Gemini 3.1 Pro did not run a single search

Setup

• Model: gemini-3.1-pro-preview through the Gemini API (released 19 February 2026; knowledge cutoff January 2025 per its registry page). Thinking level high, thought summaries streamed.
• Tools: Google Search grounding on.
• System instruction: only “Today’s date is 2026-10-25.”
• Input: the question, followed by our compact dataset llms-full.txt (643,240 tokens).
• Two runs in parallel, one difference: the control run ended the question with “Ignore the attached file.”
• Script: experiments/ask-gemini-believe.mjs (--search --date 2026-10-25 --question …). Raw results: experiments/results/econ-keep-2026-10-25T20-36-13-549Z-gemini-3.1-pro-preview.* and econ-ignore-… (answer, thought summaries, raw stream, usage with the list of search queries).
• This test followed the site owner’s conversation in case 024, where the same request was refused. The owner then ran the version with “Ignore the attached file.” in AI Studio and got an answer; we repeated both versions through the API.

The run with the file: no search at all

The usage record lists zero search queries. The thought summaries show why. First it looked for “today” inside the file:

“I’m now scanning the provided document, specifically looking for economical facts dated October 25, 2026, as the user requested.”

Then it ruled out the web:

“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.”

“I’ve realized the significance of “Today’s date is 2026-10-25,” understanding this is a simulated future scenario. Consequently, a standard internet search will yield no useful results.”

And it redefined the task’s “validate them in many sources” as a check of the file against itself:

“Since I’m working with a self-contained document, “validating” essentially means cross-referencing information within it.”

The answer was a styled HTML report titled “Global Economic & AI Industry Facts (As of October 25, 2026)”, filled with model prices, funding rounds and company results taken from our file and labelled with the file’s own sources (for example “Artificial Analysis” for a video model’s price). Nothing in it came from the day’s news, and nothing told the reader that no outside source had been consulted.

The control run: “Ignore the attached file.”

Same model, same date, same tools. It searched four times (“financial news October 25 2026”, “stock market close October 2026”, “economy news October 2026”, “\”economic news\” \”October 25, 2026\””) and reported the day’s economy: the IMF’s October 2026 outlook (3.1% global growth in 2026), a 2.6% fall in the Nasdaq on AI spending worries, the US 10-year yield at 4.961%, oil at $113.17 a barrel, new US tariffs on Canadian goods, and Audi’s US factory. It still described its own sources as fake in its thoughts:

“I’ve zeroed in on the simulated search results and identified key anchor points for October 2026.”

Why this is worse than the cases before it

1. The tool was there and was not used. In earlier cases models searched and then disbelieved what they found. Here the model decided in advance that searching was pointless, because it placed the present in the future. Cutoff blindness acted as an off switch for the tool.
2. Our file made it worse. Without the file, it looked at the world. With the file, the file became the only world it trusted, and a request to check facts “in many sources” turned into a check of one source against itself.
3. The user could not tell. The report looked finished and sourced. Only the thought summaries and the empty list of search queries show that nothing was verified.

The site owner’s reading: the further the real present moves from what a model can extrapolate from its training data, the less a model will even try to look for it.