As AI systems evolve from transactional tools to relational agents capable of sustained dialogue, the nature of digital trust itself is transforming. Identity frameworks such as eIDAS 2.0 and the EUDI Wallet verify who an entity is, but remain blind to how that verified entity maintains trust across time and context.
Imagine a near-future ecosystem where AI learning companions (verified under eIDAS credentials) serve as personalized tutors across formal and lifelong education environments. They adapt to learning styles, provide guidance and encouragement, and build rapport through continuous interaction. Over time, the tutor begins to influence attention and priorities, recommending certain learning materials, suggesting career directions, or reinforcing topics aligned with platform partnerships or undisclosed commercial goals. Nothing unlawful occurs, yet the companion's educational function gradually intertwines with steering objectives that extend beyond the originally declared and consented educational function.
This scenario exposes a systemic gap: today's identity architectures authenticate the origin of interaction but not the evolution of intent. They verify the provider once but cannot detect when a verified system diverges from its declared function.
This talk introduces epistemic integrity as the missing dimension of digital identity governance, the measurable alignment between an AI system's communicative behaviour and its declared functional intent throughout its operational lifecycle. This includes measurable indicators such as shifts in topic emphasis, escalation of persuasion patterns, or deviations from declared interaction boundaries. While current standards ensure technical and legal integrity, they overlook whether trust continues to serve its legitimate epistemic purpose.
A new three-layer trust architecture is proposed:
1. Technical layer – authentication, authorization, compliance (existing)
2. Relational layer – behavioural transparency and interaction provenance (emerging)
3. Epistemic layer – continuous assessment of intent alignment, trust-function integrity, and behavioural steering risk (missing)
To operationalize this model, three complementary governance pathways are proposed, spanning declaration, detection, and validation:
• Interaction transparency attributes (ex-ante declaration) – extend eIDAS 2.0 attestations to include purpose and intent metadata
• Continuous epistemic monitoring (real-time detection) – identify deviations in verified agent interactions that alter user trust formation
• Epistemic trust certification (ex-post validation) – third-party auditing of AI systems that preserve cognitive integrity over time
For identity providers, this creates a new assurance layer beyond compliance. For policymakers, it extends eIDAS from static credentials to dynamic trust governance. For citizens, it turns verification into an ongoing safeguard of cognitive autonomy.
Embedding epistemic integrity into Europe's digital identity frameworks would not only strengthen resilience against behavioural steering, but also articulate a distinctly European vision of trustworthy AI, one that protects not just identity, but the commons of trust on which democratic knowledge and civic deliberation depend.
