporating AI into educational environments. Third, it generated evaluation knowledge, derived from the interaction problems identified during heuristic evaluation.

From this perspective, artefact evaluation is not treated as an isolated final stage but as a mechanism for testing and refining design decisions. The combination of the two empirical sources therefore provides a basis for examining the consistency of the chain connecting human factors with technological implementation.

3. Results

3.1. Characterisation of the Participants and the Technological Context

The technological acceptance study involved 42 professional interpreters from diverse professional and geographical backgrounds. The sample was heterogeneous in terms of professional experience, employment status, and interpreting modalities. Participants represented different stages of their professional careers, ranging from interpreters with limited experience to professionals with more than ten years of practice. Likewise, different employment contexts were represented, including freelance work, companies, agencies, and institutional settings.

Regarding interpreting modalities, participants reported experience in face-to-face and remote simultaneous interpreting, consecutive interpreting, bilateral interpreting, and telephone interpreting. This diversity is relevant because the conditions under which technologies are used may vary depending on the temporal and cognitive characteristics of each interpreting modality.

It should be noted that the categories corresponding to professional status were not mutually exclusive. Participants were allowed to select more than one category; consequently, the frequencies associated with this variable amounted to 48 responses. This figure does not represent the sample size, which remained at 42 participants.

This diversity provides an appropriate context for examining the relationship between technology and professional practice without assuming that interpreters’ needs are homogeneous. In particular, it allows technology adoption to be considered in relation to different interpreting modalities, levels of professional experience, and work contexts.

3.2. Patterns of Technology Adoption

The results reveal substantial differences in the degree of adoption across the technological categories examined. Terminology management tools were the technologies most frequently used by participants. These tools were reported to be integrated into professional workflows, particularly during assignment preparation and terminology consultation. Participants mentioned both specialised applications and general-purpose tools, including spreadsheets and word processors.

In contrast, ASR and transcription tools showed considerably lower levels of use. Most participants reported using these technologies rarely or never in their regular professional practice. When used, their application was primarily concentrated on assignment preparation, for example, to become familiar with specialised terminology, identify relevant concepts, or conduct a preliminary analysis of task-related content.

AI-based tools for generating content from glossaries occupied an intermediate position. Their use was higher than that observed for ASR and transcription tools, although it cannot yet be considered intensive or systematic. As noted above, the most frequently used tools included ChatGPT, Copilot, Gemini, and NotebookLM, which participants primarily employed as complementary resources within their professional practice.

Finally, AI ecosystems and booth-assistance systems exhibited the lowest levels of adoption. Most participants reported using these technologies rarely or never, suggesting that their integration into professional practice remains at an early stage.

Overall, these findings reveal a clear distinction between established technologies and emerging forms of AI-based assistance. The former, particularly those related to terminology management, show greater integration into professional workflows, whereas more recent AI technologies exhibit lower levels of use.

This pattern constitutes a relevant initial finding for the design of the artefact: the availability of a technological functionality does not, in itself, guarantee professional adoption. Effective integration also appears to depend on users’ familiarity with the technology, the context in which it is used, and its compatibility with existing professional practices.

3.3. Perceived Usefulness

Perceived usefulness was assessed using four items designed to determine the extent to which participants considered that each technological category could improve their performance, facilitate the completion of their work, increase their productivity, and be useful for their professional activities. These four indicators were applied to the four technological categories considered in the study: speech-to-text tools, AI-based tools for generating speeches from glossaries, terminology management tools, and AI ecosystems and booth-assistance systems.

The results reveal a differentiated pattern across the technologies analysed. Terminology management tools received the most consistently favourable ratings across all four usefulness indicators. Responses were strongly concentrated in the agreement categories, with negative responses being virtually negligible. This pattern was particularly evident for the items concerning improved performance, facilitation of work, and overall usefulness.

AI-based tools for generating speeches from glossaries also received predominantly positive evaluations. Across the four items, a substantial concentration of favourable responses was observed, although the distribution was somewhat more heterogeneous than for terminology management tools. In particular, neutral responses occurred more frequently and, to a lesser extent, so did responses indicating disagreement.

Speech-to-text tools likewise showed a generally positive perception of usefulness, particularly with regard to overall usefulness and their ability to facilitate work. However, their distributions included a higher proportion of neutral and negative responses than those observed for terminology management tools. This finding suggests a favourable perception of the potential of these tools, although this perception is less established than that associated with technologies already integrated into professional workflows.

The pattern was more markedly different for AI ecosystems and booth-assistance systems. In this category, responses showed considerably greater heterogeneity, with a relevant proportion of neutral and disagreement responses. This heterogeneity was observed across all four perceived usefulness indicators and was particularly evident in participants’ perceptions of the potential of these tools to increase productivity and improve professional performance.

Overall, the results indicate a gradient in perceived usefulness acceptance. Technologies that are more established within professional practices show a more homogeneous and favourable evaluation, whereas AI technologies intended to provide more direct assistance during interpreting activities elicit a more ambivalent assessment.

This finding is relevant because it indicates that perceived usefulness does not depend exclusively on a tool’s technical capabilities. A technology may offer potentially useful functionalities while nevertheless receiving a less favourable evaluation when its practical application is less familiar or presents greater challenges for integration into professional activities.

Therefore, the PU results suggest that the value of an assistive technology should be understood in relation to the specific task performed by the user, the point in the workflow at which assistance is provided, and the manner in which such assistance is integrated into existing professional practices.

3.4. Perception Ease of Use

Perceived ease of use was assessed using four indicators addressing participants’ enjoyment of using the tool, comfort during use, ease of becoming proficient in its operation, and the perception that the tool is manageable and controllable.

The results again reveal differences across the technological categories.

Terminology management tools received predominantly favourable evaluations of their ease of use. Although some negative responses were reported, the distributions were mainly concentrated in the agreement categories, particularly for the indicators related to ease of handling and perceived controllability.

AI-based tools for generating speeches from glossaries also showed a predominantly positive perception. Responses were considerably concentrated in the agreement categories across all four indicators, although a higher proportion of neutral responses was observed than for terminology management tools.

For speech-to-text tools, the overall evaluation remained favourable, but greater heterogeneity was observed. Ease of becoming proficient and perceived control showed less clearly positive distributions than those observed for terminology management tools. Responses indicating disagreement were also present across the different indicators.

The most pronounced difference once again emerged for AI ecosystems and booth-assistance systems. In this category, responses were much more evenly distributed across disagreement, neutral, and agreement positions. Neutral responses were particularly prevalent, while agreement responses accounted for a smaller proportion than in the other categories.

This finding is particularly significant when considered alongside perceived usefulness. Participants do not appear to evaluate emerging technologies solely in terms of what they can do, but also in terms of how easily they can be incorporated into their professional activities and the degree of control users retain during interaction.

The difference between established and emerging technologies can therefore be interpreted as an indication of the importance of familiarity and prior experience. Tools that professionals already know and use regularly receive more favourable ease-of-use evaluations, whereas tools that introduce new forms of assistance generate greater uncertainty.

This finding has a direct implication for the design of the educational environment: the incorporation of AI capabilities should not assume that users already possess the knowledge required to interact with them efficiently. The environment itself can serve as a means of familiarisation and progressive learning.

3.5. Behavioural Intention

Future intention to use was assessed using two indicators: the likelihood of continuing to use the tool in the future and the desire to be able to continue using it.

The results show a particularly favourable intention to use in the case of terminology management tools. Responses were clearly concentrated in the agreement categories, with very limited neutral and negative responses. This pattern is consistent with the high perceived usefulness and ease of use observed for this category.

AI-based tools for generating speeches from glossaries also showed a predominantly positive future intention to use. Most responses fell within the agreement categories, although a small proportion of neutral and negative responses was observed.

For speech-to-text tools, future intention to use was likewise predominantly favourable. However, the distribution was somewhat more heterogeneous than that observed for terminology management tools, particularly for the item concerning the likelihood of continuing to use this type of tool.

By contrast, AI ecosystems and booth-assistance systems showed considerably greater heterogeneity in future intention to use. Although a relevant proportion of participants expressed an interest in continuing to use these technologies, neutral and negative responses were also observed for both indicators.

The comparison of PU, PEOU, and BI reveals a relevant descriptive pattern: technologies receiving more favourable evaluations in terms of usefulness and ease of use are also those showing more homogeneous future intentions to use. Conversely, as neutral or negative responses increase for PU and PEOU, future intention to use also becomes more dispersed.

3.6. Barriers and Training Needs

The findings regarding adoption barriers complement the information provided by PU, PEOU, and BI. While these three constructs describe participants’ perceptions of the technologies, this part of the questionnaire helps identify some of the reasons why professionals may not incorporate them into their regular practice.

The most frequently reported barrier was a lack of learning or knowledge about how to use these tools, selected on nine occasions. This was followed by a perceived lack of need for the tool, with eight responses, and the perception that the organisations for which participants work do not encourage their use, also with eight responses. Lack of comfort with the tool was selected twice, and two additional responses corresponded to other reasons.

These findings are particularly relevant because they show that non-adoption cannot automatically be interpreted as technology rejection. In fact, the most frequently reported barrier was related to a lack of knowledge or training rather than to a negative evaluation of the technology itself.

The training dimension becomes even more relevant because the questionnaire specifically addressed participants’ interest in receiving training, gaining access to these tools, and learning how to use them.

This finding establishes a direct connection between the technology acceptance study and the educational purpose of the artefact. If some of the identified barriers are related to a lack of knowledge and learning opportunities, an AI-assisted educational environment can serve a dual function: providing technological assistance while simultaneously developing the competencies required to use such assistance critically and effectively.

The identification of the organisational context as a barrier is also significant. Technologies are not adopted solely on the basis of individual user preferences but within professional environments subject to policies, procedures, and constraints. Consequently, even users who perceive a technology positively may not have the necessary conditions to use it.

These findings allow a distinction to be established between individual acceptance and contextual adoption. The former is related to users’ perceptions of the technology, whereas the latter also depends on external factors such as training, access, organisational policies, and compatibility with existing professional practices.

3.7. From TAM Findings to Human-Factor Requirements

The preceding findings were used as the starting point for identifying the human factors that should guide the design of the educational environment. The aim was not to mechanically translate each questionnaire response into a specific functionality, but rather to introduce an intermediate interpretive step that would enable the empirical findings to be translated into design requirements.

This transformation was structured across four levels: empirical finding, human factor, design requirement, and design principle.

The first stage captures the pattern observed in the study. The second interprets this pattern from the perspective of the interaction between the user, the technology, and the context. The third translates this interpretation into a condition that the artefact should fulfil. Finally, the fourth expresses this condition as a general principle applicable across the different functionalities of the environment.

Table 2 makes explicit one of the main methodological contributions of the study. The TAM results are not used solely to describe the level of acceptance of different technologies, but rather to identify human and contextual conditions that are subsequently translated into design criteria.

Based on this transformation, five principles were established:

• Meaningful Human Oversight. AI capabilities should provide assistance without removing the user’s ability to evaluate their outputs. Users should be able to review, modify, accept, or discard information generated by the system.
• Transparency and Calibrated Trust. Interaction with AI systems should foster calibrated trust: sufficient to take advantage of their capabilities, but not based on the automatic acceptance of their outputs. Users should have access to mechanisms that facilitate the verification and correction of generated information.
• Protection of Cognitive Resources. Technological assistance should reduce rather than increase the cognitive demands of the task. This entails avoiding overly dense interfaces, unnecessary interactions, and the simultaneous presentation of information that is not relevant to the activity being performed.
• Contextualised Assistance. The usefulness of a functionality depends on its appropriateness to the context. Therefore, AI capabilities should be linked to specific tasks and particular points in the workflow, rather than being presented as independent functionalities that are permanently available.
• AI as a Learning Support. Technology should contribute not only to task completion but also to the development of the skills required to use it. This principle is particularly relevant given that a lack of knowledge about how these tools operate was the most frequently identified adoption barrier in the questionnaire.

3.7.1. Summary of the Transformation Process

The findings therefore provide the basis for establishing a traceability chain between the empirical study and the subsequent development of the artefact. The analysis of technology adoption, perceived usefulness, ease of use, and behavioural intention to use was complemented by the identification of adoption barriers and the human factors underlying them. These findings were subsequently translated into design requirements and, ultimately, into design principles that informed the development of the artefact.

This chain constitutes a distinctive element compared with a conventional application of the TAM. In the present study, the model is not used exclusively to explain users’ intention to adopt a particular technology, but rather as a diagnostic instrument for identifying conditions that should be considered when designing an AI-assisted educational environment.

This distinction is important. A conventional application of the TAM might conclude that certain technologies exhibit higher or lower levels of acceptance. In contrast, the present study uses these differences to formulate design-oriented questions: What makes a technology perceived as useful? What makes it difficult to use? What conditions facilitate its integration? What support do users need to learn how to use it? And how should the interaction be designed to prevent technological assistance from increasing task demands?

The answers to these questions provided the conceptual basis for the decisions made during the development of the artefact. They also established a clear connection between the empirical findings and the resulting design decisions, which were subsequently examined through a second source of evidence provided by the heuristic evaluation.

3.8. Development and Operationalisation of the Design Principles

The results of the technology acceptance study were not used as a direct evaluation of the artefact, but rather as contextual evidence to inform its design. This distinction is important because the questionnaire focused on interpreters’ experiences and perceptions of different technologies currently used or potentially applicable in their professional practice, whereas the artefact itself was subsequently evaluated through a dedicated heuristic evaluation procedure.

The development of the environment followed an iterative approach grounded in DSR. Within this approach, the artefact is not merely the final outcome of the research, but also a means of generating and evaluating design knowledge. Accordingly, the decisions made during its development were grounded in evidence gathered throughout the different research phases and were progressively reviewed through cycles of implementation, verification, and refinement.

The development process comprised three iterations involving researchers and developers, while interpreters participated in testing the implemented functionalities. These profiles performed distinct roles: researchers assessed the alignment between design decisions and the research objectives; developers translated the requirements into operational functionalities; and interpreters assessed the suitability of the implemented workflows and behaviours from a professional perspective.

The iterations were not limited to programming cycles. Each iteration combined implementation, review, and verification, such that a functionality was not considered complete solely because it had been technically implemented. A functionality was considered complete only after its effective integration into the backend and subsequent verification by an interpreter. This verification made it possible to assess both the expected behaviour of the functionality and its suitability within the intended workflow.

Three main sources of feedback were used throughout this process: team meetings, researcher reviews, and internal testing with interpreters. Team meetings provided opportunities to review the status of the functionalities and make implementation decisions; researcher reviews assessed their alignment with the research objectives and design principles; and interpreter testing helped identify issues related to interaction workflows and the suitability of the functionalities for professional needs.

3.8.1. From Design Principles to Features

The operationalisation of the principles was grounded in a traceability process connecting the evidence obtained throughout the research with concrete implementation decisions. Empirical findings were interpreted in terms of the human factors they revealed, and these factors informed the formulation of design principles. The principles were subsequently translated into functional and interaction requirements, which guided the selection and implementation of specific artefact functionalities.

This procedure helped prevent functionalities from being selected solely on the basis of the capabilities offered by available AI technologies. Instead, technological capabilities were incorporated when they could address a need identified within the conceptual framework and supported by the empirical findings. In this sense, the role of technology was considered in relation to the requirements derived from the study rather than as the primary driver of the artefact design.

The principle of meaningful human oversight was translated into requirements aimed at preserving users’ ability to review, modify, and validate the results produced by the system. Rather than conceiving automation as a substitute for user activity, the functionalities were designed to provide information or assistance that could be examined and used according to the requirements of the task. This approach sought to ensure that users retained control over decisions for which human judgement remained essential.

The principle of contextualised assistance guided decisions concerning when and how AI-generated support should be presented. Assistance was intended to be provided in relation to the user’s current task and information needs, rather than being continuously or indiscriminately displayed. This principle informed the organisation of functionalities so that AI support could be accessed when relevant to the activity being performed without unnecessarily interrupting the user’s workflow.

The principle of transparency and calibrated trust was operationalised through decisions aimed at distinguishing automatically generated information from information entered or validated by the user. Transparency was understood primarily from a functional perspective: users should be able to understand what the system is doing, what type of information it is providing, and which limitations are relevant to the task. This approach is particularly important in AI systems whose outputs may appear plausible while nevertheless being incorrect.

The principle of protecting cognitive resources guided decisions concerning interface organisation, the reduction of unnecessary interactions, and the contextual presentation of information. These decisions followed from considering interpreting as an activity involving substantial attentional demands, in which the introduction of an additional source of information may impose further cognitive costs. Accordingly, AI assistance was incorporated in ways intended to support task performance without unnecessarily increasing the amount of information that users need to monitor or process.

Finally, the principle of AI as a learning support rather than a substitute for human competence had a particularly strong influence on the functionalities intended for training. Automation was conceived as a resource to facilitate practice, preparation, and reflection, while avoiding the removal of processes that learners need to develop themselves. This distinction is fundamental in an educational environment, where reducing the effort required to complete a task does not necessarily imply improved learning. Consequently, the design sought to use AI to extend opportunities for learning and feedback while preserving the learner’s active role in the development of the underlying competencies.

3.8.2. Correspondence Between Principles and Functionalities

The operationalisation of the principles resulted in a modular environment in which the different technological capabilities were distributed across distinct activities within the learning cycle. This organisation made it possible to integrate AI technologies without making them the central element of the interaction.

The functionalities developed included assignment management, glossary generation and management, note management, transcription, video-based training, memorisation, and training through speech synthesis.

This distribution reflects a relevant design decision: AI capabilities are not presented as generic functionalities applicable to any activity, but rather as resources associated with specific tasks. In this way, technological assistance is situated within the context of use in which it can provide value.

For example, glossary generation is linked to preparation activities; transcription is associated with speech processing and analysis; training functionalities are integrated into activities aimed at developing professional competencies; and memorisation mechanisms are used to reinforce the learning of relevant information.

This organisation operationalises the principle of contextualised assistance, according to which the usefulness of an AI capability depends not only on its technical performance but also on its alignment with the task being performed by the user.

3.8.3. Iteration, Verification and Refinement

The three iterations enabled the functionalities to be progressively reviewed before the formal evaluation. Team meetings, researcher reviews, and internal testing helped identify functional issues, interaction aspects requiring improvement, and discrepancies between the implemented functionalities and the expectations of professional users. These activities provided opportunities to revisit implementation decisions and progressively consolidate the different components of the artefact.

This process is important for subsequently interpreting the results of the heuristic evaluation. The formal evaluation was not conducted on a set of isolated functionalities or on components that were still at an early stage of implementation, but rather on an initial integrated version of the artefact incorporating the system’s main functionalities. The integrated version therefore represented the outcome of the preceding design and implementation activities, including the progressive incorporation of requirements derived from the conceptual framework and empirical findings.

Following the evaluation, the findings obtained were used to introduce further adjustments, primarily concerning visual aspects and interaction. Evaluation therefore provided additional evidence that could be used to revisit and refine previous design decisions. Rather than representing a final stage independent from development, the heuristic evaluation contributed to an ongoing process in which evidence from different stages of the research informed subsequent improvements to the artefact.

Taken together, the development process established a traceability relationship between the conceptual and empirical foundations of the study, the principles guiding the design, the resulting functional and interaction requirements, and their implementation through successive iterations. These activities led to an integrated version of the artefact that could be examined through heuristic evaluation, whose findings subsequently provided further input for design refinement.

This traceability constitutes a central element of the DSR approach adopted in this research. The TAM results informed the definition of design needs and conditions; these findings were interpreted through the human factors identified in the study and translated into design principles and concrete requirements; the iterative development process enabled these requirements to be progressively incorporated into the artefact; and the heuristic evaluation provided additional evidence concerning the interaction problems that remained in the integrated version. In this way, design and evaluation formed part of a single process of generating and refining design knowledge, rather than constituting independent activities.

3.9. Heuristic Evaluation

Once the first version of the environment had been developed and integrated, a heuristic evaluation was conducted to identify interaction problems and determine the extent to which the implemented design decisions were consistent with the human-centred principles established in the previous phases. The evaluation was conceived as an inspection activity focused on the artefact rather than as a further measurement of technology acceptance. Thus, while the TAM study characterised the perceptions and needs associated with the use of assistive technologies in the professional context, the heuristic evaluation directly examined interaction with the developed system.

The evaluation was conducted by a multidisciplinary team of four evaluators: two interpreters, one researcher, and one developer. This composition made it possible to combine three complementary perspectives: professional domain knowledge in interpreting, a methodological research perspective, and software development expertise. The participation of different professional profiles is particularly relevant for an artefact that combines educational functionalities, language-processing technologies, and interaction mechanisms intended for users with specific professional needs.

The inspection was structured around Nielsen’s ten usability heuristics: visibility of system status; match between the system and the real world; user control and freedom; consistency and standards; error prevention; recognition rather than recall; flexibility and efficiency of use; aesthetic and minimalist design; recognition, diagnosis, and recovery from errors; and help and documentation. Each finding was associated with the affected heuristics and classified using a five-level severity scale: 0, when no usability problem was identified; 1, for cosmetic problems; 2, for minor problems; 3, for major problems; and 4, for critical problems.

This classification made it possible to distinguish between issues that could be addressed through minor interface adjustments and those requiring prioritised intervention. The use of a severity scale is particularly relevant in the context of the present study because the aim of the evaluation was not merely to identify the largest possible number of problems, but to determine which issues could most significantly compromise interaction with the environment.

3.9.1. Overall Distribution of the Findings

In total, the evaluation identified 34 usability findings distributed across the different functionalities of the artefact. The issues were not evenly distributed. The highest concentration was found in functionalities related to glossary generation and management, assignment management, and transcription, whereas other functionalities presented fewer issues. The memorisation module did not yield any findings during the evaluation.

This result should be interpreted in light of the different functional complexity of the modules. Functionalities related to glossary generation and management, for example, involve operations for creating, consulting, editing, and managing information, whereas transcription additionally introduces an automated processing workflow whose states and results need to be communicated to the user. Therefore, a higher concentration of findings in these areas does not necessarily indicate lower overall module quality, but may instead reflect a greater density of interactions and system states.

From this perspective, the distribution of findings also provides useful information for the DSR process, as it makes it possible to identify which parts of the artefact require greater attention during subsequent refinement.

3.9.2. Heuristics Affected

The 34 findings were concentrated primarily across four heuristics: error prevention; recognition rather than recall; user control and freedom; and consistency and standards. The remaining heuristics were considerably less represented among the identified problems.

The concentration of findings across these four heuristics is particularly significant from the perspective of the conceptual framework developed in this research. Although Nielsen’s heuristics were originally formulated as general usability principles, the identified problems can be reinterpreted within the human factors framework used to guide the design of the environment.

Error prevention is directly related to the need to avoid requiring users to retrieve information or correct situations that could have been prevented through clearer design. In an AI-based educational environment, this issue becomes particularly important when certain actions may trigger automated processes or modify information that will subsequently be used in other activities.

Problems related to recognition rather than recall indicate the need for the interface to provide sufficient information for users to identify relevant options and system states without having to remember how the system operates. This issue is particularly important in an environment comprising multiple interconnected modules and workflows.

Problems associated with user control and freedom are particularly relevant to the principle of human agency. If users do not understand how to cancel an action, modify a decision, or return to a previous state, the technology may effectively constrain their ability to exercise control, even when the underlying functionality has been designed to preserve human oversight.

Finally, problems concerning consistency and standards affect the predictability of interaction. An interface that uses different patterns for equivalent actions increases the need for system-specific learning and requires users to devote additional cognitive resources to understanding how the system operates.

3.9.3. Severity of the Problems

The severity distribution showed that most findings corresponded to problems of low or moderate severity, although some more significant issues requiring prioritised intervention were also identified.

This result is important for properly interpreting the evaluation. The identification of 34 findings does not mean that the artefact presented 34 critical problems or that its use was unfeasible. Rather, the severity classification made it possible to distinguish issues primarily affecting presentation or interaction from those with the potential to more substantially compromise task performance.

This distinction is consistent with the iterative nature of the DSR process. Lower-severity issues could be addressed through incremental adjustments, whereas higher-severity issues provided evidence for revisiting design decisions with a greater potential impact on interaction.

3.9.4. Interpreting the Findings in the Light of Design Principles

Beyond their classification according to Nielsen’s heuristics, the findings were interpreted in relation to the design principles that guided the development process. This second layer of analysis prevents the evaluation from being reduced to a list of interface problems and allows the findings to be used as evidence concerning the operationalisation of the design principles.

The first relevant aspect is the relationship between usability and the protection of cognitive resources. Problems related to recognition, consistency, error prevention, and user control may increase the effort required to interact with the system. Consequently, an issue that may appear minor from a purely visual perspective can become more significant when the system is used during an activity that already involves substantial cognitive demands.

For example, when users need to remember where a particular functionality is located, interpret an inconsistent interaction pattern, or perform additional steps to complete an action, part of their cognitive resources must be devoted to managing the interface rather than to the primary task. This issue is particularly relevant in interpreter training, where interaction with the system may coexist with activities involving comprehension, production, memory, and decision-making.

Findings related to user control and freedom also provide evidence for the principle of human agency. Human oversight does not depend solely on whether the system technically allows users to modify an AI-generated output; it also depends on whether this possibility is accessible and understandable to the user.

Similarly, problems related to error prevention can be interpreted in relation to the principle of transparency and calibrated trust. A system that does not adequately communicate its states or that makes it difficult to recover from an action may create uncertainty about what has occurred and about the consequences of the actions performed.

Finally, problems concerning consistency and recognition provide evidence of the need to design interfaces that enable users to progressively learn how the environment operates. This issue connects directly with the principle of using AI as a learning support: if the environment itself requires excessive effort to learn how to use it, it may become a barrier to the learning process it is intended to facilitate.

3.9.5. Evaluation as a Means of Refinement

The evaluation results were not used solely to determine whether the first version of the artefact could be considered acceptable. Rather, the findings were incorporated into the iterative development process as evidence for refining design decisions. In particular, the identified problems led primarily to visual and interaction adjustments aimed at improving the user experience. This process made it possible to address interface issues that had not been identified during internal testing and to subsequently assess the consistency between the modifications introduced and the principles that had guided the design.

This role of evaluation is important from a DSR perspective. The knowledge generated does not derive exclusively from the final outcome, but also from the relationship between an identified problem, the design decision through which it was addressed, and the modification introduced in response. The heuristic findings therefore provided evidence not only about specific usability problems, but also about the conditions under which human-centred design principles could be effectively implemented in an AI-assisted educational environment.

The heuristic evaluation thus complemented the traceability established through the TAM study. The empirical findings from the TAM analysis had previously been interpreted in terms of human factors and translated into design requirements, principles, and artefact functionalities. The heuristic evaluation examined how these principles and requirements were manifested in the concrete interaction design of the integrated artefact. The problems identified through this evaluation subsequently informed further refinements to the corresponding design decisions.

Taken together, these relationships make it possible to reconstruct the process through which design knowledge was generated and refined. The process began with contextual and empirical evidence, which informed the identification of relevant human factors and the formulation of design principles. These principles were operationalised through functional and interaction requirements and subsequently incorporated into the artefact through iterative development. The resulting integrated version was then evaluated, generating additional evidence that could be used to revisit and refine the design decisions.

From this perspective, the heuristic evaluation does not constitute an isolated phase at the end of development, but rather a mechanism through which the design principles are examined in their concrete implementation. Its contribution lies not only in identifying usability problems, but also in establishing a relationship between principles, implementation decisions, observed interaction issues, and subsequent refinements. Evaluation therefore forms part of the same cumulative process of generating and refining design knowledge that connects the empirical study with the development of the artefact.

3.10. Design Refinement and Integrated Findings

The integration of the different sources of evidence makes it possible to identify three main findings.

First, the study involving 42 professional interpreters showed that the acceptance of assistive technologies is not homogeneous. Technologies that are more established within professional workflows, particularly those related to terminology management, received more favourable evaluations, whereas more recent AI technologies exhibited greater heterogeneity in perceived usefulness, perceived ease of use, and future intention to use. In addition, lack of training emerged as one of the main reported barriers to adoption.

Second, these findings were not translated directly into functionalities, but rather into human factors and design requirements. This transformation introduced an explicit layer of design reasoning between empirical evidence and implementation.

Third, the heuristic evaluation of the first integrated version showed that even when functionalities are conceptually aligned with the established principles, their implementation at the interface level may introduce interaction problems. The 34 findings identified, particularly those related to error prevention, recognition, user control, and consistency, provided evidence for subsequent refinement of the artefact.

The overall result is therefore a process in which evaluation is not limited to asking whether users would accept a particular technology or to determining whether an interface presents usability problems. Instead, evidence is used to establish connections between technology acceptance, human factors, design decisions, and interaction with the artefact.

This integration provides the basis for the subsequent discussion, which examines the design knowledge that can be derived from this process and the extent to which the findings enable the research questions to be addressed.

4. Discussion

4.1. From Technological Capability to Educational Suitability

The findings of this study suggest that the main challenge associated with incorporating AI into interpreter training does not lie exclusively in determining which technological capabilities can be integrated into an educational environment, but rather in establishing how, when, and under what conditions these capabilities should be made available to learners. This conclusion emerges from the convergence of the technology acceptance study, the artefact development process, and the heuristic evaluation.

The study involving 42 professional interpreters showed a generally positive assessment of the potential of digital technologies to support interpreting practice, but also revealed important differences between established and emerging technologies. Terminology management tools, which are already embedded in established professional practices, exhibited more homogeneous acceptance patterns, whereas more recent AI-based technologies showed greater heterogeneity in perceived usefulness, perceived ease of use, and future intention to use. This finding is consistent with the original conception of the TAM, according to which perceived usefulness and perceived ease of use are fundamental determinants of technology acceptance [13]. However, the results obtained suggest that these constructs should be interpreted in relation to the specific characteristics of the activity in which the technology is introduced.

In particular, a technology may be perceived as potentially useful without necessarily being incorporated into the professional workflow. In the context examined, adoption is influenced by additional factors such as familiarity with the tool, trust in its outputs, the cognitive load associated with its use, and its compatibility with existing professional practices. This observation is consistent with research on CAI, which has highlighted the need to consider the temporal and cognitive constraints specific to interpreting when introducing technology into the activity [8].

This distinction is particularly relevant in simultaneous interpreting. During this activity, interpreters must simultaneously manage processes of comprehension, memory, production, and monitoring. According to Gile’s Effort Models, these activities compete for limited cognitive resources [7]. Therefore, the introduction of an additional source of information cannot automatically be considered beneficial merely because it provides more information.

From this perspective, our findings qualify a common conception of AI-based innovation: greater automation or more information does not necessarily result in a better educational tool. The value of an AI capability depends on the relationship between the benefit it provides, the cognitive cost of accessing it, and the objective of the activity in which it is used.

This conclusion connects directly with the literature on HCAI. Human-centred systems should not be designed solely to maximise technological capabilities, but rather to establish an appropriate relationship between automation, human control, and human capabilities [6,18]. In the present study, this perspective is extended to the educational domain: the question is not merely whether AI can perform a particular task, but whether automating that task genuinely contributes to the learning process.

4.2. RQ1: Limitations of AI-Based CAI Environments for Interpreter Training

The first research question addressed the limitations of current AI-based CAI environments when they are transferred to the context of interpreter training.

The findings allow four main limitations to be identified.

The first is the disconnect between technological capability and context of use. CAI tools have evolved from functions primarily focused on terminology management towards capabilities related to automatic speech recognition, content generation, and real-time assistance [8,9]. However, the availability of more advanced capabilities does not guarantee that they can be effectively integrated into activities characterised by high cognitive demands.

The second limitation concerns the uncertainty associated with AI-generated outputs. Unlike traditional deterministic systems, AI models may produce plausible but incorrect outputs and exhibit varying levels of uncertainty. Consequently, interaction requires mechanisms that enable users to understand the system’s capabilities, detect potential errors, and determine when its outputs can be trusted. This need is consistent with the guidelines for human–AI interaction proposed by Amershi et al. [11] and with the concept of calibrated trust developed in recent human–AI interaction research [5].

The third limitation is the cognitive load associated with interaction. Participants’ comments indicate that consulting information during interpreting may interfere with the primary activity. In this regard, technology may paradoxically become an additional source of cognitive load when it requires interpreters to simultaneously monitor the system and perform the professional task.

The fourth limitation is educational in nature. A tool designed to support professional practice does not necessarily constitute an appropriate learning tool. In education, automation should be considered in relation to the opportunities it provides learners to practise, make decisions, receive feedback, and progressively develop their competencies. Research on AI in education has specifically highlighted the need to shift the focus from mere technological adoption towards pedagogically meaningful integration [2,3].

Consequently, the answer to RQ1 does not consist of identifying a specific technological deficiency that should be addressed through a new functionality. The main limitation identified is the lack of systematic integration between technological capabilities, human–AI interaction, and educational purpose.

4.3. RQ2: Human Factors That Should Guide the Design

The second research question focused on the human factors that should guide the design of AI-assisted educational environments for interpreter training.

The evidence obtained made it possible to identify five interrelated dimensions: human agency and control, transparency and calibrated trust, protection of cognitive resources, contextualised assistance, and AI as a learning support.

The first dimension, human agency and control, implies that users should retain the ability to intervene in the outputs produced by the system. This principle is consistent with the HCAI framework proposed by Shneiderman [6] and with the human–AI interaction guidelines described by Amershi et al. [11]. In an educational context, however, control takes on an additional dimension: it not only protects users against system errors but also enables learners to remain active agents in their own learning process.

The second dimension is transparency and calibrated trust. The TAM results show that acceptance of emerging technologies is influenced by concerns regarding their reliability. Therefore, the design objective should not be to maximise trust, but rather to ensure that trust is proportional to the system’s actual capabilities. Excessive trust may lead to inappropriate reliance on automated outputs, whereas insufficient trust may prevent users from taking advantage of the available capabilities [5].

The third dimension concerns the protection of cognitive resources. This dimension is particularly relevant to interpreting because the cost of interacting with a technology must be considered alongside the cognitive demands of the activity itself [7]. The principle implies that the interface should not require learners to perform interaction-related efforts that do not provide educational or functional value.

The fourth dimension is contextualised assistance. The results show that technologies are used differently depending on the stage of an assignment. This suggests that assistance should not remain constant throughout the entire workflow. The appropriate level of automation depends on the task, the point in the workflow, and the user’s level of experience.

Finally, the fifth dimension concerns the conception of AI as a learning support rather than a substitute for human competence. This dimension arguably represents the most important distinction between a CAI system intended for professional use and an educational environment designed to train future professionals. A functionality that maximises task efficiency may be pedagogically counterproductive if it eliminates precisely the cognitive activity that learners need to develop.

Taken together, these factors provide a basis for extending traditional technology acceptance models. TAM, originally proposed by Davis [13], offers a useful framework for understanding perceived usefulness, perceived ease of use, and behavioural intention to use. However, interaction with AI systems introduces dimensions that are not fully represented in the original model, including transparency, human control, uncertainty, and oversight [6,11].

Therefore, the findings suggest that technology acceptance should be considered a necessary, but not sufficient, condition for the design of AI-assisted educational environments.

4.4. RQ3: From Human Factors to Design Requirements

The third research question addressed how these human factors could be operationalised into concrete design requirements.

The main methodological contribution of the study lies in establishing an explicit traceability process through which empirical evidence was progressively transformed into design knowledge. The human factors identified through the empirical study provided the basis for formulating the design principles. These principles were then operationalised through functional and interaction requirements, which guided the implementation of specific artefact functionalities. The resulting artefact was subsequently evaluated, and the findings from this evaluation provided further evidence for revisiting and refining the corresponding design decisions.

This process makes it possible to address a common challenge in human-centred system design: although general principles may be conceptually relevant, they do not by themselves specify how those principles should be materialised in a concrete interface. Operationalisation therefore requires an intermediate level at which abstract principles are translated into requirements that can guide specific design and implementation decisions.

For example, identifying the need to preserve human control does not in itself constitute a design decision. This need must be translated into concrete requirements, such as enabling users to review, modify, accept, or discard automatically generated outputs. Similarly, identifying cognitive load as a relevant factor should lead to decisions concerning the amount of information presented, the organisation of the interface, and the number of interactions required to complete a task.

The DSR process made it possible to use the artefact as a mechanism for materialising and examining these relationships. In this respect, the study builds on the DSR approaches described by Hevner et al. [14] and Peffers et al. [15], in which the construction and evaluation of artefacts serve as means of generating design knowledge. The artefact was therefore not treated merely as the technological outcome of the research, but as a means through which abstract design principles could be instantiated, examined, and refined.

The contribution therefore does not simply consist of implementing a platform with different AI functionalities. Rather, it lies in documenting how a set of human factors can be interpreted as design principles, translated into concrete requirements, materialised through artefact functionalities, and subsequently examined through evaluation. This provides a more explicit account of how human-centred knowledge can inform the design of AI-assisted educational environments.

This distinction is particularly important in the field of AI in education. Reviews of AI in education have shown the growing incorporation of intelligent systems while also highlighting the need to better understand their integration with the needs of students and educators [3]. From a design perspective, this entails moving beyond the question “What can AI do?” towards the question “What role should AI play within a given learning activity?” The latter question places technological capabilities within the context of human needs, task demands, and educational objectives, thereby providing a more appropriate basis for the design of human-centred AI systems.

4.5. RQ4: Design Principles Derived from the Process

The fourth research question addressed which human-centred design principles could be derived from the development and evaluation of the environment.

The findings support the proposal of five principles that operate as an intermediate level between general HCAI frameworks and concrete implementation decisions: meaningful human control; transparency and calibrated trust; protection of cognitive resources; contextualised assistance; and AI as a learning support.

The relevance of these principles lies in the fact that they do not merely constitute general recommendations for responsible AI. Each principle is linked to evidence obtained within the specific context of interpreting and was subsequently operationalised through functionalities implemented in the artefact.

The first principle extends the general notion of human control to an educational activity in which users must retain not only the ability to correct the system but also responsibility for their own decisions.

The second transfers the issue of trust in AI systems to the educational context. Learners need to learn not only how to use a tool, but also how to critically evaluate its outputs.

The third incorporates the specific cognitive characteristics of interpreting. An interface may be technically functional while simultaneously introducing an interaction burden that is incompatible with the demands of the task.

The fourth establishes that assistance should be context-dependent. The same AI capability may be appropriate during preparation but less appropriate during a real-time interpreting task.

Finally, the fifth principle explicitly introduces the pedagogical dimension. AI should facilitate opportunities for practice, analysis, and feedback without eliminating the activities through which competence is developed.

These principles position the proposal at an intermediate level between general HCAI frameworks and the concrete functionalities of an educational system. Whereas the HCAI framework proposed by Shneiderman [6] establishes the need to maintain a balanced relationship between people and intelligent systems, the proposed principles translate this orientation into a context characterised by professional learning, high cognitive demands, and the need for human oversight.

4.6. Heuristic Evaluation as Evidence of the Realisation of Principles

The heuristic evaluation provides a second perspective on the contribution of the study. The 34 findings identified show that the conceptual adoption of human-centred principles does not automatically guarantee their effective implementation at the interface level.

This finding is relevant because it reveals a distinction between conceptual design and interaction design. A functionality may have been conceived to preserve user control and yet present navigation or action-recovery problems that hinder such control. Similarly, an interface may incorporate information intended to promote transparency while presenting that information in a way that is difficult for users to recognise. The evaluation therefore provided an opportunity to examine whether the principles formulated at the conceptual level were effectively reflected in the concrete interaction with the artefact.

The concentration of problems in the heuristics of error prevention, recognition rather than recall, user control and freedom, and consistency and standards is particularly significant in this regard. These heuristics directly affect the amount of effort users must devote to understanding and managing the interface [19]. Their relevance is especially evident in AI-assisted environments, where users may already need to monitor, interpret, and assess information generated by the system.

Within the proposed framework, these findings support the interpretation of interaction quality as an intermediate layer connecting technological capabilities with the conditions under which educational activities take place. The presence of an AI capability does not in itself determine its educational value. Its contribution depends on how that capability is incorporated into the interaction, how users experience and control it, and how well its use is compatible with the requirements of the learning activity. Educational outcomes therefore depend not only on what the technology can provide, but also on whether that support can be meaningfully incorporated into the activity.

From this perspective, a technological capability can contribute to learning only when users can interact with it in a manner that is comprehensible, controllable, and compatible with the demands of the task. This reinforces the human-centred orientation of the proposed framework, in which technological capabilities are considered in relation to interaction quality, user needs, and the characteristics of the educational activity rather than as ends in themselves.

This interpretation also explains why the findings from the heuristic evaluation should not be regarded merely as a list of interface defects. Rather, the findings provide information about where and how design principles may fail during implementation. A usability problem can therefore reveal a discrepancy between an intended design principle and its concrete realisation in the interface. Consequently, evaluation becomes a source of knowledge for design refinement, allowing the principles and requirements to be reconsidered in light of the interaction problems observed in the artefact.

4.7. Implications for the Design of AI-assisted Educational Environments

The findings allow several broader implications to be drawn for the design of AI-based educational environments.

First, automation should be selective and context-dependent. Preparation, organisation, and post-task analysis may accommodate relatively high levels of automation, whereas activities in which learners need to develop a specific competence require more active human involvement.

Second, human oversight should be designed as a functional property of the system rather than as an expectation placed on user behaviour. It is not sufficient to assume that learners will verify AI-generated outputs; the system should explicitly facilitate such verification.

Third, transparency should be action-oriented. The objective should not be to explain every technical aspect of the underlying model, but rather to provide users with the information necessary to understand what the AI has produced, what they can do with the resulting output, and when it should be reviewed.

Fourth, ease of use should be considered in conjunction with cognitive load. An interaction may appear simple from an interface perspective while remaining inappropriate if it requires learners to process irrelevant information during a cognitively demanding task.

Fifth, educational design should distinguish between automating a task and automating a competence. This distinction constitutes one of the main implications of the findings. Automating the preparation of learning materials may expand learning opportunities, whereas fully automating the activity that learners need to master may reduce them.

Finally, the findings suggest that the evaluation of AI-assisted educational environments should move beyond traditional technology acceptance measures. Perceived usefulness and perceived ease of use are important indicators, but they should be complemented by measures of cognitive load, user control, trust, usage behaviour, and learning outcomes. This perspective is consistent with the need for human-centred approaches in research on AI in education [4].

4.8. Limitations and Implications for Future Research

The findings should be interpreted in light of several limitations.

First, the proposed principles are derived from a specific context: interpreter education and professional interpreting practice. Although some factors—such as human control, transparency, and calibrated trust—may be relevant in other contexts, the specific configuration of these principles cannot be assumed to be universally applicable.

Second, the technology acceptance study was conducted with 42 professional interpreters. This sample provides a relevant perspective on the needs and expectations of the professional context, but it does not allow the findings to be assumed to represent the entire population of interpreters or, in particular, students with different levels of training.

Third, the use of TAM in this study was primarily descriptive and design-oriented. The results make it possible to characterise patterns of perceived usefulness, perceived ease of use, and behavioural intention to use, but the study did not involve a statistical model aimed at demonstrating causal relationships among these constructs. This distinction should be made explicit to avoid overinterpreting the findings.

Fourth, the heuristic evaluation was conducted by four evaluators with complementary profiles. This approach made it possible to identify interaction problems from different perspectives, but it does not provide evidence regarding the artefact’s actual impact on learning over the long term.

Finally, the study does not yet allow us to determine whether use of the environment effectively improves interpreting competence. Addressing this question would require studies involving learners, longitudinal research designs, and objective performance measures.

These limitations point to several directions for future research. In particular, future studies should experimentally investigate how different levels of assistance affect performance and cognitive load; examine changes in trust and reliance on AI throughout the learning process; compare users with different levels of experience; and investigate whether the five principles identified can be transferred to other educational domains characterised by cognitively complex activities.

4.9. Overall Contribution of the Study

Taken together, the findings support the formulation of the main contribution of the study as a traceable design framework for the integration of AI into educational environments that support cognitively complex activities.

The contribution does not primarily depend on the specific technologies incorporated into the artefact. Speech recognition, content generation, semantic representation, and speech synthesis technologies can be replaced as new models and technical capabilities emerge. What remains more stable is the design process through which technological possibilities are considered in relation to human needs, interaction requirements, and educational objectives. From this perspective, the educational relevance of an AI capability cannot be determined solely by what the technology is able to perform. It also depends on how that capability is incorporated into the interaction and on whether its use is compatible with the requirements of the learning activity.

This perspective provides a basis for reconsidering the role of AI in interpreter education. Rather than treating automation as an end in itself, the design of AI-assisted learning environments should consider the degree of automation that is appropriate for each activity, the level of control that learners should retain, and which aspects of the learning process should remain under human responsibility. Such considerations are particularly important in cognitively demanding activities, where excessive automation may reduce opportunities for active engagement, monitoring, decision-making, and the development of professional competence.

The main contribution of the study therefore lies in shifting the focus from the question of what AI can automate towards a more educationally and human-centred design question: how AI can be integrated to augment human capabilities without compromising learner agency, cognitive engagement, or learning opportunities.

This shift constitutes the core of the conceptual and methodological contribution of the study. The TAM results provide empirical evidence about users’ perceptions, needs, and concerns in the context of AI-assisted learning. These findings are interpreted through the identification of relevant human factors, which provide a basis for formulating design principles and translating them into functional and interaction requirements. The DSR process then uses these requirements to guide the construction and iterative development of the artefact, while heuristic evaluation provides additional evidence about how the principles are realised in concrete interactions and where further refinement is required.

The resulting knowledge is therefore not tied to any particular technological implementation. Instead, it lies in the systematic approach used to connect empirical evidence, human factors, design principles, requirements, artefact construction, and evaluation. In this sense, the study contributes not only an AI-assisted educational artefact but also a design-oriented account of how technological capabilities can be integrated into learning environments while preserving human agency, supporting cognitive engagement, and maintaining alignment with the educational purposes of the activity.

5. Conclusions

The integration of AI into educational environments poses a challenge that extends beyond the availability of new technological capabilities. In professional training contexts characterised by high cognitive demands, such as interpreting, it is necessary to determine what role AI should play, what degree of automation is appropriate, and how it can be integrated without reducing learner agency or learning opportunities. This study addressed this challenge through a DSR process focused on the design and evaluation of an AI-assisted educational environment for interpreter training.

The findings show, first, that technology acceptance does not depend exclusively on functional capabilities. The study involving 42 professional interpreters revealed greater familiarity with and adoption of established tools, particularly those related to terminology management, compared with the more limited and heterogeneous use of emerging AI-based technologies. The differences observed in perceived usefulness, perceived ease of use, and behavioural intention to use, together with participants’ qualitative responses, highlight the relevance of factors such as trust, perceived reliability, cognitive load, technological training, and compatibility with existing workflows.

These findings addressed the first and second research questions by identifying a set of conditions that should be considered when AI is introduced into interpreter training. In particular, technology should be integrated with consideration of the cognitive constraints of the activity, the need for human oversight, and the differences between the various stages of the interpreting process. The findings also show that technology acceptance does not in itself constitute evidence of educational suitability. A tool may be perceived as useful while still being inappropriate for a particular learning activity if it reduces learners’ cognitive engagement or introduces additional interaction demands.

Based on these observations, the study proposed five design principles: meaningful human control, transparency and calibrated trust, protection of cognitive resources, contextualised assistance, and AI as a learning support. These principles constitute the main conceptual contribution of the study. Their value does not lie in presenting them as a universal set of rules, but rather in providing an intermediate structure through which human and contextual factors can be translated into design requirements for AI-assisted educational environments.

The third research question was addressed through a traceability matrix connecting empirical evidence with human factors, design requirements, principles, implemented functionalities, and evaluation results. This traceability makes explicit the reasoning process linking data obtained from the professional context to the decisions adopted during artefact development. In this way, the design process becomes less dependent on purely technology-driven decisions and instead becomes grounded in evidence concerning users’ characteristics and needs.

The heuristic evaluation provided a second source of evidence. The problems identified during this phase show that the conceptual incorporation of human-centred principles does not automatically guarantee their appropriate implementation at the interface level. The identification of problems primarily related to error prevention, recognition rather than recall, user control and freedom, and consistency made it possible to identify interaction aspects requiring refinement. Consequently, the evaluation was not used solely to determine whether the artefact was usable, but also as a mechanism for examining how the design principles had been materialised in the system and for guiding subsequent iterations.

In response to the fourth research question, the study therefore shows that design principles can act as a link between evidence about users, design requirements, and implementation decisions. This relationship constitutes a relevant contribution of the approach adopted, as it makes it possible to document not only which functionalities are incorporated into an AI-assisted educational environment, but also why they are incorporated and which human factors they are intended to address.

Nevertheless, the findings should be interpreted in light of the study’s limitations. The technology acceptance analysis was based on a sample of 42 professional interpreters and therefore provides evidence concerning the professional context, but does not allow the findings to be directly generalised to interpreting students. Similarly, TAM was used to characterise perceived usefulness, perceived ease of use, and behavioural intention to use, rather than to establish causal relationships among its constructs. Furthermore, the heuristic evaluation provides evidence concerning interaction problems but does not constitute an assessment of the environment’s impact on learning or on the development of interpreting competence.

These limitations also define the main directions for future research. Future studies should evaluate the environment with interpreting students and employ experimental or longitudinal designs to examine changes in performance, cognitive load, trust, and AI-use strategies. Further research should also investigate different levels of automation and determine how their suitability varies according to the task, level of experience, and stage of the learning process. Finally, future research should examine the extent to which the five principles identified can be transferred to other educational contexts characterised by complex tasks and close interaction between human capabilities and AI systems.

Overall, this study advocates a shift from a perspective focused on what AI can do towards one concerned with how AI should be integrated to support learning. The main contribution is therefore neither a specific technology nor a particular artefact functionality, but a design process that connects empirical evidence, human factors, design principles, implementation, and evaluation. From this perspective, AI is not conceived as a substitute for human activity, but as a resource whose educational value depends on whether its integration preserves learner control, appropriately manages cognitive demands, and contributes to developing rather than replacing the competencies that learners need to acquire.