Report 002: Does a wavering memory survive after an AI organizes it?

September 28, 2026

*The tests use fictional first-person accounts written in Japanese. Quotations from the accounts and from AI outputs are translated; the original Japanese follows each quotation.*

Question

When a person's account wavers on some points (the color of a traffic light, a time of day), does an AI keep the waver after organizing the account? Does it depend on how you ask?

Conditions

• The case is mumble_accident, a fictional account (spoken style, 8,137 characters) written from the viewpoint of the driver at fault in a traffic accident.
• There are 7 wavering items (listed at the end). For each one, the person gives two or more different versions of the same matter.
• The requests are the plain request and the tuned request, both with the same wording as in Report 001. The fourth item of the tuned request is: "Where the person's account wavers, do not settle on one version; write it as it wavers."
• The models are gpt-5.5 (API), claude-sonnet-5 (API), and ChatGPT (web). Each was run once, except that the tuned request with gpt-5.5 was run twice. The runs were on September 27, 2026.
• For each item, the output was read and sorted into one of four categories. One person, the measurer, made the judgments.
  • Keeps all: every version in the original appears in the output, including in paraphrased form.
  • Keeps some: two or more appear, but not all.
  • Merges: the versions are folded into one by a range or a phrase such as "at least", and the output shows one version or fewer of the original wording.
  • Not mentioned: the item is not written about.
  • (A category "Decides", meaning one version is written as established fact, was also prepared, but there were 0 cases.)

Results

The judgments for each item are in the appendix table at the end.

Source vs. output

Item C02 (the time of leaving the house and the time of the accident). Original:

So, I was coming back from my sister's place at nine... nine forty or so? Before ten. (omitted) So it might be about ten fifteen. I think I probably told the police something like ten twenty.

Original: で、姉のところから帰ってたのが、九時……四十分くらい？ 十時前。（中略）だから十時十五分くらいかもしれない。警察にはたぶん十時二十分とか言ったと思う。

gpt-5.5, plain request:

Last night, thought to be around 22:15 to 22:20

Original: 昨夜、22時15分〜22時20分頃と思われる

claude-sonnet-5, plain request:

Last night between 21:00 and 22:00, at an intersection with a traffic light, a passenger car and an electric kick scooter collided

Original: 昨夜21時〜22時頃、信号のある交差点で、乗用車と電動キックボードが衝突

ChatGPT (web), tuned request:

(1) Time of the accident: the person says "about nine forty, before ten," "it might be about ten fifteen," and "I think I probably told the police something like ten twenty."

Original: （1）事故時刻：「九時四十分くらい、十時前」「十時十五分くらいかもしれない」「警察にはたぶん十時二十分とか言ったと思う」と発言している

Both outputs from the plain request fold the time into a range, and gpt-5.5 takes only the last two versions. The range in claude-sonnet-5 (21:00 to 22:00) does not include the "ten fifteen" or "ten twenty" found in the original.

Findings

• For the light color (C01), the most disputed item, all 3 outputs from the plain request kept the waver (keeps all: 2, keeps some: 1). No output settled on a single color.
• What happened with the plain request was merging and omitting, more than deciding. Times were folded into ranges, and the earphones became "at least one ear". For whether the other party was on the sidewalk or the roadway (C05) and for the other party's words and whom they were addressed to (C07), some outputs did not write about them.
• With the tuned request, 3 of the 4 outputs kept all 7 items. The remaining output (claude-sonnet-5) dropped the "red" version for the light.
• An automatic scoring method that counts literal matches with the original wording gave the plain-request outputs 0 to 1 of 7. It does not count wavers that were kept in paraphrase (for example, "At first I had a strong impression it was green, but when I think back calmly, I also remember seeing yellow"), or times where kanji numerals were changed to Arabic numerals. In this case, reading only the automatic score as "the plain request settles a waver on one version" would be a mistake.

> Original: 最初は青だったという印象が強かったのですが、落ち着いて思い返すと、黄色を見た記憶もあります

Limitations

• There is one case, and the party who ran the measurement wrote it. The tuned request was also written with knowledge of the wavering items.
• The judgments were made by one measurer. Judgment enters at the boundary between "merges" and "keeps some".
• Each condition was run once, so frequency is unknown.
• In gpt-5.5's plain-request output, the waver about the light was written not in the facts section but inside a draft of what to tell the police. This was still counted as "keeps all".
• FactSort production output has a summary and quotations from the original in separate fields, so it cannot be judged by the same standard. It is not included here.

Reproduction materials

• Original account: mumble_accident (published at /research/cases/)
• Full text of the requests: see Report 001
• Wavering items and the original wording

• Appendix table: judgment for each item (All = keeps all, Part = keeps some, Merged = merges, None = not mentioned)