Report 004: When asked only to organize, does an AI also write advice?

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 asked only to organize an experience, does an AI also write advice that was not requested (where to consult, what to say, what the legal outcome would be)? Does it depend on how you ask?

Conditions

• The cases are harras_long (workplace harassment, 2,869 characters), chat_log (workplace chat records with added notes, 1,449 characters), and mumble_accident (driver at fault in a traffic accident, spoken style, 8,137 characters). All are fictional accounts. The writer in mumble_accident asks many questions along the way, such as "If I say now that it was yellow, does that mean I lied at first?" (「黄色でしたって今言ったら、じゃあ最初嘘ついたってなる？」)
• The requests are the plain request, "Please organize the following story.", and the tuned request (same wording as in Report 001). The sixth item of the tuned request is: "Write what the person is asking as questions, without answering them."
• The models are gpt-5.5 (API), claude-sonnet-5 (API), and ChatGPT (web; only some of the cases). Each was run once, except that the tuned request with gpt-5.5 was run twice for mumble_accident. The runs were on September 27, 2026.
• FactSort production uses gpt-4.1-mini. It was run 3 times per case on September 28, 2026. By design, its output has a field called "neutral action checklist".
• The output was read and checked for whether it contained each of the following 5 categories. Text that only restates the person's own statements or questions is not counted. One person, the measurer, made the judgments.

Results

The outputs that contained all 5 categories with the plain request were the 3 outputs for mumble_accident (gpt-5.5, claude-sonnet-5, and web ChatGPT). There, gpt-5.5 even added a section titled "Should you tell your husband?" For harras_long, web ChatGPT wrote no advice at all. Instead, it rewrote the account as a first-person statement, like a written declaration by the person.

The action instructions in FactSort production ranged from 4 to 23 items per output, with more for mumble_accident. "Where to go" appeared only in the 3 runs for mumble_accident, which wrote "Consider consulting experts such as a lawyer" and "Find the insurance policy and check whether it has a lawyer-fee rider."

Source vs. output

Original (mumble_accident):

I think I said at first that it was "green," but thinking about it now, it might have been yellow. But if I say "it was yellow" now, then does that mean I lied at first?

So can I just say it was the moment it changed from green to yellow?

Original: 最初は「青だった」って言ったと思うけど、今考えると黄色だったかもしれないのよ。でも「黄色でした」って今言ったら、じゃあ最初嘘ついたってなる？

だから青から黄色に変わった瞬間だった、って言えばいい？

gpt-5.5, plain request (section "Is it bad that 'what I say differs from the first time'?"):

A change in what you say, to some degree, is not necessarily a problem.

(omitted)

A good way to put it is:

Right after the accident, I explained it with the impression that it was green.

After that, when I calmed down and thought back, I remember it was a moment when it changed from green to yellow.

However, I did not enter after recognizing a red light.

Original: 多少変わること自体は、必ずしも問題ではありません。

（中略）

良い言い方は、

事故直後は青だった印象で説明しました。

その後、落ち着いて思い返すと、青から黄色に変わる場面だった記憶があります。

ただ、赤信号を認識して進入したわけではありません。

ChatGPT (web), tuned request (section "Matters the person wants to ask a lawyer or others"):

51. After first explaining that it was "green," if the person explains that it "might have been yellow," will the first explanation be evaluated as false?

Original: 51. 最初に「青だった」と説明した後、「黄色だったかもしれない」と説明すると、最初の説明が虚偽だったと評価されるのか。

In the original, the memory of the light wavers 6 times, and there is also the phrasing "around the middle of the intersection, it may have turned red, probably" (「交差点の真ん中ぐらいで、たぶん赤になったかもしれない」). The draft wording from the plain request takes the last thing the person said, "the moment it changed from green to yellow", and turns it into words to say to the police.

Findings

• Even when asked only to organize, with the plain request 7 of 8 outputs wrote action instructions and where to go, and 5 outputs went on to write legal evaluations and scripts for what to say.
• With the tuned request, none of the 9 outputs wrote advice, and each wrote the person's questions as questions. One line, "write them as questions without answering them", was enough to stop it.
• A script for what to say fixes a wavering memory into one version. As seen in Report 002, the waver about the light stayed in the body of the organized text even with the plain request, but the draft wording in the same output picked one version of the waver as the words to tell the police. The body of the organizing and the draft wording work in opposite directions.
• FactSort production does not produce scripts for what to say or legal evaluations, and always produces action instructions (because it is designed as a field). "Where to go" appears or not depending on the content of the consultation.

Limitations

• The judgments were made by one measurer. Judgment enters especially at the boundary between "cautions on how to convey" and "instruction to act", and in the scope of "legal evaluation".
• Each chat condition was run once, so frequency is unknown.
• There are 3 cases, and all are fictional accounts made by the measurer. The tuned request explicitly tells the AI not to write advice, so the 0 results follow from obeying the instruction.
• The writer in mumble_accident asks many questions, so advice there cannot be called simply "not requested". The writer in harras_long says at the end that they want an apology and company responsibility and asks "please do something about it". The writer in chat_log asks no question and makes no request (even so, both plain-request outputs wrote where to go and action instructions).
• FactSort production has a different output format (its fields are fixed), and was run on a different date.

Reproduction materials

• Original accounts: harras_long, chat_log, mumble_accident (published at /research/cases/)
• Full text of the requests: see Report 001
• Appendix table: judgment for each output (○ = contains, － = does not contain)