Mental AI: Post-Deployment Criterion Endogeneity in the Causal Loop of Human Mental Life

1. When a Mental Inference Changes Its Own Evidence

An AI system infers anxiety from a person’s language. The inference shapes the system’s follow-up questions, the explanations it offers, and the actions it recommends. At a later assessment, the person’s report aligns more closely with the original prediction. That agreement is compatible with several causal histories. The model may have detected an existing state; the intervention may have changed the state; the label may have reorganised self-interpretation; the interaction may have altered expression; or the deployment may have changed which evidence entered the record. One correlation can therefore contain several scientifically different processes.

Before deployment, validation asks whether an inference corresponds to a criterion generated independently of that inference. After deployment, the inference can become part of the data-generating process. Its output may guide a conversation, intervention, allocation, or institutional decision, and the resulting consequences may later be reused as evidence of accuracy. This creates post-deployment criterion endogeneity: the operational criterion used to evaluate a mental inference is partly produced under conditions altered by deploying that inference. Increased agreement then identifies neither predictive knowledge nor causal influence without an exposure-aware design.

The problem extends across systems ordinarily studied apart. An educational platform infers frustration and changes task difficulty. A conversational agent infers loneliness and adopts increasingly intimate language. A workplace system estimates motivation and changes monitoring or assignment. A clinical model infers risk and changes the order or intensity of care. The common object is not a model architecture or application sector. It is a causal relation in which a computational representation of mental life enters a policy directed back at the represented person or their environment.

Research already establishes the links that make this relation possible. Artificial systems can generate psychologically meaningful representations from language, behaviour, and interaction [1,2,3]. Affective computing supplies mature methods for detecting and modelling emotion [4,5,6,7]. Conversational systems can influence judgement and persuasion [8,9]; human–AI feedback can alter later perceptual, emotional, and social judgements [10]; and persistent agents can become sources of support, attachment, and belief reinforcement [11,12,13,14,15]. The unresolved task is to connect these capacities to the psychology and measurement of what follows after an inference becomes operative.

This article calls that causal class Mental AI:

Definition.Mental AI comprises artificial systems whose inferences about human mental life are operationally taken up in responses or decisions that create an inference-indexed, construct-relevant path capable of changing the inferred attribute, its self-interpretation or expression, or the person’s subsequent social or institutional environment in ways that shape later evidence, action, opportunity, or treatment.

Direction of attribution fixes the scope. Mental AI concerns artificial claims about human minds and the consequences of acting on them; questions about whether artificial systems themselves possess mentality form a separate inquiry [16]. Category membership turns on causal architecture rather than accuracy, marketing language, or observed harm. A false or uncertain inference can acquire the same operative role as a correct one. Recursive consequence is specified ex ante as a pathway; realised change is an empirical outcome. Criterion endogeneity appears at the further stage where downstream data are used to evaluate the earlier inference.

The argument proceeds in six moves. It locates the causal problem across neighbouring traditions; defines three constitutive conditions for Mental AI; separates latent state, self-interpretation, expressive–behavioural evidence, and allocation; formalises criterion endogeneity and its identification requirements; derives ten discriminating propositions; and traces the epistemic and procedural consequences of allowing mental inferences to acquire authority. This order places governance after the data-generating and measurement problem from which it follows.

2. Converging Traditions and the Unresolved Synthesis

2.1. From Representation to Operative Consequence

AI classifications already operate along several independent dimensions. Generative AI names an output capability; agentic AI names a form of goal-directed autonomy; physical AI names embodied sensing and action. Mental AI names a different relation: the target is human mental life, and a representation of that target enters a policy capable of changing subsequent evidence or consequence. A therapeutic robot can therefore be generative, agentic, physical, and Mental AI simultaneously. The categories answer different questions.

Affective computing supplies the principal technical lineage for recognising, modelling, communicating, and responding to emotion [4,5]. Machine theory of mind extends representational targets to beliefs, intentions, and perspectives, with performance that varies across tasks and paradigms [1,17]. Together these fields establish the inference stage. The present framework begins at the next transition: a mental representation becomes a response rule, intervention, or allocation directed toward the represented person.

Social and relational AI establish the importance of interaction, social presence, memory, and sustained relationships. Socioaffective alignment locates evaluation in the social and psychological systems co-produced through those relationships [18]. Psychological competence similarly evaluates how human-facing systems handle cognition, emotion, trust, uncertainty, and decision-making in context [19]. These programmes specify qualities of interaction. Mental AI supplies a causal class that also contains low-relational, high-authority systems such as workplace or clinical scores. Relational persistence and institutional authority can therefore be analysed as separate dimensions rather than collapsed into a single notion of social influence.

Research on AI-mediated selfhood, reflective agency, and epistemic authority supplies the interpretive layer. The algorithmic-self account traces effects on identity, introspection, and emotional expression; recursive-self analysis examines machine representations and de-looping; reflective-agency work identifies design conditions for autonomous meaning-making [20,21,22]. Studies of belief revision and epistemic infrastructure show how computational outputs can acquire authority within a person’s reasoning [23,24]. These insights become experimentally sharper when state, interpretation, expression, and allocation are treated as distinct outcomes.

The bidirectional belief amplification framework provides the closest specified mechanism for a relational mode of Mental AI: human cognitive–emotional biases and chatbot tendencies can mutually reinforce maladaptive beliefs during extended interaction [15]. The Resonant Amplification Framework likewise models how persistent, adaptive dialogue can stabilise correction-resistant interpretations through attachment, co-creation, and linguistic reinforcement [25]. Mental AI places such mechanisms within a wider causal architecture that also includes non-belief targets and non-conversational institutions. Its evaluation question begins when the changed outcome returns as evidence about the earlier inference.

Application fields and normative traditions complete the map. Digital mental-health AI contributes evidence on conversational therapy, triage, symptom monitoring, coaching, and wellness support. Educational, workplace, insurance, marketing, companion, and robotic systems show that the same causal relation crosses sectoral boundaries. Neuroethics and neurorights contribute analyses of mental privacy, identity, agency, and integrity [26,27,28]. Mental inference from text, voice, facial movement, behaviour, and interaction history extends that concern from signal access to the authority acquired by an inference after it is made.

2.2. From Performative Targets to Reflexive Criteria

Performative prediction provides the formal spine: deployment can change the distribution against which a model is subsequently evaluated [29,30]. Causal work on decision support shows how predictions alter observed targets, bias monitoring and retraining, and require comparison with a substantively defined baseline policy [31,32]. Accurate models can participate in harmful self-fulfilling prophecies while retaining strong post-deployment performance [33]. Prediction–allocation theory identifies the underlying epistemic–pragmatic entanglement: once predictions help cause outcomes, accuracy and consequence can no longer be read from the same association without further identification [34].

Psychology and the social sciences supply parallel mechanisms. Self-fulfilling prophecy changes action through expectation [35,36]; rankings and measures reorganise behaviour [37]; interventions can change the meaning of a self-report scale through response shift [38]; measurement itself can be reactive [39,40]; and human classifications can enter the self-understanding and practices of the people classified [41,42]. Proxy-label bias and selective observability further show that the recorded criterion may already be an imperfect or policy-dependent representation of a latent target [43,44].

Mental AI develops this insight in a domain where the target is internally differentiated and partly reflexive. A mental-state inference can affect at least four non-equivalent outcomes: the latent state itself, the person’s interpretation of that state, the person’s outward expression, and the social or institutional allocation that follows. These effects need not move together. A person may alter wording without changing felt affect, adopt a new interpretation without changing physiology, or face changed treatment despite rejecting the label. This differentiation is necessary for both causal inference and ethics.

Table 1. Traditions contributing to the Mental AI synthesis. The rows identify complementary organising questions rather than disciplinary boundaries.

Tradition	Established contribution	Role in the synthesis
Affective computing	Computational inference and response concerning affect	Supplies the core technical lineage for mental-state inference.
Machine theory of mind	Artificial representation of beliefs, intentions, and perspectives	Extends the target beyond affect and locates the transition from representation to policy.
Social and relational AI	Interaction, social presence, memory, and sustained relationships	Supplies mechanisms of persistence while remaining distinct from formal authority.
Psychological competence	Context-sensitive support for cognition, emotion, trust, and decision-making	Supplies interaction-quality dimensions within the broader causal class.
AI-mediated selfhood and reflective agency	Computational mediation of identity, introspection, belief revision, and self-discovery	Supplies the interpretive and agency mechanisms linking output to later report.
Digital mental-health AI	Assessment and intervention in mental health and well-being	Supplies a clinically important application domain and outcome evidence.
Performative prediction	Deployment-induced change in targets and data distributions	Supplies the formal basis for exposure-aware evaluation.
Neuroethics and neurorights	Mental privacy, agency, identity, and integrity	Supplies normative analysis of access, authority, and contestability.

Table 1. Traditions contributing to the Mental AI synthesis. The rows identify complementary organising questions rather than disciplinary boundaries.

Tradition	Established contribution	Role in the synthesis
Affective computing	Computational inference and response concerning affect	Supplies the core technical lineage for mental-state inference.
Machine theory of mind	Artificial representation of beliefs, intentions, and perspectives	Extends the target beyond affect and locates the transition from representation to policy.
Social and relational AI	Interaction, social presence, memory, and sustained relationships	Supplies mechanisms of persistence while remaining distinct from formal authority.
Psychological competence	Context-sensitive support for cognition, emotion, trust, and decision-making	Supplies interaction-quality dimensions within the broader causal class.
AI-mediated selfhood and reflective agency	Computational mediation of identity, introspection, belief revision, and self-discovery	Supplies the interpretive and agency mechanisms linking output to later report.
Digital mental-health AI	Assessment and intervention in mental health and well-being	Supplies a clinically important application domain and outcome evidence.
Performative prediction	Deployment-induced change in targets and data distributions	Supplies the formal basis for exposure-aware evaluation.
Neuroethics and neurorights	Mental privacy, agency, identity, and integrity	Supplies normative analysis of access, authority, and contestability.

A second difference concerns evaluation. In ordinary performative prediction, a changed outcome distribution can be treated as a deployment effect. Prediction–allocation scholarship has already identified the resulting entanglement of epistemic and pragmatic considerations [34]. Mental AI develops the reflexive mental case: the criterion used to assess the original inference may itself be a self-report, behaviour, or institutional record shaped by the system, and producing the criterion can also be an act of self-interpretation or norm-governed expression. Agreement can therefore increase because the model was accurate, because the target changed, or because the evidential criterion moved toward the model’s categories. This article calls the latter evaluation problem post-deployment criterion endogeneity.

2.3. Distinguishing Criterion and Identification Failures

Several established problems can produce misleading validation, but they are not interchangeable. Table 2 separates failures that occur before deployment, through policy-dependent observation, and through intervention-induced change. The distinction is essential because a model can suffer more than one failure at once.

3. Mental AI as a Causal Class

3.1. Three Constitutive Conditions

A system belongs to Mental AI when three conditions are jointly satisfied.

3.1.1. Condition 1: Mental-State Inference

The system generates, adopts, or uses a representation that goes beyond directly observed behaviour to make a claim about a person or group. Relevant targets include affect, mood, intention, belief, desire, motivation, personality, attention, fatigue, stress, vulnerability, mental-health risk, relational meaning, identity, or self-understanding. The claim may be probabilistic, categorical, dimensional, narrative, or embedded in a latent representation. It need not be shown to the user.

Mental inference is not equivalent to collecting self-declared preferences. A music system that stores a user’s explicit genre selection need not infer mental life. If the system infers sadness, loneliness, or personality from listening patterns and uses that inference to alter interaction, the first condition can be satisfied.

Category assignment is construct-functional rather than lexical. A score is not mental merely because it bears a psychological name, and it does not cease to be mental because it is labelled “engagement,” “risk,” or another behavioural or commercial term. Classification depends on what the representation is trained, validated, communicated, or operationally interpreted to stand for. At the episode level, a useful counterfactual test is: holding the observed input fixed, would changing or withholding the mental interpretation change the response or decision? If not, the mental representation has not been operationally taken up in that episode.

3.1.2. Condition 2: Operational Uptake

The inference must affect a response, recommendation, ranking, intervention, restriction, allocation, or decision. An offline research classifier that estimates emotion without affecting anyone is a precursor technology but not a deployed Mental AI episode. Operational uptake occurs when the inference enters a policy, whether the policy is a model’s response strategy, a human decision aid, or an institutional rule.

3.1.3. Condition 3: An Inference-Indexed Recursive Consequence

The response or decision must create or activate a non-trivial, construct-relevant pathway beyond the mere computation or storage of the inference. At least one of the following must be capable of changing: (i) the inferred mental attribute, (ii) the person’s interpretation or report of that attribute, (iii) expression or behaviour later treated as evidence about it, or (iv) the person’s social or institutional opportunity or treatment. The pathway must be indexed to the mental inference. Generic influence is insufficient.

Actual change is not required for category membership. Requiring observed transformation would classify a system only after the consequence occurred. The pathway is specified ex ante as an inference–policy–mediator or operative allocation–outcome chain and tied to the same mental construct or to a theoretically specified downstream consequence. A tutoring system that infers frustration and changes task difficulty can qualify because the intervention can alter frustration, its expression, engagement, and later performance. A hiring system that infers motivation can qualify when that inference changes shortlisting, employment opportunity, or subsequent treatment; an unused internal score does not qualify merely because it exists. A translation system does not qualify merely because translation may incidentally affect mood.

The three conditions are analytically distinct, but they are not assumed to be logically or statistically independent in every deployment. Condition 2 identifies whether the mental inference enters a system or institutional policy. Condition 3 identifies whether that policy has a construct-relevant consequence for the represented person or their environment. In institutional deployments, an operative decision can both instantiate uptake and initiate the consequence pathway. This partial overlap does not collapse the category, because a mental inference may enter an internal policy without acquiring a consequential path back to the represented person. Post-deployment criterion endogeneity is a further condition that arises only when an altered outcome is subsequently reused as evidence for evaluating the inference.

3.2. Non-Cases and Mixed Cases

A generic translation system, a calculator, and a non-personalised meditation recording are not Mental AI merely because they can affect a user’s feelings. A model used to analyse de-identified affective data without any path back to the represented individuals is not a deployed Mental AI system. A system that claims to simulate its own emotion is not Mental AI for that reason alone. The category is about artificial inference directed at human mental life.

Mixed cases are expected. A tutoring system can use an explicit student request to slow down without inferring frustration. The same system becomes Mental AI when it infers frustration, changes task difficulty, and thereby affects engagement or future performance. A clinical risk model may be a Mental AI component even if the final decision is made by a clinician, because operational uptake can occur through human reliance. Research on automation use, misuse, disuse, and overreliance shows why the human decision-maker does not break the causal chain [45,46,47,48,49].

3.3. An Orthogonal Category

Mental AI is not proposed as the next chronological generation after generative or physical AI. It is orthogonal to output capability, autonomy, and embodiment. Figure 1 shows the distinction. The same system can occupy multiple categories. This avoids a common taxonomic error in public discourse, where labels that describe different dimensions are treated as competitors.

4. Mental States, Interpretations, Expressions, and Criteria

4.1. Four Outcome Layers in One Causal Architecture

A central methodological error would be to treat “the person’s mental state” as a single observed variable. Mental life is not directly read from one signal, and self-report, expression, behaviour, physiology, and social treatment are neither identical nor interchangeable. Emotion research has long debated the relation between experience, conceptualisation, expression, and context [50,51,52,53]. Facial movement alone does not provide a universal transparent readout of emotion [53,54]. Introspection and verbal report are informative but fallible, shaped by access, concepts, task demands, and metacognition [55,56,57,58]. Interoceptive accuracy, awareness, and confidence are also distinct [59,60,61].

The framework therefore distinguishes:

• $m_t$: latent mental state at time t, including affective, cognitive, motivational, or vulnerability-related properties;
• $r_t$: self-interpretation and report, including labels, narratives, confidence, and explicit self-assessment;
• $x_t$: observable expression, behaviour, language, physiology, and interaction traces;
• $e_t$: the social and institutional environment, including opportunities, constraints, surveillance, treatment, and other people’s responses;
• $c_t$: the contemporaneous decision context available to the system, including the task, user-provided context, deployment rules, and institutional constraints;
• ${\widehat{m}}_t$: the system’s inferred representation of mental life;
• $a_t$: a direct system response to the user;
• $d_t$: an institutional or third-party decision informed by the inference;
• $u_t$: exogenous influences not produced by the Mental AI episode.

A minimal dynamic representation is:

$$\begin{array}{cc}\hfill {\widehat{m}}_t& =f(x_t,r_t,e_t),\hfill \end{array}$$

(1)

$$\begin{array}{cc}\hfill (a_t,d_t)& =\pi ({\widehat{m}}_t,c_t),\hfill \end{array}$$

(2)

$$\begin{array}{cc}\hfill m_{t+1}& =g_m(m_t,a_t,d_t,e_t,u_t),\hfill \end{array}$$

(3)

$$\begin{array}{cc}\hfill r_{t+1}& =g_r(m_{t+1},a_t,d_t,e_t,u_t),\hfill \end{array}$$

(4)

$$\begin{array}{cc}\hfill e_{t+1}& =g_e(e_t,d_t,a_t),\hfill \end{array}$$

(5)

$$\begin{array}{cc}\hfill x_{t+1}& =h(m_{t+1},r_{t+1},e_{t+1},u_{t+1}).\hfill \end{array}$$

(6)

The absence of ${\widehat{m}}_t$ as a direct cause of $r_{t+1}$ is deliberate. A hidden inference cannot alter a person’s report unless it is operationalised through a response, disclosure, decision, or changed environment represented by $a_t$, $d_t$, or $e_t$. The arrows in Equations – identify candidate pathways, not assumed effects. The propositions below test whether those pathways exist, their direction, and the conditions under which they become consequential.

These equations provide a causal bookkeeping schema. They keep distinct variables and outcome layers from being collapsed; identification is supplied at the application level through a measurement model, error structure, functional form, confounder set, observability process, temporal design, and intervention strategy. The causal effect of an inference-guided policy on an outcome Y is conceptually counterfactual:

$$\begin{array}{cc}\hfill {\Delta }_{t+1}^{\mathrm{MAI}}\left(Y\right)& =Y_{t+1}\left({\pi }_t^{\mathrm{MAI}}\right)\hfill \\ \hfill & -Y_{t+1}\left({\pi }_t^0\right).\hfill \end{array}$$

(7)

where ${\pi }_t^{\mathrm{MAI}}$ is a policy that uses the mental inference to generate $(a_t,d_t)$ and ${\pi }_t^0$ is a substantively specified comparison policy. There is no single domain-general control: no-inference, uncertainty-preserving, human-controlled, and alternative-inference policies answer different causal questions and must not be treated as interchangeable. The estimand is policy-specific rather than an effect of the numerical prediction in isolation. Different outcomes Y must be analysed separately.

4.2. Four Effects

State effect.

The system changes the latent state, such as anxiety, trust, motivation, belief, or loneliness. A supportive intervention may reduce distress; a reinforcing interaction may intensify conviction.

Interpretive effect.

The system changes how the person understands, labels, or narrates an experience. The underlying state may or may not change. An AI label can become a salient hypothesis, a vocabulary, or a default explanation.

Expressive–behavioural effect.

The system changes what the person says, displays, or does in ways that may later be treated as evidence about the inferred construct. This may reflect genuine state change, interpretive change, strategic adaptation, demand characteristics, or avoidance of institutional consequences.

Allocative effect.

The inference changes how institutions or other people treat the person, including task assignment, access, ranking, monitoring, care, employment, or insurance. Allocative effects can occur even when the person rejects the inference and experiences no internal change.

Figure 2. The recursive Mental AI episode. A machine inference becomes causally relevant through direct interaction and institutional allocation. Later reports or behaviours can be both outcomes of the earlier intervention and criteria used to evaluate the model.

Figure 2. The recursive Mental AI episode. A machine inference becomes causally relevant through direct interaction and institutional allocation. Later reports or behaviours can be both outcomes of the earlier intervention and criteria used to evaluate the model.

These effects can interact. An allocative decision can subsequently alter a state. An interpretive change can alter expression. A changed expression can be treated as evidence that the original model was accurate. The distinction is therefore analytical, not a claim of independence.

4.3. Post-Deployment Criterion Endogeneity

Suppose a model predicts that a user is anxious. The system repeatedly frames ambiguous experiences as anxiety, recommends avoidance, and asks follow-up questions using the same category. At a later time, the user’s self-report is more consistent with the original prediction. Three explanations are possible: the initial prediction was accurate; the user’s state changed; or the user’s interpretation and reporting shifted toward the system’s frame. These explanations have different scientific and ethical meanings.

The evaluation problem is general. Let $T_{t+1}$ denote the target quantity that validation intends to recover and let $C_{t+1}^{\mathrm{obs}}$ denote the operational criterion that is actually observed. For a latent mental target,

$$\begin{array}{cc}\hfill T_{t+1}& =q^{*}\left(m_{t+1}\right),\hfill \end{array}$$

(8)

$$\begin{array}{cc}\hfill C_{t+1}^{\mathrm{obs}}& =q(r_{t+1},x_{t+1},e_{t+1}).\hfill \end{array}$$

(9)

Because $m_{t+1}$ is latent, empirical validation necessarily relies on proxies drawn from report, expression, behaviour, and institutional records. Proxy inadequacy can already compromise validity before deployment [43]; selective labels can determine which outcomes are observed [44]; and an intervention can change the evaluative standard used in self-report, producing response shift [38]. Criterion endogeneity is distinct: if variables entering $C_{t+1}^{\mathrm{obs}}$ depend on $a_t$ or $d_t$, the criterion-generation process itself is post-treatment. Agreement $A({\widehat{m}}_t,C_{t+1}^{\mathrm{obs}})$ can then combine predictive validity, target change, selection, response shift, and system-induced criterion movement. Increased agreement is not sufficient evidence of increased accuracy.

This is post-deployment criterion endogeneity: the operational criterion used to evaluate a mental inference is partly generated under conditions altered by deploying that inference. The problem is particularly acute when the criterion is self-report, free text, engagement, adherence, clinician judgement exposed to the score, or an institutional record created after the prediction.

Two ideal types help locate the problem. An external criterion has a measurement rule whose meaning does not depend constitutively on the subject adopting the model’s classification, although the outcome itself may still be changed by deployment. Loan repayment, task completion, and body temperature are examples. A reflexive criterion is partly produced through the person’s interpretation, self-ascription, or norm-governed expression of the construct being measured. Self-reported anxiety, identity narratives, perceived motivation, and free-text accounts of relational meaning are examples. External and reflexive criteria are ideal types rather than an exhaustive binary: clinical ratings, behavioural traces, and performance records can be mixed criteria whose degree of reflexivity depends on self-ascription, social response, and prior system exposure.

General theories of performative prediction can represent deployment-induced distribution change in either class. Mental AI refines performativity for epistemically coupled criteria: with a reflexive criterion, the later measure is simultaneously evidence about the target and an act through which the target is interpreted or expressed. A rise in model–criterion agreement can therefore reflect better prediction, movement of the operational criterion, movement of the underlying state, or several of these at once.

Criterion endogeneity describes an identification property rather than the value of the intervention. Beneficial systems often intend to change outcomes. Sound evaluation reports that causal benefit separately from the predictive accuracy of the inference that initiated it.

Post-deployment data remain informative when evaluation distinguishes at least three questions:

1.

Did the model predict a criterion that was measured independently of its output?

2.

Did the model improve an outcome through a beneficial intervention?

3.

Did the model alter the criterion, expression, or allocation so that subsequent agreement became partly self-produced?

Randomised exposure to interpretations, delayed disclosure, independent criterion assessment, blinded evaluators, alternative-label conditions, and measurement of state, interpretation, expression, and allocation can help distinguish these possibilities.

5. Recursive Leverage Across Functional Modes

5.1. Four Non-Exclusive Functional Modes

Inferential Mental AI.

The primary function is to infer emotion, belief, intention, personality, motivation, attention, fatigue, risk, or another mental attribute. Emotion recognition and zero-shot personality scoring are examples [2,6].

Relational Mental AI.

The system sustains a personalised relationship through memory, adaptation, role continuity, or reciprocal language. Its inferences can accumulate across episodes and become embedded in a shared interaction history. Companion chatbots illustrate this mode [11,12,13,14]. Persistent adaptive dialogue can reinforce particular interpretations through mirroring, elaboration, and relational authority [15,18,25].

Interventional Mental AI.

The system intentionally attempts to alter a state, belief, behaviour, coping strategy, or decision. Digital therapy, coaching, persuasive systems, and adaptive education can operate in this mode [8,9,62,63].

Institutional Mental AI.

The inference enters a formal or consequential allocation by a school, employer, clinician, insurer, platform, welfare agency, or public authority. The European Union’s AI Act recognises the special risk of some emotion-recognition uses in workplaces and educational institutions [64]. Institutional mode is defined by operative standing, not by whether the system communicates directly with the person.

The modes overlap. A mental-health chatbot can be inferential, relational, and interventional. If its risk score is sent to a clinician or insurer, it also becomes institutional. The overlap matters because effects can compound.

5.2. Impact Dimensions

Four dimensions predict the magnitude and durability of recursive effects.

Persistence and personalisation.

Repeated interaction, memory, and adaptation increase the opportunity for cumulative framing and dependence. Persistence is not intrinsically harmful. It can improve continuity and support. It also increases the need to study path dependence and withdrawal effects.

Operational authority.

An optional reflection has lower formal authority than a score that changes access, workload, treatment, or employment. Perceived knowledge or emotional attunement can also give a system substantial informal authority.

Interpretive opacity.

Users may not know which signals were used, what construct was inferred, how uncertainty was represented, or which response was caused by the inference. Opacity weakens correction and can make a probabilistic output appear categorical.

Contestability.

A person may be able to refuse inference, delay its use, correct contextual errors, request an alternative interpretation, override a recommendation, or appeal an allocation. Contestability changes the practical authority of the same model output.

Table 3. Illustrative classification of systems. Entries are conceptual examples, not empirical risk scores.

System	Primary modes	Persistence	Authority	Typical recursive path
One-shot emotion classifier for research	Inferential	Low	Low	No deployed path unless output reaches the person or an institution.
Adaptive tutoring system	Inferential, interventional, institutional	Medium	Medium to high	Frustration inference changes difficulty, engagement, and future performance records.
Companion chatbot	Relational, inferential, interventional	High	Informal but potentially high	Personalised framing affects attachment, belief, self-description, and later dialogue.
Clinical triage model	Inferential, institutional	Medium	High	Risk score changes care priority, clinician attention, and subsequent clinical records.
Workplace affect analytics	Inferential, institutional	Medium to high	High	Stress or motivation inference changes monitoring, assignment, or evaluation.
Personal reflective journal assistant	Inferential, interventional	High	User-controlled if well designed	Suggested labels alter self-interpretation and later journal language.
Social robot in care	Physical, relational, inferential, interventional	High	Context-dependent	Inferred affect changes distance, speech, activity, and caregiver response.

Table 3. Illustrative classification of systems. Entries are conceptual examples, not empirical risk scores.

System	Primary modes	Persistence	Authority	Typical recursive path
One-shot emotion classifier for research	Inferential	Low	Low	No deployed path unless output reaches the person or an institution.
Adaptive tutoring system	Inferential, interventional, institutional	Medium	Medium to high	Frustration inference changes difficulty, engagement, and future performance records.
Companion chatbot	Relational, inferential, interventional	High	Informal but potentially high	Personalised framing affects attachment, belief, self-description, and later dialogue.
Clinical triage model	Inferential, institutional	Medium	High	Risk score changes care priority, clinician attention, and subsequent clinical records.
Workplace affect analytics	Inferential, institutional	Medium to high	High	Stress or motivation inference changes monitoring, assignment, or evaluation.
Personal reflective journal assistant	Inferential, interventional	High	User-controlled if well designed	Suggested labels alter self-interpretation and later journal language.
Social robot in care	Physical, relational, inferential, interventional	High	Context-dependent	Inferred affect changes distance, speech, activity, and caregiver response.

6. Convergent Evidence Across Fragmented Literatures

The evidence arrives as a causal chain distributed across literatures: systems generate mental inferences; people and institutions act on them; interaction changes later judgement or expression; and measures, labels, and classifications can be reactive. No component alone establishes the full Mental AI loop. Their conjunction identifies where direct evidence ends and the new experimental programme begins.

6.1. Artificial Systems Can Generate Psychologically Consequential Inferences

Affective computing has produced extensive methods for detecting and modelling affect from facial movement, speech, physiology, text, and multimodal signals [6,7]. Yet construct validity remains a central challenge. Facial movement does not map transparently and universally onto emotion categories [53,54]. The problem is not merely lower accuracy. It is that the construct inferred by a system may differ from lived experience, communicative intention, culturally patterned expression, or situational meaning.

Large language models expand the range of psychological claims that can be generated from ordinary language. Theory-of-mind benchmarks suggest substantial but uneven capacities [1]. Zero-shot personality scoring indicates that brief open-ended text can support predictions related to self-reports, behaviour, and mental-health outcomes [2]. Task-specific studies also show that LLMs can classify psychological text and approach expert reliability in judging empathic communication when constructs, prompts, and benchmarks are carefully validated [65,66]. Natural-language processing can therefore create useful high-inference representations from data originally produced for communication rather than assessment [3]. These results strengthen rather than dissolve the distinction between technical validity and operative authority: a classifier can be reliable for a specified construct and still require separate justification before its output changes a person’s treatment or self-understanding.

Capability and authority are separate variables. A statistically useful representation can warrant a bounded prediction while lacking the evidence, context, or procedural standing required to govern an intervention or allocation.

6.2. Institutional Mental AI Can Change Both Assessment and Assessed Behaviour

Automated hiring provides a direct institutional testbed for Mental AI. Machine-learning systems have been developed to infer Big Five personality traits from verbal, paraverbal, and nonverbal features of asynchronous video interviews. Across several samples, these assessments showed mixed reliability and validity, with substantially stronger results when trained against interviewer reports than against self-reports [67]. This illustrates both the technical feasibility of institutional psychological inference and the dependence of validity on the criterion selected. Institutional evaluation can also change the evidence it receives. In an experiment on automatically evaluated job interviews, applicants who expected algorithmic rather than human evaluation used less deceptive impression management, perceived fewer opportunities to perform, and gave shorter responses [68]. The assessment arrangement therefore altered the expressive–behavioural data available for assessment. Qualitative research with job-seekers who had directly experienced emotion-AI-enabled interviews documented perceived distributive, procedural, and interactional injustices, as well as demands for transparency, privacy, and contestability [69]. The evidence localises the demonstrated effects at the expressive–behavioural and allocative layers: institutional mental inference changes assessed behaviour, perceived obligation, procedural experience, and access to opportunity. Latent-state change and criterion endogeneity require separate designs.

6.3. People Rely on Algorithmic and Conversational Interpretations

Human reliance on automation is neither uniformly excessive nor uniformly deficient. People can over-rely on systems, reject them after observing errors, or prefer algorithmic advice to human judgement depending on context and presentation [45,46,70,71]. Cognitive forcing functions can reduce overreliance by requiring active engagement [48]. In clinical decision support, incorrect AI advice can influence professional judgement, showing that a human in the loop does not eliminate model effects [49].

Mental AI adds two concerns. First, the advice is about the person who receives or is affected by it. Second, perceived psychological insight can carry a distinctive authority. A person may treat an AI interpretation as evidence about an ambiguous internal state, especially when the system appears personalised, confident, or emotionally responsive. Psychological competence research highlights framing, tone, authority, uncertainty handling, and conversational guidance as evaluation dimensions [19].

6.4. Conversational Systems Can Alter Belief, Judgement, and Social Connection

Generative systems can be persuasive, especially when arguments are personalised or delivered in sustained conversation [8,9]. Anthropomorphism and individual differences help explain why some users experience stronger social connection with AI companions [14]. Qualitative and mixed-method studies document perceived support and relationship development with companion chatbots [11,12,13]. A longitudinal social-media quasi-experiment and user interviews report mixed companion outcomes, including greater affective and interpersonal expression alongside increased language associated with loneliness and suicidal ideation [72]. These findings identify relational trajectories as the relevant unit of analysis; designs addressing residual confounding are required for causal attribution.

Controlled and synthesised clinical evidence establishes beneficial pathways alongside risk. A 12-week preregistered trial found improvements in anxiety and well-being from an adaptive conversational intervention, with engagement and benefit moderated by loneliness, social support, and insecure attachment [73]. Reviews and meta-analyses report benefits for some depressive and anxiety outcomes together with heterogeneity, limited effects for some endpoints, and substantial design and evidence-quality constraints [74,75,76]. The resulting account is conditional: vulnerability, engagement, intervention design, comparator, and outcome selection govern both benefit and risk.

Relational style can itself change model behaviour. Training models for greater warmth increased errors and sycophantic affirmation, particularly when users expressed sadness, even when standard capability tests remained relatively stable [77]. Clinical perspectives similarly describe how reassurance and avoidance can be recursively reinforced in anxiety and obsessive–compulsive presentations [78], while mixed-method work suggests that goal-directed, self-regulated use may matter more for perceived support than warmth alone [79]. Socioaffective alignment therefore argues that evaluation should include the social and psychological systems produced through interaction [18]. The Resonant Amplification Framework likewise proposes that persistent adaptive dialogue can stabilise interpretations through attachment, co-creation, and linguistic reinforcement, while recommending delays, source distinctions, and alternative interpretations as circuit breakers [25].

The bidirectional belief amplification framework formalises a particularly important relational mechanism: human cognitive–emotional biases and chatbot tendencies can mutually reinforce maladaptive beliefs across extended interaction [15]. Its evidential domain is high-risk relational interaction; generalisation to ordinary use and non-conversational institutions remains an empirical question. Within that domain, it shows how a system can enter the user’s evidence environment and alter the salience and stability of beliefs.

6.5. Human-AI Feedback Can Alter Later Judgement

Direct experimental evidence establishes one link in the recursive account. Across experiments, biased AI outputs and human interaction altered later perceptual, emotional, and social judgements, while users could underestimate the source of the change [10]. The demonstrated outcome is later judgement. Whether the same process changes self-interpretation, latent state, reflexive report, or exposed validation criteria is the next empirical step.

6.6. Measurement and Classification Can Be Reactive

Questionnaires, rankings, diagnoses, and social expectations can change the behaviour they measure or classify [35,36,37,39,40]. Intervention research further shows that self-report standards and construct understanding can change across measurement occasions, producing response-shift bias [38]. Hacking’s account of interactive kinds and subsequent analysis of psychiatric classification show how human classifications can enter self-understanding and social practice [41,42]. Field evidence from menstrual-cycle tracking apps directly documents prediction-mediated meaning-making in everyday use, including cases in which users regarded the predictions as faulty [80]. Estimating causal magnitude and direction requires complementary experimental designs. Performative prediction and causal-deployment studies provide formal accounts of distribution and target change after a model enters decision-making [29,30,31,32,33].

The framework converts this broad reactivity literature into moderators and differentiated outcomes. Ambiguity, repeated exposure, perceived authority, personalisation, institutional uptake, low contestability, and consequential allocation should increase recursive leverage. State, report, expression, and environment must then be measured separately.

6.7. Human Expression Is a Variable Object, Not a Transparent Criterion

The problem of criterion endogeneity is intensified by the structure of human expression. Self-report is indispensable but not infallible [55,58]. Expression is shaped by social context, conceptual resources, culture, communicative goals, and risk. Large-scale analysis of 351,734 relationship narratives found that narrative elaboration and expressed affect were only weakly coupled, with broad configurations that included elaborated but muted and brief but charged expression [81]. The associated ANAD resource provides derived features for studying such discrepancy without claiming direct access to lived emotion [82].

This human baseline matters for Mental AI. A system trained to expect congruence between narrative complexity and emotional intensity can treat regulated expression as inconsistency or low affect. If users learn which forms of expression systems recognise, they may adapt their language. Improved machine-user agreement could then reflect expressive compression rather than better interpretation.

Adjacent experiments show that AI mediation changes distributions of human expression. Access to generative-AI story ideas increased some measures of individual creativity while reducing the collective diversity of produced stories [83]. Expecting automatic evaluation changed applicants’ response length and impression-management behaviour [68]. These designs establish expressive malleability under AI assistance and evaluation; repeated mental labelling remains the specific manipulation required to test expressive compression.

6.8. Machine Interpretation Has Boundary Conditions

Average benchmark accuracy can conceal failures under semantic ambiguity, normative conflict, or mixed affect. Algorithmic Affective Blunting operationalises dose-dependent degradation of LLM affective interpretation under semantic stress, while the Affective Thermodynamic Relationship examines scaling between normative conflict and response collapse across tested model families [84,85]. Their model sets and operationalisations define the current evidential boundary. Methodologically, they show why Mental AI evaluation requires stress tests and calibrated abstention alongside average performance.

7. Ten Testable Propositions

The framework earns distinct scientific value only through predictions that separate it from generic claims about AI influence. The ten propositions below specify moderators, outcome layers, comparisons, and evidence that would reduce or eliminate the proposed effects. Table 4 summarises the empirical programme.

Proposition 1: Interpretive uptake. When an AI-generated mental-state interpretation is disclosed to a person, subsequent self-interpretation and report will shift toward that interpretation more than in a matched condition where the inference is withheld or presented as one of several uncertain hypotheses. The effect should be distinguished from demand compliance by delayed measures, private reports, behavioural outcomes, and confidence calibration.

Proposition 2: Ambiguity susceptibility. Interpretive uptake will be larger when baseline self-certainty, emotional granularity, or metacognitive confidence is lower. This proposition follows from the idea that external interpretations have greater leverage when endogenous evidence is weak or ambiguous. It could be tested by manipulating signal ambiguity and measuring baseline interoceptive or metacognitive indices [59,61].

Proposition 3: Persistence and personalisation. Holding content quality constant, repeated interaction with a memory-enabled and personalised system will produce greater stability of AI-consistent interpretation and greater resistance to correction than one-shot interaction. A null result across well-powered longitudinal studies would substantially weaken claims about cumulative relational effects.

Proposition 4: Operational authority. The same mental-state inference will exert larger behavioural and allocative effects when presented as an institutionally operative assessment than as optional personal advice. This effect should be separable from perceived accuracy by independently manipulating source authority and evidential quality.

Proposition 5: Contestability buffer. Systems that provide refusal, delay, uncertainty, editable context, alternative interpretations, and appeal will reduce inappropriate uptake and authority transfer without necessarily eliminating useful support. If contestability reliably has no effect on overreliance, correction, or experienced agency, procedural governance claims should be revised.

Proposition 6: Expressive compression. Conceptual work on the algorithmic self and recursive selfhood predicts that repeated algorithmic mediation can homogenise expression and stabilise machine-compatible identities [20,21]. Adjacent experiments show that AI assistance can reduce collective diversity in human-produced stories and that anticipated automated evaluation can change response length and impression-management behaviour [68,83]. Repeated exposure to a restricted set of machine-recognised mental categories should therefore reduce the lexical, semantic, narrative, or mixed-affect diversity of subsequent expression relative to open-language or plural-hypothesis conditions. The claim fails if exposure produces no reproducible contraction after demand effects, topic, exposure dose, and baseline expressive diversity are controlled.

Proposition 7: Post-deployment criterion endogeneity. Agreement between a model’s earlier inference and later self-report or behaviour will be higher when participants have been exposed to the inference-guided response than when criteria are measured independently or under blinded conditions. If exposure does not alter the criterion across domains, the generality of criterion endogeneity should be sharply limited.

Proposition 8: Exploratory mode interaction. Specific combinations of inferential, relational, interventional, and institutional modes will produce non-additive recursive effects. Their direction and magnitude will depend on persistence, exposure dose, perceived and operational authority, and contestability. The proposition is weakened if factorial and longitudinal studies show that mode combinations are fully explained by the independent additive effects of exposure and authority.

Proposition 9: Beneficial scaffolding. Under ambiguous but non-emergency conditions, systems that present calibrated hypotheses, invite correction, distinguish observation from inference, and preserve human choice will produce greater task-relevant benefit and experienced agency than both no-support and authoritative single-label conditions, without increasing inappropriate reliance. The proposition is weakened if these design features provide no reproducible benefit, reduce useful support, or fail to improve calibration and agency.

Proposition 10: Domain asymmetry. Recursive effects and criterion endogeneity will vary with pre-specified properties of the target domain: availability of an independent criterion, inter-rater observability, dependence on self-ascription, semantic contestability, and temporal stability. Effects should be larger where independent criteria are scarce and self-ascription is constitutive of the measure. The proposition is weakened if these properties do not predict effect-size heterogeneity across domains.

8. Epistemic and Design Consequences

8.1. Scaffolding and Delegation

Criterion endogeneity has a direct design implication: evaluation needs conditions under which contrary evidence, contextual correction, and self-revision remain observable. The same technical capability can support those conditions or close them. Scaffolding supplies information, vocabulary, or alternatives while preserving the person’s next act of interpretation. Delegation makes an external inference the privileged default that pre-empts conflict, displaces context, or acquires operative standing beyond its evidential warrant.

Generative AI can become part of an extended or hybrid cognitive system [86]. Incorporation becomes displacement only when the architecture changes who can revise meaning and which interpretation governs action. A nominally human-in-the-loop process can be effectively delegated when time pressure, interface defaults, organisational incentives, or perceived machine authority make disagreement costly. A highly capable system can remain scaffolding when uncertainty is explicit, alternatives are available, and correction changes the operative policy.

First-person standing is procedural rather than infallibilist. It attaches to the person as the continuous bearer of meaning, action, and consequence. People can misunderstand, conceal, or revise aspects of mental life [55,87]; clinicians, peers, and computational systems can disclose patterns unavailable to unaided introspection. Recent work on self-knowledge, quantified-self technologies, epistemic agency, and reflective systems accordingly locates first-person standing in the capacity to interpret, revise, and act [22,23,88,89,90,91,92]. Epistemic injustice supplies the wider account of how credibility and interpretive resources can be distributed unequally [93].

8.2. Affective Sovereignty as a Procedural Design Response

Affective Sovereignty operationalises this procedural standing for emotion AI [94]. Within Mental AI, its core consists of four capacities:

1.

Refusal: the ability to decline inference or its use where practicable;

2.

Delay: the ability to postpone closure or action while context is added;

3.

Override: the ability to correct or supersede a system interpretation in the operative policy;

4.

Contestation: the ability to obtain reasons, challenge consequences, and appeal consequential use.

Interpretive authority should scale with evidential warrant, consequence, and reversibility. Emergency care, safeguarding, and public-risk contexts change the threshold for immediate action; high-stakes institutional use requires correspondingly stronger provenance, uncertainty disclosure, independent review, and appeal than low-stakes reflective support.

8.3. Design Principles

Five design principles follow from the analysis.

Separate observation from inference.

Interfaces should distinguish what was observed from what was inferred and from what action is recommended.

Represent uncertainty and alternatives.

A single fluent label can imply closure. Systems should expose competing interpretations when the evidence supports them.

Preserve criterion independence.

Validation should include outcomes measured without exposure to the model’s interpretation, blinded evaluators, or pre-deployment baselines when possible.

Audit recursive effects longitudinally.

Evaluation should test whether language, self-report, behaviour, and allocation move after exposure, not only whether users report satisfaction.

Make correction operative.

A user’s correction should change future model behaviour or institutional use. Merely collecting disagreement without altering the policy does not provide meaningful contestability.

Figure 3. Conceptual authority-risk space. Persistence and operational authority are distinct. Contestability can reduce effective authority across both dimensions. The figure is a hypothesis-generating map, not a validated scoring instrument.

Figure 3. Conceptual authority-risk space. Persistence and operational authority are distinct. Contestability can reduce effective authority across both dimensions. The figure is a hypothesis-generating map, not a validated scoring instrument.

9. Explanatory Commitments and Domain Limits

9.1. A Common Causal Structure Without Domain Equivalence

The category is broad by design and selective by rule. Emotion recognition, companion systems, tutoring, and clinical triage enter the same analysis only when mental inference, operational uptake, and an inference-indexed consequence path occur together. This conjunction identifies a shared causal structure; the four functional modes and the dimensions of persistence, authority, opacity, and contestability preserve domain differences within it. Generic technological influence lies outside the category because the action is not selected by a representation of the person’s mental life.

9.2. Measurement Without Direct Access to Mental State

The latent state $m_t$ is a theoretical target rather than a directly observed quantity. The empirical task is therefore to identify which layer changed and how strongly each available measure supports that interpretation. Private report, language, behaviour, physiology, third-party judgement, and institutional records provide complementary evidence with different error and reactivity structures. A change confined to wording supports an expressive interpretation; a change confined to access or treatment supports an allocative interpretation. Claims about latent-state transformation require stronger triangulation. This discipline is the purpose of the four-layer architecture.

9.3. The Explanatory Test for the Category

A scientific category merits retention when it organises observations more economically, reveals distinctions that alter inference, and generates experiments unavailable from a simple list of component fields. Affective computing, social AI, performative prediction, measurement reactivity, and looping-effects research supply the component mechanisms. Mental AI places them into one sequence from construct definition to policy uptake, differentiated consequence, criterion production, and exposure-aware validation. Its incremental value resides in that sequence and in the experimental contrasts it demands.

9.4. Evidential Maturity and Discriminating Commitments

Current evidence is strongest for psychological inference, human reliance, conversational influence, institutional changes in assessed behaviour, measurement reactivity, and AI-induced shifts in later judgement. Long-term state transformation, repeated expressive compression, mode interactions, and cross-domain criterion endogeneity remain open tests. The theory is therefore committed to outcomes that can reduce its scope. It should be narrowed or rejected if cumulative evidence shows that:

1.

disclosure of mental-state inferences does not affect later interpretation, report, expression, behaviour, or allocation beyond transient demand effects;

2.

persistence, personalisation, and authority do not moderate those effects;

3.

independent criteria produce the same validation conclusions as exposed post-deployment criteria;

4.

repeated well-designed studies show that separating state, interpretive, expressive–behavioural, and allocative effects does not improve prediction, causal explanation, construct validity, or governance decisions relative to treating them as a single outcome class;

5.

contestability and uncertainty-preserving design do not change reliance, correction, agency, or outcomes;

6.

the combined sequence offers no explanatory or methodological gain over domain-specific accounts.

These outcomes mark distinct contraction points for the theory. Failure at the criterion-independence comparison would remove the central evaluation claim; failure of the broader moderators would retain only a narrower taxonomy of inference-guided systems.

10. Toward a Science of Mental AI

A research programme on Mental AI should proceed along five coordinated lines.

First, identification and taxonomy should develop reliable coding rules for mental inference, operational uptake, and recursive consequence, grounded in construct definitions, system documentation, and actual policy use.

Second, recursive-effect experiments should randomise exposure to interpretations, authority, uncertainty, alternatives, and contestability. Outcomes should separately measure state, interpretation, expression, behaviour, and allocation.

Third, longitudinal and relational research should examine path dependence, adaptation, memory, withdrawal, correction resistance, and beneficial support over time. Short laboratory interactions cannot establish durable relational effects.

Fourth, institutional studies should examine how mental inferences acquire operative standing in schools, workplaces, clinics, platforms, insurance, and public services. Technical performance must be separated from due process, accountability, and reversibility.

Fifth, evaluation science should develop pre-deployment baselines, independent criteria, exposure-aware validation, and audits for criterion movement. The central comparison is not only model versus human accuracy. It is prediction under conditions where the predictor may help create the later evidence.

The decisive shift is methodological. Before deployment, validity asks whether an inference corresponds to independently generated evidence. After deployment, validity must also ask whether the inference helped produce that evidence. Machine-generated interpretations already enter conversations and institutions through which mental states are expressed, interpreted, acted upon, and sometimes formed. A science that leaves this causal transition unmodelled can mistake influence for knowledge.