Prolonged human–AI interaction as a post-deployment governance challenge

Dear members of the Apply AI Alliance,
 

in the context of the implementation of the EU AI Act and the Apply AI Strategy, I would like to share a set of recent open-access preprints (January 2026) that I have made available on Zenodo.

The papers explore several under-discussed aspects of prolonged interaction with LLM-based systems and AI systems classified as high-risk and low-risk, with a specific focus on post-deployment effects. 

In particular, the work addresses:
 

• Digital philotherapy: a low-intensity, non-clinical, non-substitutive framework for mental support in contexts with limited access to psychotherapy, inspired by DBT and bibliotherapy;
• Human cognitive adaptation and cognitive safety behavior in high-risk AI systems, as a potential regulatory blind spot in the current formulation of the AI Act;
• Deterministic governance architectures (SUPREME-1 v3, OMNIA-1, EROS-1), designed to support identity stability, proportional restraint, auditability, and post-market governance;
• Cognitive Access Parity (CAP) as a principle to mitigate cognitive exclusion in low-connectivity or resource-constrained settings;
• Failure modes in low-risk systems, including over-reliance, channel-induced exclusion, and temporal accumulation of risk.

All contributions are fully open access and intentionally adopt a non-anthropomorphic, risk-proportionate approach, with emphasis on human agency, auditability, and long-term governance, rather than performance or short-term outcomes.

Links to the main collection (recent uploads):

• https://zenodo.org/records/18385137
• https://zenodo.org/records/18390878
• https://zenodo.org/records/18310366 (SUPREME-1 v3 core)
• https://zenodo.org/records/18390100
• https://zenodo.org/records/18375149
• https://zenodo.org/records/18369262
• https://zenodo.org/records/18360968
• https://zenodo.org/records/18347345
• https://zenodo.org/records/18335916
• https://zenodo.org/records/18331459
• https://zenodo.org/records/18327255

I would be very interested in discussing within the Alliance how cognitive and relational dimensions could be more explicitly integrated into sector-specific AI governance, particularly with respect to post-deployment monitoring and to AI systems formally classified as low-risk but capable of generating cumulative or delayed risks over time.

Feedback and discussion are very welcome, and I remain open to dialogue or collaboration.

 Kind regards,

Stefano Valente,MD