Early Warning Signals for Behavioural Drift in General-Purpose AI Systems

General-purpose AI models are increasingly deployed in critical infrastructures, public services and decision-support systems.

While regulatory frameworks such as the EU AI Act address risk classes and prohibited practices, one technical risk remains insufficiently mapped: behavioural drift in black-box GPAI systems over time.

I would like to share a concise forensic perspective on early-warning indicators that can help detect emerging instabilities before they escalate into safety, reliability or compliance failures.

Observed weak signals include:

– Cross-format inconsistencies (text, code, reasoning, symbolic output)

– Semantic meaning-shifts under stable prompts

– Degradation of deterministic task patterns

– Compression artefacts in reasoning chains

– Emergent self-correction or meta-answer loops

Such indicators are highly relevant for post-market monitoring, technical documentation and continuous risk assessment under Articles 52–56 and 62–66 of the AI Act.

From a European perspective, systematically capturing these signals can strengthen trustworthy, predictable and human-centric AI deployment, especially for public administrations and high-impact use cases.

I would be interested in exchanging with others working on AI testing, monitoring or governance on how behavioural drift is currently addressed in practice, particularly in European regulatory and deployment contexts.

Marion Koziol