When we approve an AI summary, are we still checking our own memory?

A proposal becomes a decision

Imagine this explicitly fictional situation. After a meeting, the minutes say a launch has been approved. During the discussion, it was only a proposal, subject to a further estimate. Participants read the text and ask whether it matches their recollection. The wording feels familiar. Nobody notices the shift.

The problem extends beyond the incorrect sentence. The document is now present as everyone reconstructs the meeting. Collective approval may give authority to an account the sources do not support. To check a summary, we need an opportunity to form a judgement before it shapes that judgement.

When later information shifts recollection

The misinformation effect describes how misleading information received after an event can influence memory of that event. Mattea Sim, Yael Eiger and Tadayoshi Kohno investigate a form involving AI summaries in a paper posted on 23 September 2026 [1].

The analysis includes 328 adults recruited online in the United States. They watch a short animated accident video and read a summary 24 to 48 hours later. Correct recognition of the traffic sign reaches 83.6% after an accurate summary and 44.8% after a misleading one. The erroneous detail is experimentally manipulated; all texts originate from AI generation with controlled adaptations.

There is no statistically significant difference between labels attributing the text to AI or a human. This does not establish universal equivalence. Without a group reading no summary, the percentages also do not measure decline against an assessed initial memory.

A memory test is not professional verification

Participants complete a recognition test, not a professional review with access to sources. The study therefore establishes neither a general impossibility of validating AI nor the effectiveness of the procedure proposed here. Applying the question to meetings remains an editorial inference.

In an organisation, verification might involve returning to a recording, comparing contemporary notes or locating a condition attached to an agreement. That differs from a general impression that the text sounds like what people said. It requires documents, time and permission to challenge the proposed account.

‘Reviewed and approved’ describes an action. It does not say which sources were consulted, in what order or which disagreements were resolved. Approval becomes more useful when it identifies the facts actually checked.

Two different risks

Memory contamination concerns what someone believes they remember. Excessive reliance on automated output concerns the weight they give that output. A person may distrust a tool yet already have read wording that influences recollection. Conversely, someone may accept text for convenience without their memory changing.

The NIST profile for generative AI addresses human–AI configurations and risks of excessive reliance [2]. It complements the cognitive question: requesting approval does not yet define the conditions for exercising oversight. For a contractual commitment or assigned responsibility, those conditions deserve to be explicit.

Putting it into practice: sources before the summary

Here is a proposal to evaluate in proportion to the stakes. The aim is not to double every reading task, but to protect the few facts whose distortion would change a decision.

1. Identify facts that entail a decision, responsibility or promise: who committed to what, and under which condition?

1. Check those facts in the sources before consulting the summary. Dated notes or a recorded passage allow examination of what was actually said.

1. Record an initial judgement, including uncertainty. A missing condition or an agreement that cannot be found must remain marked as such.

1. Read the summary and examine discrepancies. Distinguish acceptable rewording from a change in status, scope or responsibility.

1. Ask a second reader with access to the sources to resolve important contradictions. Let them form their view before seeing the disputed conclusion too.

Selected excerpts can themselves steer reading. Keep access to the full context so that an influential summary is not replaced by equally directive selections. If the summary has already been read, acknowledge that instead of claiming unaltered independence.

Keep the reference account open to challenge

This requirement connects with the positioning of Mnemory Notes and Mnemory Knowledge: preserving sources, context and the possibility of challenging a summary before it becomes the reference account. It does not mean these products already provide blind validation or tracking of exposure to summaries.

The practical aim is simple: retain what allows a question to be reopened. A summary helps people work; its approval should remain connected to what was checked. The quality of human oversight also depends on what happens before the first reading.

Philippe Contal