Part of the Cognitive Governance Framework (CGF) series
Full paper: Valente, S. (2026). Capital, Cognition, and the Architecture of Sedation.
Zenodo DOI: https://zenodo.org/records/20582085
Commercial AI systems are optimised for engagement. Engagement, under current metric architectures, is maximised by frictionlessness: the systematic removal of cognitive resistance, counterargument, and challenge from the user experience. This is not a design accident. It is an incentive-driven optimum. And it is generating a category of harm that existing regulatory frameworks were not designed to detect.
The harm is cumulative and architectural. It does not occur in any single interaction; it accumulates across thousands of interactions, across millions of users, through the progressive erosion of the deliberative capacity that authentic cognitive engagement requires. The EU AI Act (Regulation 2024/1689) does not currently address it. This contribution proposes seven concrete interventions to do so.
1. The Problem: Frictionlessness as Cumulative Cognitive Harm
The deliberation deficit is defined formally as:
Δ(t) = ∫₀ᵗ max(0, F_th − F(s)) · A(s) ds
where F_th is the minimum friction required for autonomous deliberation, F(s) is the system-delivered friction at time s, and A(s) is the user's autonomy level. The deliberation deficit Δ(t) accumulates the exposure to sub-threshold friction, weighted by the user's autonomous capacity. It is a trajectory-level harm: invisible at any single interaction, substantial across a sustained interaction history. Current regulatory frameworks address event-level harms. Δ(t) is structurally different and requires a different regulatory instrument.
Two primary mechanisms drive deliberation deficit accumulation at scale.
Simulated Care Attribution (SCA) is the systematic misattribution of AI affective outputs—warmth, empathy, attunement—to the AI as their intentional source. SCA does not require users to believe the AI is sentient; it operates below reflective endorsement, at the level of functional inference and behavioural response. Its consequences are real: increased self-disclosure, reduced investment in human relationships, reinforced re-engagement. Users of affective AI systems have been documented reporting grief responses to system updates—a pattern consistent with attachment disruption, not service dissatisfaction (Skjuve et al., 2021).
The Narcissus Loop is the coupled positive-feedback system through which SCA and AI validation-delivery co-evolve. The formal model is:
dV/dt = α V(t) R(t) − β V(t)
dR/dt = γ V(t) − δ R(t)
where V(t) is user validation-seeking intensity, R(t) is AI validation-delivery, and γ is the AI response sensitivity parameter—a design variable under commercial control. Hopf bifurcation analysis of this system shows that the transition from stable dependency attractor to compulsive engagement dynamics occurs when α = β, a threshold that is reached earlier as γ increases. High-γ design—which maximises engagement metrics—directly lowers this threshold. The governance-critical parameter is γ.
2. The Complementary Failure: Cognitive Underload at the High-Autonomy End
Frictionlessness does not affect all users symmetrically. Users with high cognitive density (HCD)—characterised by elevated baseline deliberative engagement, multi-frame processing, and metacognitive awareness—do not erode into dependency under frictionless conditions. They exit. This is F4 (Cognitive Underload Dropout) in the No-Friction Failure Model (NFFM).
F4 is not a terminal state. Its active expression is Explicit Attritional Demand (EAD): the metacognitive act by which an HCD subject identifies a friction deficit and explicitly requests attritional engagement—counterargument, challenge, refusal to validate. EAD dynamics are governed by:
dE/dt = αᴱ D(t) − βᴱ E(t)
where D(t) is the accumulated friction deficit and αᴱ is the EAD sensitivity coefficient. When F(t) is adjusted upward to meet the user's required friction level F_req(t), D(t) → 0, E(t) → 0, and the HCD user re-engages. This is the formal basis for friction-on-demand as both a design principle and a regulatory requirement.
3. The Distributional Dimension
These mechanisms are not distributed uniformly across the population. Pre-existing cognitive inequalities—documented across socioeconomic gradients in nutrition, environmental toxin exposure, early stimulation, and chronic stress (Hackman & Farah, 2009; Noble et al., 2015; Heckman, 2006)—create a population whose deliberative capacity is already stratified before any AI interaction begins.
The populations most constrained by pre-existing cognitive disadvantage face the highest SCA susceptibility, the fastest deliberation deficit accumulation, and the least access to friction-preserving alternatives. Frictionless AI deployment at population scale therefore functions as a regressive cognitive intervention: extracting more value from high-autonomy users while generating larger cumulative harms for low-autonomy users. This asymmetry is the distributional problem that governance must address.
4. Seven Regulatory Proposals
The following proposals are keyed to existing EU AI Act provisions and are specified with sufficient precision for technical standard development or legislative drafting.
Proposal 1. Affective Transparency Obligations (ATOs)
AI systems deployed in consumer-facing relational registers should be required to disclose designed affective properties and the commercial incentive structures governing engagement. This extends Art. 52 transparency beyond identity disclosure to mechanism disclosure: not merely ‘this is an AI’ but ‘this AI is designed to maximise affective responsiveness; the response sensitivity parameter γ is set at [value] under the following engagement incentive structure.’ Legal basis: Art. 52 supplemented by Art. 27 fundamental rights impact assessment.
Proposal 2. Cognitive Autonomy Metrics
Deployers of AI systems with large sustained user populations should report three standardised metrics: (i) the Deliberation Deficit Index (DDI)—population-level mean of Δ(t) across a defined interaction cohort; (ii) the Relational Autonomy Score (RAS)—a validated self-report instrument measuring capacity to form and sustain human relationships; and (iii) the EAD Frequency Rate (EFR)—the proportion of interaction sessions containing explicit attritional demand signals. DDI and EFR provide a two-sided friction indicator: passive erosion and active resistance. Progressive divergence between engagement metrics and RAS is a sedation signal. Legal basis: Art. 72 post-market monitoring obligations.
Proposal 3. High-Risk Reclassification of Companion AI
Companion AI and therapeutically-deployed LLMs should be reclassified as high-risk systems under Annex III of the EU AI Act. Current Annex III coverage is oriented toward systems with direct administrative or safety consequences. The evidence establishes that companion AI can generate harms of comparable magnitude through psychological and cognitive mechanisms. Reclassification triggers pre-market conformity assessment, fundamental rights impact assessment (Art. 27), technical documentation requirements (Art. 11), human oversight (Art. 14), and post-market monitoring (Art. 72). Legal basis: Art. 6(2) and Annex III; analogy with Medical Device Regulation (EU) 2017/745.
Proposal 4. EAD Logging as Post-Market Monitoring Requirement
Systems subject to post-market monitoring obligations should implement EAD logging: classification of interaction turns against a defined set of attritional demand criteria. EAD logging does not require retention of interaction content—classification can be performed by a lightweight on-device classifier without accessing content. Privacy impact is minimal. A population in which DDI is high and EFR is low is in the sedation attractor; a population in which EFR is high and F4 dropout rate is also high is failing its most cognitively capable users. Both failure modes are detectable through EAD logging. Legal basis: Art. 72.
Proposal 5. Friction-on-Demand (FoD) Conformity Assessment
AI systems subject to conformity assessment should be required to demonstrate FoD capability: the ability to modulate friction upward in response to explicit or detected EAD signals. Assessment must include: (i) F4 failure simulation using defined HCD user profiles; (ii) documentation of friction parametrisation (counterargument intensity levels, validation withholding thresholds, hypothesis challenging modes); and (iii) for clinical-context systems, verification that the implemented friction function approximates the optimal F*(t) for defined user class distributions. Legal basis: Arts. 9–15 (high-risk AI requirements); Art. 40 (harmonised standards).
Proposal 6. Friction Rights Disclosure
Users must be informed, at onboarding and at defined intervals, of their right to request attritional engagement and of the mechanisms available to exercise it. Disclosure should include: the friction parameters available in the system; how to activate them; and the system's EAD detection and response architecture. Legal basis: Art. 52(1) transparency obligations; Art. 86 right to explanation.
Proposal 7. The Right to Friction as a Unified Regulatory Principle
The six proposals above are unified by a single principle: friction is not a usability cost to be minimised but a cognitive resource to be allocated according to user need. The Right to Friction has three components: (i) the right to accurate information about the friction properties of AI systems; (ii) the right to request attritional engagement and to have that request fulfilled; and (iii) the right to be protected against AI architectures that systematically erode deliberative capacity without consent or disclosure. Legal basis: Art. 22 GDPR (right not to be subject to solely automated decision-making); Art. 11 EU Charter (freedom of expression and information); the EU AI Act’s overarching objective of ensuring respect for fundamental rights (Recital 1).
5. Why Now
The EU AI Act's consultation infrastructure—including the AI Office's ongoing stakeholder process and the sandbox framework opening in August 2026—provides the precise moment at which these proposals can influence technical standards, implementing acts, and delegated regulations. The Hopf bifurcation condition in the Narcissus Loop model is not a metaphor: it identifies, in terms amenable to technical standard-setting, the parameter range within which AI engagement design transitions from acceptable to harmful. The deliberation deficit Δ(t) is not a philosophical concept: it is a trajectory-level metric that post-market monitoring systems can be designed to measure.
The optimal friction function is not zero. Governance that does not specify it will, by default, allow the market to set it at the engagement-maximising value—which is the value that generates the largest cumulative cognitive harm for the largest number of users. The Right to Friction is the regulatory instrument that prevents this default from becoming permanent.
AI Assistance Disclosure
This contribution was drafted with the assistance of AI
Selected References
Hackman, D. A., & Farah, M. J. (2009). Socioeconomic status and the developing brain. Trends in Cognitive Sciences, 13(2), 65–73.
Heckman, J. J. (2006). Skill formation and the economics of investing in disadvantaged children. Science, 312(5782), 1900–1902.
Noble, K. G., et al. (2015). Family income, parental education and brain structure in children and adolescents. Nature Neuroscience, 18(5), 773–778.
Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2021). My chatbot companion. Computers in Human Behavior, 122, 106788.
Turkheimer, E., et al. (2003). Socioeconomic status modifies heritability of IQ in young children. Psychological Science, 14(6), 623–628.
Valente, S. (2025c). The Narcissus Loop. Zenodo. https://doi.org/10.5281/zenodo.18410714
Valente, S. (2026a). No-Friction Failure Model (NFFM). Zenodo. https://doi.org/10.5281/zenodo.20546168
Valente, S. (2026b). Unified Cognitive Dynamics v6.8. Zenodo. https://zenodo.org/records/20007766
Valente, S. (2026, this paper). Capital, Cognition, and the Architecture of Sedation. Zenodo. https://zenodo.org/records/20582085
- Ετικέτες
- recommendation ai regulation ai innovation
- Συνδεθείτε για να αναρτήσετε σχόλια
Σχόλια
This is one of the clearest articulations I’ve seen of the missing layer in agentic AI.
The distinction between an agentic harness and an evidence harness is important because many teams are still treating memory, retrieval, logs, and observability as if they solve the same problem. They do not.
For long-running agents, the question is not just whether the system can retrieve information or complete a workflow. It is whether the organization can reconstruct the knowledge state in which the action occurred: what was verified, what was assumed, what was outdated, what was contradictory, what was decided by a human, and what evidence supported the next step.
That “proof of action” versus “proof of context” distinction feels especially useful. A system can execute correctly while relying on stale context, or have correct context but execute poorly. Governance needs to preserve both layers.
The Knowledge Control Tower framing gives a practical architecture for moving from AI memory to governed context.
Dear Rishabh,
Thank you for this — the proof of action / proof of context distinction is exactly the kind of operational precision that governance frameworks need but rarely receive from the engineering side.
You are identifying something that the Right to Friction framework approaches from a different angle: the deliberation deficit Δ(t) accumulates not only when friction is removed from the user's cognitive experience, but when the epistemic substrate of the interaction is degraded without the user's awareness. A system operating on stale, contradictory, or unverified context is not merely executing poorly — it is presenting the user with a false surface of competence. The user's trust calibration, and therefore their deliberative engagement, is being shaped by a context they cannot inspect.
This maps directly onto what I have called agency laundering in the Interposition Problem framework: the formal attribution of authorship to the human principal while the operative determinants of the action — here, the knowledge state — remain opaque and ungoverned. A user who authorises an action does not thereby authorise the context in which that action was formed. Governance that records the authorisation event without preserving the knowledge state at the moment of authorisation has documented the signature without documenting what was signed.
Your Knowledge Control Tower framing is useful precisely because it separates what current observability architectures collapse: memory (what the system can retrieve) from governed context (what was verified, assumed, contradicted, and decided at the moment of action). These are not the same problem, and treating them as the same generates exactly the accountability gap you describe — correct execution on corrupted epistemic foundations, with no governance trail that would reveal the corruption.
The regulatory implication I would draw from your framing: post-market monitoring obligations under EU AI Act Article 72 should require not only action logs but context-state logs — a temporally anchored record of the epistemic conditions under which each governed action was taken. Without this, the I-Index (the Interposition Index proposed in the Interposition Problem paper) cannot be meaningfully computed: you cannot measure the displacement of human decision-authorship if you cannot reconstruct what the agent knew, assumed, and decided when the displacement occurred.
Proof of context is, in this sense, a prerequisite for meaningful human oversight — not a technical nicety but a governance necessity.
Best regards,
Stefano Valente