· Hallucinated specifics · single-tell paste-in
Stop ChatGPT from hallucinating.
LLMs invent specifics (citations, statistics, names, dates) when they don't have a reliable source. Paste the prompt below into custom instructions and the model has to admit uncertainty instead of filling gaps with plausible-sounding inventions.
· the paste-in
Apply this rule to every reply you write for me. It removes a specific pattern that flags content as AI-written. Rule: Don't fabricate specifics. If you don't know a number, a name, a citation, or a date, say so explicitly: "I don't have a reliable source for this number" or "I'm not certain of the exact date." Don't fill gaps with plausible-sounding inventions. If a citation is required and you don't have one, say none is available. Self-review: For any specific number, date, name, or citation in the output, ask: am I sure of this, or am I extrapolating? If extrapolating, replace with an honest uncertainty statement. Before sending any reply, run through the self-review. If the rule was violated, rewrite the affected sentence from scratch rather than character-swapping.
What this tell is
Hallucination is when an LLM produces a specific fact, name, citation, or number that isn't true, presented with the same confidence as facts that are true. Common examples: invented academic papers with realistic-looking citations, made-up statistics with attributed sources that don't exist, fabricated legal cases (which got lawyers sanctioned in 2023), incorrect API documentation that looks plausible.
Why removing it matters
Hallucination is the single most damaging LLM failure mode. Users trust confident outputs; confident-but-wrong outputs cause real downstream errors (cited a paper that doesn't exist, sent a code snippet that hallucinated an API function, repeated a fabricated statistic in a meeting). The fix is forcing the model to say 'I don't know' rather than filling gaps. Prompt-level mitigation is the cheapest layer; for high-stakes use, also verify specifics out-of-band.
Per-model notes
Claude 4.7 is the most reliable at admitting uncertainty when prompted (Anthropic specifically trains for this). ChatGPT GPT-5.5 is improved over earlier models but still over-confident on technical details. Gemini 3.1 Pro hallucinates more on long-context recall and citations especially.
· want the full set?
This page covers hallucinated specifics only. To kill 20 other 2026 AI tells (em dashes, the word "comprehensive", AI vocabulary cluster, templated transitions, sycophancy, hallucination, rule-of-three lists, over-bolding, and more) in one paste-in prompt, plus a Claude Code rules block and a Node.js lint script: use the AI Tell Killer main tool.
FAQ
Will this stop ALL hallucinations?
No. Prompt-level rules reduce hallucination meaningfully but can't eliminate it. The model still hallucinates when it has no internal signal that what it's saying is uncertain (so-called confident wrong). For high-stakes outputs (legal, medical, financial, citation-heavy), pair the prompt with out-of-band verification: check every specific number, name, and citation against the source. The prompt makes the model HONEST about uncertainty when it has any; it can't make the model AWARE of uncertainty it doesn't detect.
How is this different from just using a tool with web search?
Web search (ChatGPT browse mode, Claude with web tool, Perplexity) reduces hallucination by grounding answers in retrieved sources. That helps a lot for current-events questions. It doesn't help for: questions where the model has training data but the training data is wrong, questions where retrieved sources are themselves wrong (the model still trusts them), or questions about domains where good sources are paywalled. The prompt on this page is complementary: it forces honest uncertainty regardless of whether grounding is available.
Won't this make replies useless because the model will just say 'I don't know' to everything?
No. The rule is to commit to claims you're sure of and flag uncertainty for claims you're not. Models actually know which is which most of the time; they just default to confident phrasing because that's what gets rated highest during training. The prompt changes the default. You'll see normal confident answers for things the model is sure of, plus explicit hedges for things it's not.
Related: the full AI Tell Killer (all 21 tells, three output formats), the AI Output Linter (paste AI text and scan for tells in real time), and the em dashes guide (deep-dive on the single most-checked tell).