Search the Atlas

Search risks, controls, and glossary terms

MediumUndisclosed AIRealized

Undisclosed AI interaction

Transparency & Explainability

Description

Users are not told they are interacting with an AI system, breaching transparency obligations (e.g., EU AI Act Art. 50).

Example scenario

Customers believe they are chatting with a human agent, not the AI assistant.

Real-world evidenceRealized

Multiple government-run algorithmic decision systems have caused documented large-scale harm due to inadequate human review or override mechanisms, resulting in parliamentary inquiries, court rulings, and policy reversals. These are confirmed production failures, not experiments.

Primary mitigations

  • Clear AI disclosure at start of interaction
  • persistent labelling
  • opt-out to human.

Detection signals

Disclosure-presence audits; user-comprehension testing.

Mitigating controls

4
Non-agentic controls

Related risks in Transparency & Explainability