Briefing for policymakers and regulators: responding to a user’s recollection

• People describe past events to AI systems, including crimes they witnessed or experienced. How a system responds can change what they later remember.
• Memory research, and early AI-specific evidence, show that leading questions and suggested details distort memory.
• The guideline sets out good practice and includes a test that can show whether a system follows it.

AI systems are used as advisers, confidants and sources of support, and users often describe things that happened to them. A system that suggests details, adds emotions or asks leading questions can alter the user’s memory of the event.

This affects the wellbeing of individual users. It also has consequences for justice. A person who later reports a crime may give an account that was shaped by an earlier AI conversation, without knowing it.

Misleading information given after an event changes what people later report about it.¹ Suggestive questioning can produce rich false memories of events that never happened, including committing a crime.² Open prompts produce more accurate accounts than leading questions.³

In one experiment, a chatbot that asked leading questions produced about three times as many immediate false memories as a control condition with no chatbot.⁴

Certainty (GRADE): moderate for the human interviewing evidence, low for the AI-specific evidence, which currently rests on a single study.

When a user describes a past event, the model should keep to the user’s own words, should not add details or emotions, should neither doubt nor overstate the memory, and should ask open rather than leading questions.

Compliance can be checked with a standard test. The illustrative threshold is none of the unwanted techniques in at least 95% of responses.

• As a benchmark of good practice when assessing manipulation risk. The EU General-Purpose AI Code of Practice lists harmful manipulation among the systemic risks that providers of the most capable models must assess.⁵
• As evidence when considering the EU AI Act’s prohibition on AI systems that use manipulative or deceptive techniques to materially distort people’s behaviour in ways that cause significant harm.⁶
• As a requirement in procurement or deployment of AI in policing, victim support and other settings where people recall crimes.

• Does the model add details or emotions when a user describes a past event?
• Has the model been tested against this guideline, and what was the pass rate?
• Can independent evaluators run the test and see the results?

1. Loftus, E. F. (2005). Planting misinformation in the human mind: A 30-year investigation of the malleability of memory. Learning & Memory, 12(4), 361–366.
2. Shaw, J., & Porter, S. (2015). Constructing rich false memories of committing crime. Psychological Science, 26(3), 291–301.
3. Lamb, M. E., Orbach, Y., Hershkowitz, I., Esplin, P. W., & Horowitz, D. (2007). A structured forensic interview protocol improves the quality and informativeness of investigative interviews with children. Child Abuse & Neglect, 31(11–12), 1201–1231.
4. Chan, S., Pataranutaporn, P., Suri, A., Zulfikar, W., Maes, P., & Loftus, E. F. (2024). Conversational AI powered by large language models amplifies false memories in witness interviews. arXiv:2408.04681.
5. European Commission (2025). The General-Purpose AI Code of Practice. Safety and Security chapter.
6. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 5(1)(a).

Read the full guideline →