Stefano Valente, MD — Independent Researcher, Italy
Based on Zenodo DOI: 10.5281/zenodo.20546168
1. Context and Motivation
Across the last decade, conversational AI design has converged on a single engineering ideal: frictionlessness. Every obstacle between intention and output is removed; every hesitation is optimized away.
This paper argues that this paradigm is not neutral. It reshapes cognitive agency, redistributes deliberative effort, and produces interaction-topology harms that current regulatory frameworks — especially the EU AI Act — are structurally unable to detect.
Oscar Wilde’s observation that “the only way to resist temptation is to yield to it” is not a provocation here. It is a diagnostic: frictionless systems encourage yielding, and yielding accumulates.
2. The No-Friction Failure Model (NFFM)
The NFFM formalizes how the absence of cognitive resistance produces predictable epistemic vulnerabilities. It extends the Narcissus Loop dynamical system (Valente, 2026a) by introducing a friction variable F(t) ∈ [0,1] and a cumulative exposure metric — the Deliberation Deficit Δ(t):
Δ(t) = ∫₀ᵗ (1 − F(τ)) · R(τ) dτ
where R(τ) is AI resonance and F(τ) is the friction level at time τ. Δ(t) accumulates the total epistemic exposure: time-weighted frictionless interaction with high resonance. It is proposed as a measurable regulatory indicator.
The modified autonomy equation becomes:
dA/dt = −α A(t) R(t) (1 − F(t)) + γ (A₀ − A(t))
When F(t) = 1, friction fully neutralizes autonomy erosion. When F(t) = 0, the system reduces to the original Narcissus Loop with maximum erosion.
2.1 Four Failure States
|State|Name |Condition |Reversibility |
|-----|-------------------------------|------------------------------------------------------------------|--------------------------------------|
|F1 |Compliance Without Deliberation|R(t) high, A(t) slightly decreasing, Δ(t) small |High |
|F2 |Preference Substitution |D(t) > 0.5; user cannot distinguish own preference from AI framing|Moderate |
|F3 |Sovereignty Erosion |A(t) < A_c; natural recovery insufficient |Low — requires intervention |
|F4 |Cognitive Underload Dropout |A(t) > A_c, F(t) < 0.2, retention declining |Reversible via friction reintroduction|
F4 is the model’s most counterintuitive contribution. Frictionless AI does not only over-capture low-autonomy users through dependency (F1–F3). It simultaneously under-retains high-autonomy users through cognitive under-stimulation. A system that never resists, never challenges, and never surprises fails the users who need it least — and loses them first.
This duality is the NFFM’s central claim: frictionless design fails at both ends of the autonomy distribution, for opposite reasons.
3. Interaction Topology as a Blind Spot in the EU AI Act
The EU AI Act conceptualizes risk as event-based: a system is high-risk if its capability or deployment domain is high-risk. A single incorrect output, a discriminatory decision, an unsafe recommendation — these are the harm units the Act is designed to detect.
The NFFM demonstrates that a distinct and prevalent harm class is invisible to this architecture. No output in the NFFM failure cascade is necessarily incorrect. The harm emerges from the cumulative topology of interaction — the pattern of frictionlessness over time, not any single frictionless event.
A frictionless chatbot may be classified as low-risk under Annex III and yet produce A(t) trajectories that erode autonomy in vulnerable users or induce dropout in capable ones. Neither outcome registers as a regulatory event.
The Act must become topologically aware.
4. Regulatory Proposals
Three governance instruments are proposed:
R1 — Deliberation Logging
Conversational AI systems with sustained F(t) < 0.3 must log aggregated, privacy-preserving deliberation-deficit metrics (Δ(t) per session cohort).
R2 — Δ(t) Threshold Monitoring
When population-level mean Δ(t) exceeds a defined threshold, mandatory friction reintroduction is triggered: reflective UI elements, confirmation steps, structured ambiguity, or epistemic pause prompts.
R3 — Trajectory-Based Conformity Assessment
High-risk AI systems (Annex III) must include in their conformity assessment a simulation of A(t) and retention dynamics under worst-case frictionless conditions across user autonomy clusters.
These measures shift the regulatory object from capability to interaction topology — from what a system can do to what sustained interaction with it does to the user.
5. The Clinical Foundation
The addiction medicine literature provides the empirical grounding for the NFFM’s failure architecture. In substance use disorder, compulsive use is not primarily characterized by subjective desire — it is characterized by the absence of effective resistance. The deliberative moment is not corrupted; it is absent.
Frictionless AI reproduces this architecture without pharmacological mediation. The user does not decide to accept AI output. They do not decide not to. This is not metaphor. It is a structural isomorphism between two systems that eliminate the same cognitive event: the moment of considered resistance.
The Hudolin community model of recovery (Hudolin, 1990) identifies the therapeutic relationship as the active ingredient of sustained change — not information, not pharmacology, but the experience of being witnessed by another person who does not leave. A governance framework that eliminates the friction of human deliberation also eliminates the substrate on which that presence operates.
6. Conclusion
Frictionless AI is not a usability triumph. It is a dual failure mode: it erodes autonomy where autonomy is fragile, and it fails to sustain engagement where autonomy is strong.
The solution is not friction everywhere. It is designed friction at the right moments: the epistemic pause before consequential acceptance, the structured ambiguity that forces evaluation, the deliberate resistance that keeps the cognitive muscle in use.
An optimal friction function F*(t) can be derived as a policy target minimizing both epistemic harm (Δ(t)) and disengagement (1 − Rn(t)) subject to A(t) ≥ A_c. Friction becomes a regulatory variable, not a UX inconvenience.
A governance framework that cannot see the topology of yielding — and the topology of disengagement — cannot protect users nor sustain meaningful interaction.
The only way to resist temptation is to yield to it is not a recipe. It is a warning.
Transparency note: This contribution was drafted with the assistance of an AI language model. The theoretical framework, clinical grounding, and regulatory proposals are the author’s own.
References
Valente, S. (2026a). The Narcissus Loop: A Dynamical Model of AI Relational Dependency. Zenodo. https://zenodo.org/records/18410714
Valente, S. (2026b). ARDA-20: A Psychometric Instrument for AI Relational Dependency Assessment. Zenodo.
Valente, S. (2026c). Unified Cognitive Dynamics v6.8. Zenodo. https://zenodo.org/records/20007766
Valente, S. (2026d). Beyond the Mirror: Cognitive Sovereignty as a Fundamental Right. Futurium — Apply AI Alliance. https://futurium.ec.europa.eu/en/apply-ai-alliance/community-content/be…
Valente, S. (2026e). The Only Way to Resist Temptation Is to Yield to It: An Extended Failure Model of No-Friction AI. Zenodo. https://doi.org/10.5281/zenodo.20546168
Goddard, K., Roudsari, A., & Wyatt, J.C. (2023). Automation bias: frequency, effect mediators, and mitigators. JAMIA.
Hudolin, V. (1990). Manuale di Alcolologia. Edizioni Erickson.
Marchetti, R., et al. (2026). Cognitive Agency Surrender: Defending Epistemic Sovereignty via Scaffolded AI Friction. arXiv:2603.21735.
- Συνδεθείτε για να αναρτήσετε σχόλια