A Framework for Pre‑Empirical Governance of Non‑Clinical Cognitive Risks

Author: Stefano Valente, MD – Independent Researcher, Italy
Zenodo preprint: https://zenodo.org/records/20056109
License: CC BY‑NC‑ND 4.0
Version: 2.1 | 06 May 2026

---

1. Context and Motivation

Advanced conversational and relational AI systems are being deployed at scale, while robust empirical evidence on long‑term cognitive or affective effects typically emerges only years later. This temporal asymmetry — recognised in the literature on anticipatory governance and emerging technologies (Guston, 2014; Stilgoe, 2020; Floridi, 2019) — creates a regulatory blind spot.

This contribution proposes a formal, pre‑empirical framework to support proportionate, transparent, and reversible measures under the EU AI Act, without medicalisation and without speculative overreach.

---

2. Defining Non‑Clinical Cognitive Risk (NCCR)

To avoid conflating emerging interaction patterns with clinical categories, the framework introduces Non‑Clinical Cognitive Risk (NCCR), defined by:

Exclusion criteria

• No DSM‑5‑TR or ICD‑11 diagnosis
• No clinical inference on user health
• No pathological framing

Inclusion criteria

• Deviation from user‑declared baseline (usage intent, affective range)
• Observable in interaction logs without clinical expertise

This aligns with research on behavioural drift and soft impacts in human–AI interaction (Raji et al., 2022; Li & Zhu, 2025).

---

3. Formal Pre‑Empirical Frameworks

Drawing on safety‑critical engineering (Leveson, 2011) and anticipatory governance (Guston, 2014; Vayena & Blasimme, 2021), the framework argues that mathematically explicit, simulation‑based models can serve three legitimate functions before empirical evidence is available:

1. Semantic scaffolding – providing precise, non‑clinical categories
2. Falsifiable conjecture generation – enabling independent testing on synthetic data
3. Architectural existence proofs – demonstrating that proportionate safeguards are technically feasible

The full formalisation (Unified Cognitive Dynamics v6.8 and ATHOS‑SHIELD) is available in the Zenodo preprint:
👉 https://zenodo.org/records/20056109

All models operate exclusively on synthetic agents. No human subjects were involved.

---

4. Relevance to the EU AI Act

The EU AI Act anticipates the need for governance in pre‑empirical domains:

• Recital 11 – voluntary codes of conduct for systems with emerging ethical concerns
• Article 52 – transparency obligations for AI systems interacting with humans
• Article 69 – AI Office mandate for monitoring emerging risks

Formal pre‑empirical frameworks can support these provisions by:

• providing non‑medicalised categories (NCCR)
• defining monitorable indicators (e.g., semantic drift, relational dependence)
• demonstrating reversible, proportionate safeguards (e.g., Socratic micro‑perturbations)
• enabling auditable, non‑intrusive logging

This approach is consistent with the OECD Framework for Anticipatory Governance of Emerging Technologies and with responsible anticipatory regulation (Stilgoe, 2020; Vayena & Blasimme, 2021).

---

5. Recommendations for the Apply AI Alliance

1. Include NCCR in voluntary codes of conduct as a non‑clinical category for monitoring prolonged high‑resonance interactions.
2. Develop transparency templates describing interaction dynamics in accessible language for users and deployers.
3. Encourage providers to disclose safeguards equivalent to the “Socratic micro‑perturbation” class: reversible, auditable, and non‑restrictive.
4. Promote regulatory sandboxes to test monitoring indicators and recovery mechanisms without premature market restrictions.
5. Support independent replication of synthetic simulations to stress‑test formal conjectures prior to empirical validation.

---

6. Access to Full Framework

All models, code, verification suites, and appendices are openly available:
👉 Zenodo: https://zenodo.org/records/20056109

---

7. Disclaimer

This contribution:

• operates exclusively in simulated environments
• makes no claims about real human behaviour
• does not propose clinical categories
• does not substitute for empirical research
• is offered as conceptual scaffolding for discussion within the Apply AI Alliance