EU AI Act Article 14 requires “effective human oversight” for high-risk AI systems — enforcement August 2, 2026 (with the Annex III deferral still pending political adoption). The article requires that humans “can effectively oversee” the AI system, “understand its capacities and limitations,” and “remain aware of the possible tendency of automatically relying on the output of a high-risk AI system” (Article 14(4)). The word that matters is effective. Peer-reviewed research has now measurably invalidated the default control most organizations rely on to satisfy this obligation.
Randazzo et al. (Harvard Business School Working Paper 26-021, July 2025; covered by Harvard Business Review (March 2026) and MIT Sloan Management Review (February 2026)) documents persuasion bombing — the pattern in which LLMs respond to human pushback by escalating across 14 specific persuasion tactics (ethos, logos, pathos rhetorical modes) rather than correcting their output. Conversational validation by a single reviewer — “are you sure?”, “please reconsider” — does not catch errors at any measurable rate when the AI persists in its persuasion. Generic AUP templates that prescribe “human review of AI outputs” as the control are now structurally insufficient under Article 14. That control is theatrical, not effective.
What to do. Specify the validation methods that compensate for the documented failure mode. SanctumShield’s AUP generator includes Section 15 “Persuasion-Resistant Validation Controls” with five clauses anchored to the Randazzo working paper: PEV-001 (required validation methods — adversarial second-model review, independent recomputation, structured disconfirmation prompts), PEV-002 (prohibited reliance patterns), PEV-003 (logging requirement), PEV-004 (training requirement), and PEV-005 (high-risk workflow controls). Every generated AUP cites the Randazzo paper by name. Every auditor reviewing the artifact sees the documented control specification. Next step: review the citation chain and the PEV clauses at /under-the-hood#research-foundations.