The Measurement-Entitlement Gap: Epistemic and Ethical Limits of AI-Based Inference from Digital Traces

1. Introduction

Research on human beings frequently relies on self-report. In psychology, management, education, health research, and the social sciences, respondents are asked to describe their attitudes, experiences, intentions, and behaviour. This practice is routinely criticised because a report about behaviour is not the behaviour itself. Memory is selective, questions are interpreted differently, and responses may be affected by social desirability, strategic self-presentation, or limited introspective access. The familiar methodological aspiration is therefore to move closer to “actual behaviour” rather than merely collect verbal accounts of it [1].

The objection should not be overstated. Self-report can be the most appropriate source of evidence when the target of inquiry is a subjective experience, belief, intention, or felt state. Moreover, methodological problems attributed to common-method variance do not automatically invalidate findings based on questionnaires [2]. Nevertheless, the distinction between a person’s behaviour and a person’s description of that behaviour remains important. For much of the history of organisational research, direct and continuous observation was expensive, reactive, difficult to standardise, and feasible only for limited samples.

Digitalisation has radically altered this situation. Email systems, enterprise messengers, project-management platforms, videoconferencing systems, access-control systems, productivity applications, and wearable devices generate persistent records of organisational activity. Machine-learning methods can transform these records into predictions and classifications at a scale that would previously have been impossible. What was once primarily a methodological aspiration has therefore become a technical possibility.

Instead of asking whether an employee is engaged, researchers can analyse linguistic sentiment, response times, communication-network position, participation in meetings, or patterns of software use. Instead of asking about stress, they can analyse vocal characteristics, working hours, smartphone use, or biometric signals. Instead of relying on an employee’s declared intention to leave, they can attempt to predict turnover from communication and activity data. Under this view, digital traces promise a transition from subjective testimony to objective observation. Recent work has used GPT-3.5 and GPT-4 to infer Big Five personality traits from Facebook status updates by comparing model outputs with self-reported trait scores; the reported accuracy also varied across gender and age groups [3]. This type of study makes the distinction between predictive agreement with a label and warranted construct measurement especially consequential.

This promise is overstated for four reasons. First, digital traces are not behaviours in an unmediated form. They are records generated by technical systems designed for particular operational purposes. Second, the behavioural events registered by a system are not identical with the theoretical constructs about which researchers wish to make claims. Third, a machine-learning output is not a discovered property of a person but the result of a sequence of representational and inferential decisions. Fourth, obtaining a comprehensive record of a person’s activity may require forms of surveillance that interfere with privacy, autonomy, trust, and the social conditions under which the recorded activity occurs.

The last point creates a distinctive connection between epistemology and ethics. The problem is not simply that ethical constraints restrict otherwise epistemically ideal research. In studies of human behaviour, the method by which evidence is generated can alter the phenomenon being studied. Awareness of monitoring may change language, channel choice, effort allocation, willingness to participate, and strategies of self-presentation. Power asymmetries may affect who consents and what they communicate. Anticipated consequences may induce employees to optimise visible indicators rather than pursue the substantive aims of their work. Normatively defective data practices may therefore generate epistemically defective evidence.

This article examines the following question:

Under what conditions can an algorithmic inference from digital traces constitute a warranted measurement of a human construct, and what role do the normative conditions of data generation play in establishing that warrant?

The central thesis is that a scientifically legitimate algorithmic measurement practice concerning a person requires dual warrant. The measurement claim requires epistemic warrant: the model output must be justifiably interpreted as evidence about the target construct. The research practice requires normative warrant: researchers must be entitled to generate, transform, and use that evidence in the relevant context. These requirements are analytically distinguishable but not entirely independent.

The concept of a measurement–entitlement gap refers to the distance between three propositions:

• a system is technically capable of producing a score;
• the score is epistemically warranted as a measurement of a specified construct;
• researchers or organisations are normatively entitled to generate and use that measurement.

The first proposition does not imply the second, and neither implies the third. Technical computability, epistemic validity, and normative entitlement are different properties of a research practice. Accordingly, the measurement–entitlement gap is not a gap between truth and morality. It is a gap between technical computability, warranted interpretation, and the legitimacy of the scientific practice through which a measurement claim is produced and used.

Existing scholarship has examined construct validity in computational systems, the ethical risks of algorithmic decision-making, the contextual integrity of information flows, and the organisational consequences of workplace monitoring. However, these discussions have largely proceeded in parallel. Measurement scholarship typically concentrates on the relation between constructs, indicators, and scores, whereas algorithmic ethics concentrates on data access, fairness, privacy, and downstream use. What remains insufficiently developed is an integrated account of how the normative conditions under which digital records are generated affect both researchers’ entitlement to use them and the epistemic status of the resulting measurement claims. The dual-warrant framework is intended to address this gap.

The argument makes four contributions. First, it reconstructs an algorithmic measurement pipeline separating the target construct, behavioural context, behaviour, recording system, digital record, computational representation, label, model output, and interpreted measurement claim. Second, it identifies four gaps that must be addressed by a validation argument: recording, representation, proxy-label, and interpretation gaps. Third, it formulates the invasive proxy paradox, in which a more intrusive method remains epistemically dependent on the self-report measure it purports to replace. Fourth, it proposes a dual-warrant framework integrating construct validity, data provenance, contextual integrity, research ethics, and a praxiological distinction between efficient and permissible action.

Employee engagement provides the principal test case because it is a widely used organisational construct and an attractive target for workplace analytics. The argument is nevertheless intended to apply more broadly to algorithmic inferences about stress, depression, attention, learning, risk, productivity, trustworthiness, and other human attributes derived from digital traces.

2. From Digital Traces to Measurement Claims

Measurement is not merely the production of a number. Following Tal [4], measurement can be understood as a knowledge-producing interaction with a concrete system that supports an abstract representation of selected aspects of that system. It is an organised epistemic activity in which aspects of concrete systems are represented in an abstract form and interpreted under theoretical, operational, and inferential assumptions. This is particularly important when the target is a construct that is not directly observable. Engagement, intelligence, stress, organisational commitment, trust, and risk are not objects that can simply be read from a sensor. Their empirical meaning depends on relations between concepts, indicators, procedures, contexts, and uses.

The classic account of construct validity treats validation as the accumulation of evidence supporting an interpretation of a test as a measure of an attribute [5]. Messick [6] and Kane [7] subsequently emphasised that validity concerns the interpretations and uses made on the basis of scores rather than being a permanent property residing in an instrument. Borsboom et al. [8] formulate a stronger realist requirement: an instrument measures an attribute validly when the attribute exists and variations in it causally produce variations in the measurement outcome. Although these accounts differ, all reject the inference that numerical output by itself establishes measurement.

Contemporary computational systems frequently operate on latent, multidimensional, or normatively contested constructs. A risk score, productivity score, sentiment score, or engagement score may be represented with several decimal places while the relationship between the number and the construct remains poorly specified. Numerical precision can conceal conceptual indeterminacy. As Jacobs and Wallach [9] argue, many problems attributed to computational systems originate in mismatches between theoretical constructs and their operationalisations.

A philosophy of algorithmic measurement must therefore analyse more than model performance. It must reconstruct the chain by which a construct becomes associated with data and by which a model output is subsequently interpreted as a statement about a person.

Let behaviour denote an action or omission performed by a person in a particular temporal, social, and organisational context. Answering a client, helping a colleague, reporting an error, remaining silent in a meeting, postponing a task, and refusing an instruction are behaviours. A stored email, timestamp, cursor trajectory, audio recording, or physiological signal is not the behaviour itself. It is a record generated when behaviour or bodily activity interacts with a recording system.

The distinction is important because recording systems are selective. An email server registers messages sent through email but not the conversation that made sending a message unnecessary. A collaboration platform records visible participation but not concentrated work performed offline. Keyboard monitoring registers keystrokes but not the quality of reasoning between them. A wearable device registers selected physical signals but not the full context in which those signals arose.

The recording system determines what becomes visible, at what resolution, with what omissions, and according to which technical categories. Records may consequently be evidentially useful without being epistemically transparent. Their meaning depends on how they were generated, what the system failed to record, how users adapted to the system, and whether the record remains interpretable outside its original operational context.

Digital data should not therefore be treated as theory-free fragments of organisational reality. Their scientific utility depends on provenance, curation, representation, and the purposes for which they can be mobilized [10]. The existence of a technically accessible record does not remove the need for theoretical mediation. Numerical records do not acquire evidential meaning independently of a prior epistemic context. Rigat [11] argues that measurements become data only in relation to a previously formulated hypothesis or research problem and that numerical observations cannot determine their own scientific interpretation. Digital traces should therefore not be treated as self-interpreting evidence. This distinction is consistent with Bogen and Woodward’s [12] separation of data from phenomena. A digital record should be treated as evidence from which a phenomenon may be inferred, rather than as the phenomenon itself. The stability and scientific relevance of the inferred phenomenon must be established across records, contexts, and recording conditions. Recent empirical evidence reinforces this point. A multitrait–multimethod comparison of survey and smartphone-trace measures found very low cross-method correlations for several behaviours and showed that neither source can be treated as a perfect gold standard [13]. Digital traces may sometimes exhibit higher measurement quality, but this does not establish that they capture the same phenomenon as self-report or that they are generically superior substitutes.

An algorithmic inference is a model output interpreted as information about a person, group, behaviour, state, or future event. A language model may assign a probability that a message has negative sentiment. A classifier may predict a high probability of turnover. A dashboard may display an engagement score of 0.74. None of these outputs directly observes engagement.

The output depends on the selection of data, preprocessing, feature extraction, model architecture, objective function, training labels, thresholds, and the rule connecting the result to a substantive assertion. The transition from digital record to measurement claim is often hidden by the visual authority of the score. This authority reflects a broader tendency to associate quantification with objectivity, impersonality, and procedural neutrality even when the underlying categories remain theoretically or normatively contestable [14]. Yet the score is the endpoint of several transformations, each of which can introduce conceptual and empirical error.

The process may be represented schematically. Let $P$ denote the person’s relevant state or profile, $K$ the context, $B$ behaviour, $S$ the recording system, $D$ the resulting digital record, $\varphi$ a representational procedure, $X$ the computational representation, $f$ the model, and $\widehat{Y}$ the model output:

$$\left(P,K\right)\to B\stackrel{r_S}{\to }D\stackrel{\varphi }{\to }X\stackrel{f}{\to }\widehat{Y}.$$

(1)

The left-hand side represents a data-generating process rather than a theory of the ontology of constructs. The variable $P$ is deliberately broader than the target construct $C$, because behaviour may be produced by many interacting personal and contextual conditions. The target construct enters only at the interpretive stage, when researchers propose that the model output supports an estimate $\widehat{C}$:

$$\widehat{Y}\text{⇝}\widehat{C}.$$

(2)

The symbol $⇝$ denotes an interpretive inference rather than a causal relation.

The data-generating and interpretive processes therefore run in opposite directions. Starting from $D$, $X$, or $\widehat{Y}$, researchers attempt to infer behaviour, a latent construct, or a future outcome. This inverse inference is not licensed merely because the generative relationship exists. Many combinations of context, behaviour, system design, and personal characteristics can produce similar records. The same long response time may result from low motivation, heavy workload, deliberate concentration, illness, time-zone differences, or a malfunctioning notification system.

Four inferential gaps can be distinguished within the pipeline.

The recording gap separates behaviour from its digital record. A record is incomplete and system-dependent. Bridging the gap requires evidence that the relevant event is registered reliably, that the absence of a record is interpretable, and that plausible alternative sources of the same record have been considered.

The representation gap separates the digital record from the computational representation used by the model. Text must be tokenised or embedded, messages aggregated, networks delimited, sensor streams filtered, and categories defined. Representation determines which similarities and differences become available to the algorithm. It can preserve, erase, amplify, or create distinctions.

The proxy-label gap separates the target construct from the criterion used to train or evaluate the model. In supervised learning, the target available to the model is usually not the construct itself but a proxy: a questionnaire score, manager rating, previous organisational decision, administrative category, or behavioural outcome.

The interpretation gap separates the model output from the substantive claim made about a person. A model can predict a questionnaire score, future resignation, or linguistic category without thereby measuring engagement. The interpretation gap concerns the transition from what the model demonstrably predicts to what researchers or organisational actors claim that the result means.

A digital trace supports a measurement claim only when each transition is justified. Performance at the final stage cannot repair an unjustified transition at an earlier stage. A highly accurate model operating on a conceptually inadequate representation or invalid label does not become a valid measurement instrument merely because the prediction error is small. This pipeline-level view is consistent with recent validity frameworks for supervised learning. Although developed for education research, Anglin’s [15] framework distinguishes construct, external, internal, and statistical-conclusion validity and requires an assessment extending beyond holdout accuracy. The present account extends that discussion by locating construct-related threats at successive inferential transitions and by adding normative entitlement as a separate condition of legitimate measurement practice.

The possibility of prediction without an articulated account of the measured phenomenon is part of a broader development in data-intensive science. Napoletani et al. [16] describe an “agnostic” orientation in which extensive datasets and powerful techniques are expected to support forecasting without a structured understanding of the phenomenon. Humphreys [17] similarly distinguishes the epistemic role of models from techniques that can generate results while offering limited theoretical interpretation.

This orientation can be productive when the objective is narrowly predictive and the outcome is clearly specified. A system may predict machine failure or customer demand without representing an underlying latent attribute. The orientation becomes problematic when predictive output is redescribed as measurement of a human construct. Predictive success then creates an appearance of semantic and ontological access that has not been established. The problem is not that data-driven techniques produce no knowledge, but that the kind of knowledge produced may be misidentified.

3. Predictive Accuracy, Construct Validity, and the Invasive Proxy Paradox

Consider an attempt to infer employee engagement from workplace communication. Let $E_T$denote engagement as theoretically understood, $X$ the features extracted from messages and metadata, and $Y$ a score obtained from an engagement questionnaire. The questionnaire can be represented as a measurement procedure $m$:

$$Y=m\left(E_T\right)+{\epsilon }_m,$$

(3)

where ${\epsilon }_m$ represents measurement error and other influences on the questionnaire score.

A supervised machine-learning model is then trained to approximate:

$$f:X\to Y.$$

(4)

The model is not trained on $E_T$. It is trained on $Y$. High out-of-sample accuracy establishes that the selected features contain information useful for reproducing the questionnaire score under the sampled conditions. It does not independently establish that the model measures engagement, captures the same dimensions of engagement across groups, or remains valid when organisational conditions change.

This point is not an objection to using self-report labels. A self-report can be an appropriate indicator, especially when the target includes subjective experience. The objection concerns the claim that an algorithm has escaped the epistemic limitations of self-report merely because it does not require employees to complete the questionnaire during deployment.

This relationship gives rise to the invasive proxy paradox:

A method introduced as a more objective substitute for self-report may be more intrusive while remaining epistemically dependent on the self-report labels it was designed to replace.

The term “paradox” is used here to denote a counterintuitive conjunction rather than a formal contradiction. The method is claimed to gain epistemic independence from self-report, yet its semantic target remains fixed by a self-report or institutional proxy while the scope of data collection becomes substantially broader. The alleged gain in epistemic independence is therefore smaller than advertised, while the increase in intrusion is real.

The paradox contains three elements. First, the digital method is presented as behaviourally objective because it relies on records rather than answers to questions. Second, it requires access to a substantially broader field of personal activity than the questionnaire. Third, its semantic target is inherited from the questionnaire used as the training label. The method may therefore increase surveillance without acquiring an independent epistemic route to the construct.

If the questionnaire contains measurement error, the model may learn patterns associated with that error. If responses are affected by social desirability, the model may learn correlates of socially desirable responding. If the scale functions differently across groups or occupations, the model may reproduce those differences. The removal of the questionnaire from the deployment stage does not remove it from the epistemic foundation of the system. An earlier ML-based application to OCAI-based organisational-culture diagnosis illustrates a legitimate but limited role for prediction. The method reduces respondent and analyst burden while retaining the OCAI typology as the target-defining framework [18]. This represents a gain in efficiency and utility, but it should not be redescribed as construct access independent of the questionnaire on which the categories depend.

The same problem arises when labels are manager ratings or historical organisational decisions. A model predicting prior promotion decisions learns regularities in those decisions; it does not thereby measure merit. A model predicting previous disciplinary outcomes does not necessarily measure misconduct. A model predicting absence does not automatically measure commitment, health, or motivation. Labels are institutional products with their own histories, biases, purposes, and error structures.

Unsupervised learning does not eliminate the problem. A cluster of employees with similar communication patterns does not become a group of disengaged employees by computational necessity. The semantic label is supplied after clustering by researchers or organisational actors. Without external validation, the cluster is a regularity within a selected representation rather than a discovered psychological type.

The same applies to large language models. A language model may generate a fluent description of a person’s communication, but fluency does not establish measurement. The output remains dependent on training data, prompt design, contextual assumptions, and the evidential relationship between linguistic patterns and the attributed state. The ability to produce an intelligible explanation may increase the persuasive force of an inference without increasing its construct validity.

Four properties should consequently be kept distinct.

Predictive accuracy concerns agreement with a selected criterion. Reliability concerns the consistency or stability of results under relevant repetitions. Construct validity concerns whether interpreting the result as evidence about the target construct is warranted. Utility concerns whether use of the result serves a practical or organisational objective.

A model can be accurate but invalid when the criterion is an inadequate representation of the construct. It can be reliable but invalid when it consistently measures the wrong phenomenon. It can be useful for predicting resignation while invalid as a measure of engagement. Conversely, a self-report measure can have limited predictive power for a specific outcome while providing relevant evidence about subjective experience. Recent hybrid SEM–ML work explicitly separates structural fit from predictive performance and combines both for measurement-scale evaluation [19]. This shows how machine learning can complement rather than replace psychometric validation. It does not, however, resolve the additional recording, proxy-label, interpretation, and entitlement questions that arise when the predictors are passively collected digital traces.

The following implication is therefore invalid:

Predictive performance(f)↑ ⇏ Construct validity(f)↑.

(5)

Validation requires an argument specifying the proposed interpretation and use of a result, the evidential steps supporting that interpretation, and the plausible alternatives that must be excluded [7]. In algorithmic measurement, this argument must cover the full pipeline rather than only cross-validation metrics.

This point also explains why the contrast between “subjective surveys” and “objective behavioural data” is misleading. Both are evidential procedures involving selection, representation, interpretation, and error. They differ in the kinds of access they provide and the types of error they introduce. A questionnaire may be influenced by memory and self-presentation. A digital trace may be influenced by system design, channel selection, workplace norms, surveillance awareness, and strategic adaptation.

The relevant methodological question is not which source is generically more objective. It is which evidential route is adequate for the specified construct, population, context, and use. Direct observation can be superior when the target is a clearly defined behaviour. If researchers wish to determine whether safety equipment is used, observation may be more appropriate than asking workers whether they use it. But engagement, stress, commitment, and satisfaction are not equivalent to single observable behaviours. The epistemic priority of behavioural data must be argued rather than assumed.

Algorithmic measurement may also create a form of circularity. An organisation can first define desirable conduct through a survey or managerial evaluation, then train a system to reproduce that judgement, and finally cite the system’s output as objective confirmation of the original definition. Computational mediation can obscure rather than eliminate the evaluative assumptions embedded in the label. The measurement–entitlement gap therefore begins with an epistemic distinction. The ability to reproduce a label does not by itself warrant the interpretation of a score as a measurement of the construct named by that label.

4. Ethical Provenance and the Dual-Warrant Framework

4.1. Ethical Provenance and Epistemic Reactivity

Technical access to data does not establish either scientific or moral entitlement to use them. An organisation may administer an email system without acquiring an unrestricted entitlement to infer the psychological states of every user. A researcher may be technically able to download publicly accessible data without thereby resolving questions of contextual expectation, consent, re-identification, or harm [20]. More generally, algorithmic systems can generate opacity, responsibility gaps, unfair outcomes, and difficulties in tracing morally significant effects to particular decisions within the inferential process [21].

At least four permissions should be distinguished: permission to record an event, permission to access the resulting record, permission to infer an attribute from the record, and permission to use the inference for a specified purpose. These permissions can come apart.

$$\begin{array}{c}\hfill \mathrm{A}\mathrm{c}\mathrm{c}\mathrm{e}\mathrm{s}\mathrm{s}\left(D\right)\hspace{0.33em}\nRightarrow \hspace{0.33em}\mathrm{E}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{t}\mathrm{l}\mathrm{e}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\left(\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{e}\mathrm{r}\left(C\mid D\right)\right)\\ \hfill \hspace{0.33em}\nRightarrow \hspace{0.33em}\mathrm{E}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{t}\mathrm{l}\mathrm{e}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\left(\mathrm{U}\mathrm{s}\mathrm{e}\left(\widehat{C}\right)\right).\end{array}$$

(6)

This distinction is related to Wachter and Mittelstadt’s [22] argument that data protection must address not only access to data but also the reasonableness of inferences drawn from them. The present account adds a measurement-theoretic question: even a normatively reasonable inference may fail to constitute a valid measurement of the attributed construct. Security requirements may justify recording attempts to access a system but not evaluating employee loyalty. Storing messages for operational continuity may not justify emotion recognition. Consent to a health programme may not justify using wearable data to determine remuneration.

Nissenbaum’s account of contextual integrity treats privacy as appropriate information flow rather than mere secrecy [23,24]. Information is generated within contexts structured by actors, roles, information types, recipients, purposes, and transmission principles. A violation can occur when information is redirected or transformed contrary to the norms governing its original context, even when the information was not secret.

Workplace communication illustrates the point. A message may be written to coordinate work with an intended recipient. Its storage on organisational infrastructure does not erase the original communicative context. Using the message to infer psychological engagement changes the type of information, the recipient, and the transmission principle. Operational communication is transformed into evidence in a psychometric assessment. The principles of purpose limitation and data minimisation provide a parallel regulatory expression of this concern: data should be collected for specified purposes and limited to what is necessary for those purposes [25].

Scientific data provenance is commonly understood in technical terms: where data originated, how they were transformed, and which version was analysed. For digital traces concerning persons, provenance must also be social and ethical. It should include the circumstances of participation, the informational norms of the setting, power relations, expected uses, excluded uses, and likely behavioural adaptations. These conditions affect what the data mean and which inferences they can support. Empirical case research on AI ethics in HR and people analytics likewise indicates that formal review boards and codes of conduct can coexist with underdeveloped data governance and unresolved responsibility for personnel data [26]. Ethical provenance therefore requires more than the presence of general AI-ethics procedures; it requires an explicit allocation of authority and responsibility at each inferential stage.

Employment relationships introduce a further difficulty. Informed consent is a central protection in research involving persons, but its normative force depends on voluntariness. Employees may reasonably believe that refusing participation signals disloyalty, reduces access to benefits, or attracts managerial attention. A formally expressed agreement may therefore fail to constitute genuinely voluntary authorisation. European data-protection guidance similarly recognises that the power imbalance between employers and employees normally makes freely given consent difficult to establish [27].

The problem is intensified by continuous data. Consent to a questionnaire is relatively bounded: participants can inspect the questions and understand the immediate act of disclosure. Consent to analyse an email archive, communication history, or continuous wearable stream is open-ended. Neither researcher nor participant may anticipate every sensitive fact, third-party disclosure, or future inference contained in the material. These limitations are not fully resolved by formal consent or de-identification, because data-intensive analysis may generate new information about participants and third parties that was not available or foreseeable at the point of collection [28,29].

Normative warrant does not always require individual consent. Certain studies can be justified by public interest, safety obligations, legal authority, or the impracticability of consent, provided that independent oversight and strong safeguards are present. The important point is that a legitimate basis must be argued. It cannot be inferred from technical access or organisational ownership.

Monitoring is also an intervention. Employees who know that messages are being analysed may avoid criticism, adopt artificially positive language, move communication to unmonitored channels, or produce activity that satisfies the metric. Bernstein’s [30] transparency paradox demonstrates that increased visibility can reduce genuine transparency by inducing concealment and adaptation. Electronic performance monitoring is associated with increased stress, while the available evidence does not establish a general improvement in performance [31]. Survey evidence from 962 gig workers shows that exposure to algorithmic compensation is positively associated with time-based stress, whereas perceived transparency strengthens procedural-justice perceptions but does not reduce the stress relationship [32]. Transparency may therefore improve one dimension of normative warrant without neutralising the behavioural and experiential consequences of the system.

Three forms of reactivity are especially important.

Observational reactivity occurs when people alter their conduct because they know they are being observed. Metric reactivity occurs when people optimise the measured indicator rather than the underlying objective. Classificatory performativity occurs when a score changes how a person is treated and thereby changes subsequent behaviour.

Measures, rankings, and categories can reconstruct the social worlds they purport merely to describe [33]. A monitoring practice that undermines trust changes the communication it analyses. A system connected to sanctions may create cautious or negative linguistic markers and then interpret them as evidence of low engagement. A reward linked to visible computer activity may produce cursor movement rather than valuable work.

Ethical provenance is therefore epistemically relevant. Normative deficiencies can change the data-generating process, the composition of the sample, the meaning of absence and silence, and the distribution of error. Ethics is not merely a constraint applied after scientific validity has been established.

This does not imply that evidence is whatever ethical values declare it to be. Truth does not depend on the moral quality of the method through which it is discovered. A normatively illegitimate practice may accidentally generate a true statement. The issue is whether a research practice produces a warranted scientific measurement and whether researchers are entitled to make and use that measurement claim.

The significance of values is particularly clear where uncertainty has morally consequential effects. Douglas [34] argues that the value-free ideal is inadequate when choices about evidential standards distribute risks of error. In algorithmic measurement, false positives and false negatives are not merely statistical outcomes when they affect employment, education, health, or reputation. Researchers must therefore specify who bears the costs of error, whether the result can be challenged, and whether the available evidence is sufficient for the intended use.

The same model can be acceptable for exploratory, aggregate-level research and unacceptable for an individual disciplinary decision. Validation and normative entitlement are use-specific.

4.2. The Dual-Warrant Framework

Let $c$ denote a particular measurement claim and $M$ the research practice through which that claim is produced and used. The epistemic status of the claim and the scientific legitimacy of the practice should be represented separately:

$$\begin{array}{c}\mathrm{EpistemicallyWarranted}\left(c,M\right)=E\left(c,M\right),\hfill \\ \mathrm{S}\mathrm{c}\mathrm{i}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{y}\mathrm{L}\mathrm{e}\mathrm{g}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{t}\mathrm{e}(c,M)=E(c,M)\wedge N\left(M\right)\hfill \end{array}$$

(7)

Here, $E(c,M)$ concerns whether the evidence generated by $M$ warrants interpreting $c$ as a claim about the target construct. $N\left(M\right)$ concerns whether the practice used to generate and employ that evidence is normatively admissible.

Epistemic warrant requires at least the following:

• Construct specification: the target construct is sufficiently defined and distinguished from its indicators and neighbouring concepts.
• Recording adequacy: the recording system captures relevant events with understood omissions and error mechanisms.
• Representation adequacy: the computational representation preserves distinctions relevant to the proposed interpretation.
• Label adequacy: the training and validation criterion is justified as an indicator or outcome appropriate to the claim.
• Model robustness: performance is assessed across relevant populations, contexts, time periods, and plausible distribution shifts.
• Inferential validity: the transition from output to measurement claim is supported by an explicit validation argument.
• Reactivity analysis: the possibility that observation, measurement, or classification changes behaviour and participation is considered.

Normative warrant requires:

• Legitimate purpose: the aim is sufficiently specific and important and is compatible with the rights and reasonable expectations of affected persons.
• Necessity: the purpose cannot be achieved by a substantially less intrusive method of adequate epistemic quality.
• Proportionality: the scope, duration, granularity, and consequences of data processing are proportionate to its expected scientific or social value.
• Legitimate authorisation: participation rests on voluntary consent or another defensible basis supported by appropriate oversight.
• Contextual integrity and purpose limitation: data flows and inferences are compatible with the context in which records were generated, or any departure is separately justified.
• Risk minimisation: aggregation, local processing, access restrictions, retention limits, and separation from administrative decisions are used where possible.
• Contestability and accountability: affected persons can challenge individual claims and identify responsibility for the practice.
• Protection of third parties: the rights and interests of recipients, colleagues, clients, and others represented in the data are taken into account.

The conjunction in the second expression should not be interpreted as making moral legitimacy a criterion of truth or construct validity. A normatively illegitimate practice may produce a true and even epistemically supported claim. The framework instead distinguishes the epistemic warrant of the claim from the legitimacy of the research practice as a whole. Because surveillance, coercion, and anticipated consequences can alter the data-generating process, however, deficiencies in $N\left(M\right)$ may also undermine $E(c,M)$.

Not every ethical cost is absolutely non-compensable. Legitimate research often imposes limited burdens. Yet a decision rule under which any ethical risk can be outweighed by sufficiently high predictive accuracy would be too permissive. Conversely, a rule stating that no epistemic benefit can ever justify any increase in burden would be too rigid.

A more defensible procedure has three stages. First, exclude practices that violate hard constraints, such as severe rights violations, disproportionate harm, or uses that affected persons could not reasonably authorise. Second, retain only practices that satisfy the epistemic and normative conditions specified above. Third, among the admissible methods that provide evidence of at least minimally sufficient epistemic quality, select the least intrusive method:

$$\begin{array}{c}\mathrm{Admissible}\left(M\right)\leftrightarrow E_M\wedge N\left(M\right)\wedge \neg H\left(M\right),\\ {\mathcal{A}}_Q=\left\{M:\mathrm{Admissible}\left(M\right)\wedge Q_E\left(M\right)\ge q_{\mathrm{m}\mathrm{i}\mathrm{n}}\right\},\\ M^{*}\in \mathrm{a}\mathrm{r}\mathrm{g}\underset{M\in {\mathcal{A}}_Q}{\mathrm{m}\mathrm{i}\mathrm{n}}I\left(M\right).\end{array}$$

(8)

Here, $H\left(M\right)$ denotes a hard-constraint violation, $E_M$ the satisfaction of the epistemic conditions applicable to the method, $Q_E\left(M\right)$ its epistemic quality, $q_{\mathrm{m}\mathrm{i}\mathrm{n}}$ the minimum quality adequate for the research purpose, and $I\left(M\right)$ the degree of intrusion. The expressions impose an ordering of decisions rather than assuming that normative warrant can be measured on a cardinal scale.

Kotarbiński’s praxiology provides a useful interpretation of this structure. Praxiology distinguishes the analysis of efficient action from the moral appraisal of ends and means [35]. This praxiological perspective is particularly relevant to management science. Previous analysis has characterised management science as combining nomothetic attempts to formulate general regularities with idiographic attention to context-specific organisational action, while retaining a praxiological orientation toward effective action [36]. Algorithmic measurement in organisations lies precisely at the intersection of these orientations: it seeks generalisable regularities while producing claims about particular persons embedded in specific organisational contexts. An algorithmic method may be effective, economical, and reliable in producing predictions while remaining unjustified as a scientific or organisational action.

This distinction supports a principle of minimally intrusive sufficiency:

Among the methods capable of producing evidence adequate for a specified research purpose, researchers should select the admissible method that interferes least with the autonomy and informational environment of participants.

The relevant objective is not the maximisation of data extraction. It is the acquisition of sufficiently warranted knowledge through a defensible course of action.

5. Employee Engagement as a Test Case

5.1. Comparing Methods of Engagement Measurement

Employee engagement is an appropriate test case because it is widely studied, normatively valued by organisations, and increasingly treated as a potential target for people analytics. Yet engagement is not equivalent to visible activity. Kahn [37] associated engagement with the physical, cognitive, and emotional employment of the self in a work role. The Utrecht Work Engagement Scale conceptualises engagement in terms of vigour, dedication, and absorption [38]. These dimensions include subjective experience and cannot be reduced without further argument to message volume, positive sentiment, centrality in a communication network, or time spent at a keyboard. Research on people analytics has already identified risks of objectification, asymmetrical power, loss of contextual meaning, and threats to personal integrity [39]. The dual-warrant framework explains why these problems are not merely downstream organisational harms but may also affect the evidential status of the resulting scores. Recent HRM reviews describe sentiment analysis, text embeddings, and unobtrusive sensing as tools for assessing employee engagement, morale, well-being, and stress [40]. These applications provide the immediate practical background for the distinctions developed here, because a system may support workforce analytics while remaining ambiguous as a measure of engagement.

The measurement question should therefore precede the data question. Researchers must ask whether they seek to measure subjective work experience, engagement-related behaviour, communication structure, effort allocation, or an organisational outcome such as retention. These targets are not equivalent. A method valid for one is not automatically valid for another.

A validated questionnaire directly solicits reports about work experience. It is vulnerable to response bias, interpretive differences, memory limitations, and common-method concerns. Nevertheless, when the target includes experienced vigour, dedication, and absorption, self-report has a distinctive evidential role. Its limitations do not make indirect surveillance a superior default.

An anonymous and voluntary questionnaire with aggregate reporting may satisfy normative warrant more readily than continuous monitoring. Its epistemic adequacy depends on the instrument, sampling process, administration, response conditions, and proposed interpretation. The appropriate comparison is not subjective error versus objective truth, but different evidential routes with different error structures and different burdens.

Communication metadata can provide information about network structure, response patterns, workload concentration, and organisational isolation. At the level of teams or communication processes, such analysis may support legitimate claims. It does not directly establish individual engagement. Network centrality may reflect formal role, administrative duties, dependence of others, or organisational bottlenecks. Low message volume may reflect withdrawal, but it may also reflect uninterrupted specialist work.

The normative assessment improves when message content is excluded, results are aggregated, minimum group sizes are imposed, data retention is limited, and individual scores are unavailable to managers. Even aggregate analysis requires attention to purpose limitation and re-identification, particularly in small teams.

Full-text analysis creates substantially greater difficulties. It permits sentiment classification, topic modelling, linguistic-style analysis, and semantic search, but also exposes information about conflicts, health, family circumstances, political views, personal relationships, and confidential organisational processes. It incorporates data about recipients and third parties who may not be participants in the research.

Sentiment is not engagement. Negative language can express critical commitment to organisational improvement. Positive language can result from politeness, conformity, fear, or impression management. A generic sentiment model may be especially unreliable in professional communication, where irony, technical language, hierarchy, and local communicative conventions affect meaning.

The relationship can be stated as follows:

$$\begin{array}{c}\mathrm{linguistic}\mathrm{sentiment}\ne \mathrm{emotional}\mathrm{state}\ne \mathrm{work}\mathrm{engagement}\\ \ne \mathrm{work}\mathrm{performance}.\end{array}$$

(9)

Moving from one category to the next requires a separate empirical and theoretical argument. An individual engagement score based on message content should therefore be subject to a strong presumption against use. Overcoming that presumption would require an important and legitimate scientific purpose, a defensible basis for processing all relevant communications, robust construct validation, independent oversight, protection of third parties, and technical separation from employment decisions.

Screen capture, keystroke counts, and cursor movement are even more distant from engagement. These techniques belong to the broader class of electronic performance monitoring practices analysed by Ravid et al. [41]. They register device interaction rather than effort, reasoning, dedication, or the value of work. Cognitively demanding activity can produce little visible input, whereas trivial or simulated activity can produce abundant events. Such monitoring also encourages metric reactivity: employees may optimise visible activity rather than substantive performance.

In this case, the epistemic and ethical objections reinforce one another. The putative measure is intrusive, conceptually distant from engagement, and likely to alter the behaviour it records. Its use as a measure of engagement is difficult to justify even if it predicts managerial ratings or short-term output.

Wearable devices illustrate a related boundary. They can support legitimate voluntary health research or safety interventions. Their ethical character changes when employers receive individual physiological or activity data and connect rewards to steps, sleep, heart rate, or exercise. Financial incentives may undermine voluntariness, disadvantage people with disabilities or care responsibilities, and extend organisational influence beyond working time. Such programmes also illustrate how apparently voluntary self-tracking technologies can become instruments of workplace discipline and organisational control [42].

Physical activity may be relevant to health, but it is not employee engagement. A wellness measure becomes an employment measure only through a conceptual and normative transition requiring separate justification.

5.2. Boundary Cases and Generalisation

Several objections help delimit the argument.

First, observing behaviour can be methodologically superior to asking about it when the target is a concrete behaviour and observation is reliable and proportionate. The argument does not establish a general priority of self-report. It rejects the general priority of digital traces. The appropriate method depends on the target construct and use.

Second, organisational ownership of infrastructure does not justify every analysis of the data generated through that infrastructure. Ownership may establish duties concerning security and continuity, but it does not establish unrestricted entitlement to psychological profiling.

Third, consent can provide substantial normative warrant when it is genuinely voluntary, refusal is invisible to supervisors, withdrawal is meaningful, and data collection is limited. Consent does not, however, validate the construct, eliminate third-party interests, or make an unnecessarily intrusive design scientifically responsible.

Fourth, anonymisation and aggregation reduce risk but do not eliminate it. Linguistic style and network position can be identifying, while small-team aggregates can indirectly reveal individuals. Moreover, a violation of contextual expectations may occur at the stage of analysis even when names are removed before publication. Privacy-preserving techniques can nevertheless change the assessment. Local feature extraction, deletion of source content, differential privacy, minimum group sizes, and strict separation from human-resource systems may transform an unacceptable individual-profiling proposal into a defensible group-level study.

Fifth, safety-critical and medical purposes may justify more intrusive measurement when the purpose is compelling, the relationship between the data and the relevant harm is well established, and strong safeguards are present. Fatigue detection for an operator of dangerous machinery differs from engagement scoring for performance management. The difference lies in purpose, necessity, evidential support, governance, and consequences rather than merely in the technology used.

The European Union’s Artificial Intelligence Act reflects a related regulatory judgement by prohibiting certain emotion-recognition uses in workplaces, subject to limited medical and safety exceptions [43]. The philosophical argument developed here is broader than the scope of this legal rule. A system can fall outside a statutory prohibition and remain epistemically invalid or normatively unjustified.

Finally, normative illegitimacy does not make every output false. The point is that scientific research involves more than the accidental production of true statements. A scientific measurement practice should provide a justified interpretation of evidence and a legitimate basis for generating and using that evidence.

The employee-engagement case shows that digital traces have no default epistemic priority over self-reports. Different methods provide access to different aspects of organisational life and have different error structures. Surveys, direct observation, experiments, administrative records, metadata, and digital traces should therefore be understood as potentially complementary sources.

Triangulation is preferable when sources possess partly independent limitations and when their collection remains necessary and proportionate. The goal is not to replace allegedly subjective methods with allegedly objective methods. It is to construct an evidential network in which each source is used only for the claims it can justifiably support.

The implications extend beyond the workplace. In learning analytics, click frequency, time-on-task, navigation paths, and submission patterns may be interpreted as indicators of attention, engagement, or ability. Yet similar records may be generated by confusion, accessibility requirements, strategic test preparation, technical interruption, or differences in learning style. The recording and interpretation gaps therefore remain, while the authority relationship between educational institutions and learners raises additional questions about authorisation and downstream use.

In digital phenotyping, smartphone mobility, typing, communication, or sleep-related patterns may be used to infer depression, stress, or relapse risk. The clinical importance of the purpose may strengthen normative justification, but it does not eliminate the need to validate the relationship between trace and construct, protect third parties, and restrict secondary use. An inferential system may be clinically useful without being a direct measurement of the psychological construct used to describe its output.

Public agencies and social platforms similarly derive risk, vulnerability, trustworthiness, and preference-related inferences from administrative and interaction data. These examples show that the framework is not specific to employment, although workplace power asymmetry provides an especially clear case of the interaction between epistemic and normative warrant.

In each case, the measurement–entitlement gap separates three questions:

• What can the system compute?
• What does the computation warrant researchers in claiming?
• What are researchers or institutions entitled to compute and use?

Keeping these questions distinct is essential to responsible data-intensive research on human beings.

6. Conclusions

Artificial intelligence expands the technical capacity to derive predictions about persons from records of their daily activities. This capacity creates an understandable temptation to regard digital traces as a methodological cure for the limitations of self-report. The cure is not straightforward.

A digital record is not the behaviour that generated it. Behaviour is not identical with a theoretical construct. A computational representation is not neutral, and a model output does not become a measurement claim until an interpretation has been justified. The algorithmic measurement pipeline contains recording, representation, proxy-label, and interpretation gaps. Predictive performance addresses only part of this structure.

The invasive proxy paradox explains why replacing surveys with digital traces can fail on its own terms. An algorithm may be more intrusive while remaining epistemically dependent on the questionnaire score, manager rating, administrative category, or institutional decision used as its label. It can reproduce a proxy with high accuracy without acquiring independent access to the construct that the proxy is said to represent.

A scientifically legitimate algorithmic measurement practice concerning persons therefore requires both epistemic warrant for its construct interpretation and normative warrant for the generation and use of its evidence. Epistemic warrant concerns whether the evidence supports the intended construct interpretation. Normative warrant concerns whether researchers are entitled to generate and use that evidence within the relevant context. These dimensions of warrant are partly interdependent because surveillance, power asymmetries, and anticipated consequences can change behaviour, participation, communication, and the evidential meaning of digital traces.

The practical implication is not that behaviour should never be observed or that digital methods are inherently unethical. It is that digital traces possess no automatic epistemic privilege and that technical feasibility creates no automatic entitlement. Researchers should define the construct and intended use before selecting data, justify each inferential transition, establish a legitimate basis for data use, evaluate reactivity, and prefer the least intrusive admissible method capable of producing sufficient evidence.

Scientific maturity in the age of artificial intelligence is demonstrated not only by the ability to extract increasing amounts of information. It is also demonstrated by the ability to recognise when an apparently informative inference is conceptually unwarranted, normatively unauthorised, or both.