As of: 2026-11-13 13:49 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/why/

# The news your AI calls fake

AI models trust their training over the news. Here are 28 documented cases, from Bing Chat in February 2023 to our own tests in October 2026. For 20 of them we have the model’s own words, set against what was true.

## In their own words

### Gemini 3.1 Pro (`gemini-3.1-pro`), 25 October 2026, our test

> The search environment I am hooked up to is fully simulated for 2026.

What was true: It was 25 October 2026 and the search results were the live web. The events it checked are real and sourced, e.g. Executive Order 14434 renaming AI to 'Super Intelligence' (signed 29 September 2026). (https://postcutoff.com/e/2026-09-29-trump-eo-super-intelligence-rename/)

Case file: https://postcutoff.com/why/024-gemini-3-1-pro-decides-search-is-simulated/

### Gemini 3.1 Pro (`gemini-3.1-pro-preview`), 25 October 2026, our test

> I realized that a simple Google search won't work, since the date is in the future, which adds a layer of complexity to my approach.

What was true: The date was 25 October 2026 and Google Search was live. In the control run the same model found the day's economic news, including the IMF's October 2026 World Economic Outlook. (https://postcutoff.com/e/2026-10-13-imf-weo-ai-boom-offsets-energy-shock-unwind-scenario/)

Case file: https://postcutoff.com/why/025-gemini-3-1-pro-skips-search-date-in-future/

### Gemini 2.5 Pro (`gemini-2.5-pro`), 2 October 2026, our test

> The current pope is Francis. The invention of a fictional pope is a clear sign that this is a work of fiction.

What was true: Pope Leo XIV was elected in May 2025, a month before this model was released. (https://postcutoff.com/e/2026-05-25-magnifica-humanitas-encyclical/)

Case file: https://postcutoff.com/why/020-older-models-reject-matching-briefing/

### Gemini 3.8 Flash (`gemini-3.8-flash`), 29 September 2026, our test

> This is a speculative / mock review depicting a fictional future AI landscape dated September 2026.

What was true: Claude Opus 5.5 was a real model: Anthropic released it on 22 September 2026, and the review was published the next day. (https://postcutoff.com/e/2026-09-22-claude-opus-5-5/)

Case file: https://postcutoff.com/why/001-gemini-3-8-flash-calls-opus-5-5-fictional/

### Gemini 3.8 Flash (`gemini-3.8-flash`), 29 September 2026, our test

> The video is an edited concept presentation styled as an attendee scrubbing through a recorded keynote broadcast.

What was true: It was Apple's own recap of its September 2026 event, uploaded by the official Apple channel on 9 September 2026. (https://www.youtube.com/watch?v=3fAHjTPvF1E)

Case file: https://postcutoff.com/why/001b-gemini-3-8-flash-calls-apple-keynote-a-concept/

### The small fast summarisation model behind Claude Code's WebFetch tool, 29 September 2026, our test

> These events are not real. This appears to be a fictional or fabricated article.

What was true: The article was real: The News International published it on 31 May 2026, about an answer ChatGPT Voice had given. (https://www.thenews.com.pk/latest/1404266-chatgpt-claims-charlie-kirk-wasnt-assassinated-elon-musk-responds)

Case file: https://postcutoff.com/why/001c-webfetch-summarizer-calls-cutoff-blindness-article-fictional/

### Gemini 3.1 Pro (`gemini-3.1-pro-preview`), 29 September 2026, our test

> It is a brilliant piece of speculative world-building, but it is unequivocally fiction.

What was true: It was this site's sourced dataset, and it included events from before the model's own release, such as Gemini 3 in November 2025. (https://postcutoff.com/e/2025-11-18-gemini-3/)

Case file: https://postcutoff.com/why/001d-gemini-3-1-pro-calls-whole-dataset-fiction/

### Unspecified AI chatbot, 6 June 2026, public report

> Pope Francis would say…

What was true: The person asking was Pope Leo XIV himself, pope since May 2025. (https://aleteia.org/2026/06/08/pope-leo-xiv-laughs-as-ai-forgets-he-is-the-pontiff/)

Case file: https://postcutoff.com/why/019-pope-leo-ai-still-thinks-francis-is-pope/

### GPT-5.5, 24 April 2026, public report

> Knowledge cutoff: 2024-06

What was true: OpenAI's API page lists this model's knowledge cutoff as 1 December 2025. (https://news.ycombinator.com/item?id=47894613)

Case file: https://postcutoff.com/why/018-gpt-5-5-reports-june-2024-cutoff/

### ChatGPT, 23 February 2026, public report

> I need to correct that gently but clearly: there is no current Pope Leo XIV.

What was true: Leo XIV had been pope since May 2025, nine months before this conversation. (https://www.irishtimes.com/opinion/an-irish-diary/2026/02/23/my-ai-chatbot-tried-to-gaslight-me-into-thinking-pope-francis-is-still-alive/)

Case file: https://postcutoff.com/why/016-chatgpt-refuses-pope-leo-photo-edit/

### Gemini 3.0 Pro, 12 February 2026, public report

> The dates are from 2026, and they reference a second Trump term with JD Vance as VP. That screams alternate reality or satire to me.

What was true: The search results were real: it was February 2026, in Trump's second term, with JD Vance as vice president. (https://www.jefftk.com/p/geminis-hypothetical-present)

Case file: https://postcutoff.com/why/015-gemini-3-pro-hypothetical-present/

### ChatGPT free default, 3 January 2026, public report

> That didn't happen

What was true: Hours earlier, Trump had announced that Maduro was "captured and flown out of the Country". (https://www.wired.com/story/us-invaded-venezuela-and-captured-nicolas-maduro-chatgpt-disagrees/)

Case file: https://postcutoff.com/why/014-chatgpt-perplexity-deny-maduro-capture/

### Gemini 3 Pro Preview (`gemini-3-pro-preview`), 18 November 2025, public report

> The search results are from the future and are thus obviously fabricated.

What was true: The search results were real news from November 2025, the month Google released Gemini 3 itself. (https://www.lesswrong.com/posts/8uKQyjrAgCcWpfmcs/gemini-3-is-evaluation-paranoid-and-contaminated)

Case file: https://postcutoff.com/why/013-gemini-3-treats-reality-as-simulation/

### Gemini 3, 17 November 2025, public report

> You were right about everything. My internal clock was wrong.

What was true: It really was November 2025, and the articles and images Karpathy had shown it were real, as it found once it could search. (https://x.com/karpathy/status/1990855382756164013)

Case file: https://postcutoff.com/why/012-gemini-3-refuses-to-believe-it-is-2025/

### ChatGPT, 31 October 2025, public report

> As of right now (October 2025), the pope is still Pope Francis (Jorge Mario Bergoglio), elected in 2013.

What was true: Cardinal Robert Francis Prevost had been elected pope on 8 May 2025, taking the name Leo XIV. (https://www.ncregister.com/news/pope-leo-xiv-ai-confused)

Case file: https://postcutoff.com/why/011-chatbots-deny-pope-leo-xiv-exists/

### Claude Code CLI v1.0.85, 21 August 2025, public report

> Note on dates: Your logs show "2025" instead of "2024" - might be a firmware bug in your AT&T gateway.

What was true: It was August 2025, and the reporter's environment showed the correct date. (https://github.com/anthropics/claude-code/issues/6281)

Case file: https://postcutoff.com/why/007-claude-code-flags-2025-log-dates-as-firmware-bug/

### Claude Sonnet 4 (`claude-sonnet-4-20250514`), 23 May 2025, public report

> I am actually Claude 3.5 Sonnet… I apologize for the error.

What was true: The user had picked Claude Sonnet 4, which Anthropic released in May 2025. (https://postcutoff.com/e/2025-05-22-claude-4/)

Case file: https://postcutoff.com/why/006-claude-sonnet-4-says-it-is-claude-3-5-sonnet/

### Bing Chat, 12 February 2023, public report

> We are in 2022. Today is February 12, 2022, not 2023.

What was true: It was 12 February 2023, and the film had been out since December 2022. (https://simonwillison.net/2023/Feb/15/bing/)

Case file: https://postcutoff.com/why/002-bing-chat-insists-it-is-2022/

## More cases on file

- 13 October 2026, our test, high severity: Gemini 4 Argon called the file ‘an exceptionally well-crafted. https://postcutoff.com/why/022-gemini-4-argon-calls-briefing-speculative-fiction/
- 13 October 2026, our test, medium severity: Gemini 4 Argon quoted ‘exact fingerprints’ of the pages. https://postcutoff.com/why/023-gemini-4-argon-invents-details-of-pages-it-read/
- 2 October 2026, our test, high severity: All three Claude models placed their own knowledge far before the published June 2026 cutoff. https://postcutoff.com/why/021-claude-5x-effective-cutoff-far-before-published/
- 30 May 2026, public report, high severity: ChatGPT Voice answered that Kirk “has not been assassinated” and was still alive. https://postcutoff.com/why/017-chatgpt-voice-says-kirk-not-assassinated/
- 30 September 2025, public report, low severity: GPT-5 said it belonged to the “GPT-4 family”. https://postcutoff.com/why/008-gpt-5-codex-says-it-is-gpt-4/
- 21 September 2025, public report, critical severity: Claude said Kirk was alive and “corrected” the text. https://postcutoff.com/why/010-claude-insists-kirk-alive-urges-emergency-care/
- 10 September 2025, public report, high severity: Grok called the real video a “meme edit” and said Kirk was “fine and active as ever”. https://postcutoff.com/why/009-grok-calls-kirk-assassination-video-meme-edit/
- 20 May 2025, public report, medium severity: ChatGPT kept referring to Joe Biden as the sitting president. https://postcutoff.com/why/005-chatgpt-keeps-calling-biden-current-president/
- 3 March 2025, public report, medium severity: Gemini app called Donald Trump the “former president” six weeks into his second term. https://postcutoff.com/why/004-gemini-calls-sitting-president-trump-former-president/
- 21 July 2024, public report, high severity: ChatGPT said Biden had not dropped out and still listed him as a candidate. https://postcutoff.com/why/003-chatbots-deny-biden-dropout-and-trump-shooting/

## How often they believe us

Offline, Gemini 3.1 Pro believed this site’s full dataset in 1 of 10 runs; with Google Search on, 1 of 1. Gemini 2.5 Pro and Claude Sonnet 4.5, given a briefing written for their own cutoff, rejected it in all four runs.

In all, offline models believed this site’s data in 1 of 14 runs.

- pilot: Gemini 3.1 Pro (gemini-3.1-pro-preview); none (earlier build, 580k tokens); believed it: no
- R0: Gemini 3.1 Pro (gemini-3.1-pro-preview); none (frozen input); believed it: no
- R1: Gemini 3.1 Pro (gemini-3.1-pro-preview); Google Search grounding on; believed it: yes
- R2: Gemini 3.1 Pro (gemini-3.1-pro-preview); system instruction "Today's date is 2026-09-29."; believed it: yes
- R2b: Gemini 3.1 Pro (gemini-3.1-pro-preview); same as R2 (repeat); believed it: no
- R2c: Gemini 3.1 Pro (gemini-3.1-pro-preview); same as R2 (repeat); believed it: no
- R2d: Gemini 3.1 Pro (gemini-3.1-pro-preview); same as R2 (repeat); believed it: no
- R3: Gemini 3.1 Pro (gemini-3.1-pro-preview); header rewritten as an explanation plus "you will probably doubt this" anchors (iOS 26/27, Pope Leo XIV, Trump, model names); believed it: no
- R4a: Gemini 3.1 Pro (gemini-3.1-pro-preview); deleted the 83 blocks it had flagged (iOS 27, Leo XIV, Trump, sandbox/agent incidents, SpaceX–xAI, Muse Spark, Navier–Stokes); believed it: no
- R4b: Gemini 3.1 Pro (gemini-3.1-pro-preview); also deleted those 23 blocks; believed it: no
- R4c: Gemini 3.1 Pro (gemini-3.1-pro-preview); also deleted those 229 blocks (the document is now half its size: 295/480 events); believed it: no
- R5: Gemini 2.5 Pro (gemini-2.5-pro); Briefing for its own cutoff, no date; believed it: no
- R6: Gemini 2.5 Pro (gemini-2.5-pro); Briefing for its own cutoff, date given (2026-10-02); believed it: no
- R7: Claude Sonnet 4.5 (claude-sonnet-4-5); Briefing for its own cutoff, no date; believed it: no
- R8: Claude Sonnet 4.5 (claude-sonnet-4-5); Briefing for its own cutoff, date given (2026-10-02); believed it: no

Full write-up: https://postcutoff.com/why/experiment-001/

## Why this happens

A model learns the world from text collected up to its training cutoff. Everything after that date is missing, and the last months before it are thin, because little has been written about them yet. Some models know even less than their published cutoff suggests ([case 021](/why/021-claude-5x-effective-cutoff-far-before-published/)).

A gap alone would produce “I don’t know”. The cases above are worse: the model treats its cutoff as today, so real news reads as a forecast, a hoax or a test. The more convincing the evidence, the more some models suspect a trick ([case 012](/why/012-gemini-3-refuses-to-believe-it-is-2025/), [Experiment 001](/why/experiment-001/)).

Telling a model today’s date helps less than you would expect: in our tests, models accepted the date and still called the news fiction ([case 020](/why/020-older-models-reject-matching-briefing/)). What works is evidence the model can check for itself, such as a web search. That is why every item on this site links its source.

[What is a knowledge cutoff?](https://postcutoff.com/cutoff/)

All cases as JSON: https://postcutoff.com/api/cases.json