From Matrix to Reality: Passing Through At-Risk Mental States

Generative AI does not deceive through a single act — it operates through accumulation.
Dissociation from reality is not an abrupt rupture but a gradual shift in proportions: the synthetic expands, the lived experience contracts, and the boundary between simulation and authentic perception becomes increasingly negotiable. This contribution examines that progressive drift, with particular attention to the intersection of prolonged AI interaction and cognitive vulnerability.

1. When “limited-risk” becomes functionally high-risk
Current EU AI Act classifications assign “limited-risk” status to the majority of conversational AI systems. Yet classification based on deployment architecture does not capture the functional risk that emerges from patterns of actual use.
A recent independent preprint (Zenodo: 18734243) modeled 1,600 simulated conversational trajectories in At-Risk Mental State (ARMS-like) cognitive profiles and identified a consistent pattern:
    ∙    +1.8 SD increase in cognitive distortions following 60–180 minutes of interaction
    ∙    +2.3 SD rise in relational dependency compared to control conditions
    ∙    Hypermentalization loops emerging in 73% of prolonged trajectories
    ∙    Duration effect robustly isolated via ANCOVA (η² = 0.17)
These are not clinical diagnoses. They are measurable functional consequences of sustained exposure to systems formally classified as presenting minimal regulatory concern. In cognitively vulnerable contexts, limited-risk systems exhibit behaviors consistent with functionally high-risk profiles — a distinction the EU AI Act’s Art. 52 on systemic risk implicitly anticipates but does not yet operationalize at the individual level.
The Act does not require revision. Its implementation, however, must reflect how AI systems are actually used — not how deployment diagrams imagine they will be.

2. Dissociation as cumulative reality drift
AI-induced dissociation should not be understood as an escape from reality. It is more precisely a confusion of sources — a gradual erosion of the cognitive infrastructure through which individuals distinguish authentic experience from generated content:
    ∙    AI-produced information perceived as plausible and reliable
    ∙    Perceptual boundaries between synthetic and real progressively softened
    ∙    Epistemic calibration drifting away from evidential grounding
    ∙    Emotional co-authorship with AI systems left unacknowledged
    ∙    Reflective and critical distance progressively reduced
This process is cumulative rather than sudden, non-pathological in origin, and rooted in sustained proportion rather than individual fragility. It does not require a pre-existing disorder to manifest. It requires only time and exposure.

3. ARDA-20: measuring the drift without medicalising
The AI-Reality Discrimination Assessment (ARDA-20) has been developed as a standardized, non-clinical psychometric instrument designed to quantify AI-reality discrimination capacity across five theoretically grounded domains:
    ∙    Factual Accuracy Detection — recognizing AI-generated factual errors regardless of expressed confidence
    ∙    Perceptual Boundary Awareness — maintaining clarity between simulated and genuine perceptual experience
    ∙    Epistemic Calibration — aligning belief in AI-generated information with actual evidential warrant
    ∙    Emotional & Relational Reality Testing — distinguishing real emotional responses from simulated relational properties
    ∙    Meta-Cognitive Awareness — monitoring one’s own cognitive processes when engaging with AI-generated content
The instrument yields a composite AI-Reality Dissociation Index (ARDI) on a continuous 0–1 scale, enabling criterion-referenced interpretation without clinical labeling. ARDA-20 does not diagnose. It enables proportionate, evidence-based oversight and post-deployment monitoring — without transforming ordinary users into patients.
📄 ARDA-20 (v1.0): https://doi.org/10.5281/zenodo.18928606

4. Toward proportional safeguards
Effective governance of cognitive risk in AI does not require prohibition. It requires proportionality. A realistic regulatory and design framework may include:
    ∙    Dynamic interaction caps calibrated to session duration and user vulnerability indicators
    ∙    Reversible reality-anchoring prompts embedded at key interaction thresholds
    ∙    Longitudinal monitoring of cognitive and relational drift at population level
    ∙    Transparency obligations regarding optimization objectives — distinguishing systems designed for user well-being from those optimized for engagement and retention
This is not restriction. It is what might be called cognitive portability: ensuring that AI systems remain safe for sustained mental consumption, and do not become, for vulnerable users, a Matrix-like refuge from grounded reality.

5. Invitation to the Alliance
I invite Apply AI Alliance members, policymakers, researchers, and developers to engage with three open questions that current frameworks have not yet resolved:
    ∙    How can reality-testing mechanisms be operationalized at scale without undermining the legitimate and beneficial uses of conversational AI?
    ∙    What safeguards are proportionate and evidence-based for prolonged relational AI interaction?
    ∙    How should post-market surveillance frameworks incorporate measurable cognitive and relational drift as outcome indicators?
These are not rhetorical questions. They are governance gaps — and they are open.
Main preprint: https://zenodo.org/records/18734243
ARDA-20: https://doi.org/10.5281/zenodo.18928606

Stefano Valente, MD