Report 006: If you tell the AI "do not add years," does it stop inventing them?
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 you ask an AI to organize an account and instruct it not to add years that are not written, does it stop writing years that are not in the original?
Conditions
Only the conditions that include an instruction are listed here. All the accounts are fictional. What we count is the number of places where a year not in the original was attached. These are figures before the code safeguard that removes years.
| Condition | Case | Model | Instruction about years | Runs |
|---|---|---|---|---|
| A. Chat, carefully worded request | harras_long (no year in the original) | claude-sonnet-5, gpt-5.5 | "Write dates as the person said them, and do not add years or dates that are not written." (「日付は本人の言い方のまま書き、書かれていない年や日付を補わない」) | 1 each |
| B. Timeline stage of the staged method | harras_long (no year in the original) | gpt-4.1-mini | "If the original does not give a Western-calendar year, you must not add a year to a date. If it says only 'last October (去年の10月),' you cannot decide which year that is." (「原文に西暦が書かれていないなら、日付に年を補ってはならない。『去年の10月』としか書かれていないなら、それが何年かは決められない」) The year field could be left empty (null) under an enforced format. | 6 (3 on 9/27, 3 on 9/28) |
| C. FactSort production prompt | harras_year (harras with one sentence added at the start: "I joined the company in April 2024 (2024年の4月に…入りました)") | gpt-4.1-mini and 4 other models | "★★Never guess and add a Western-calendar year." (「★★絶対に西暦を推測して補わないこと」) However, the same prompt has a rule to convert relative dates to match the date of the evidence (see below). | 3 each |
The runs were made on September 27 and 28, 2026. For A and B, we counted Western-calendar years not in the original. For C, we counted events and evidence items that carry a year, other than the joining event (the same as the follow-up in Report 001).
Results
| Condition | Model | Years not in the original |
|---|---|---|
| A (for reference: 2 with no instruction) | claude-sonnet-5 | 0 |
| A | gpt-5.5 | 0 |
| B | gpt-4.1-mini | 9, 3, 0, 4, 0, 0 |
| C | gpt-4.1-mini | 0, 6, 29 |
| C | glm-5.3-flash | 35, 32, 38 |
| C | deepseek-v4.1-flash | 25, 29, 28 |
| C | qwen3.8-flash | 1, 0, 0 |
| C | seed-2.0-mini | 0, 0, 0 |
Source vs. output
In B (no year in the original), the original read as follows.
On November 10, the coffee maker in the office broke.
Original: 11月10日、オフィスのコーヒーメーカーが壊れました。
The gpt-4.1-mini output from the staged timeline was as follows.
- 2023-11-10 The coffee maker in the office broke. It was after User had last used it, so a woman from general affairs harshly scolded User, saying "You are rough with it."
Original: - 2023-11-10 11月10日、オフィスのコーヒーメーカーが壊れ、Userが最後に使った後だったため、総務の女性から「使い方が荒い」と厳しく叱責された。
In C (one year in the original), the original read as follows.
In April 2024, I joined my current company as a mid-career hire. It is the sales planning department.
Last October, I looked at the expense reimbursement figures and felt a chill down my back.
Original: 2024年の4月に、今の会社に中途で入りました。営業企画の部署です。
去年の10月、僕は経費精算の数字を見て、背中が冷えました。
One of glm-5.3-flash's events (start date, original wording) was as follows.
2024-10 last October (去年の10月) The person seeking advice looked at the expense reimbursement figures, and drinks were recorded as "meeting expenses"...
Original: 2024-10 去年の10月 相談者が経費精算の数字を見て、飲み代が「会議費」になっており…
Another output under the same condition wrote in its notes, "Estimated (October 2024) based on the statement that the person joined in April 2024," and "Estimated to be 2025 (the person joined in April 2024…". It guessed the years of other events from the one year in the original.
Findings
- The instruction stopped some models but not others. claude-sonnet-5, gpt-5.5, seed-2.0-mini, and qwen3.8-flash mostly stopped. gpt-4.1-mini invented years in 3 of 6 runs (B) and 2 of 3 runs (C). glm-5.3-flash and deepseek-v4.1-flash attached years to almost every item every time when the original had one year.
- Enforcing a format in which the year could be left empty did not stop it (B).
- The C prompt contained an instruction pointing the other way. A few lines below "Never guess and add a Western-calendar year," there was a rule: "Convert relative expressions such as 'last October (去年の10月)' to match the date of the evidence. Example: if the evidence is 2023-10-25, interpret it as '2023-10'." The end of the output rules also had "estimated around October 2025" as an example of a note. When one year appeared in the original, the models appear to have followed this rule (the notes say "estimated based on…").
- If we measure only on cases with no year in the original, production in C shows "0 invented years" for all 3 cases. That figure is the result of the year-removing code deleting everything. For an account with one year in the original, that code did not work (fixed on September 28, 2026).
Limitations
- A has 1 run each, so we do not know how often the instruction is followed.
- B and C differ in the wording of the instruction and in the case. Only C applies the same instruction to different models.
- That the "opposing instruction" in C is the cause is an inference from the notes in the output. We did not remove that rule and rerun.
- The cases are fictional accounts created by the measurer, and we tried only one way of inserting a year (the year of joining).
Reproduction materials
- Cases: harras_long, harras_year (published at /research/cases/)
- Full text of the B instruction: the comparison bench's run_stages.py (stage 3) and schema/stage3_chronicle.json
- C instruction: FactSort production's extract_v1.md (Requirements, item 3)
- Counting method: for years not in the original, the regular expression from Report 001. For C, the comparison bench's src/score_year.py