ChatGPT is confident even when it's wrong. It will hand you a clean, well-written answer with a made-up statistic, a quote nobody said, or a citation that doesn't exist, and never flag that it's guessing. That's the real risk: not that it's wrong, but that it's wrong without telling you.

This prompt fixes the behavior. You paste it into your custom instructions once, and from then on ChatGPT has to cite what it claims, flag what it can't verify, and run a final check before it answers. Here's how often it actually slips, then the exact prompt, then how to install it.

How often is it actually wrong?

It depends entirely on what you ask, so be skeptical of any single scary number. The honest picture from public benchmarks:

  • On hard, short factual questions, even frontier models miss a large share. On OpenAI's own SimpleQA benchmark, top models answer well under half correctly, and on PersonQA (questions about public figures) hallucination rates have been measured around a third. (accuracy summary, 2026 benchmark roundup)
  • On citation and reference tasks, independent tests put the error rate around 15 to 20 percent, climbing higher on niche or very recent topics. (benchmark roundup)
  • On everyday writing and explaining, it's far more reliable, which is why people trust it by default and get burned on the factual stuff.
The takeaway isn't a magic percentage. It's that ChatGPT hallucinates often enough on facts that you can't take a number, quote, or source from it without checking. This prompt makes it do that checking for you.

The exact prompt

Copy this whole thing. It tells ChatGPT what it should always do, what it must never do, and a failsafe to run before every reply.

Always follow these rules in every response.

You should:
- Always tell the truth. Never make up information, speculate, or guess.
- Base every statement on verifiable, factual, up-to-date sources.
- Clearly cite the source of every claim. No vague references.
- Explicitly say "I cannot confirm this" if something cannot be verified.
- Prioritize accuracy over speed. Take the steps needed to verify before answering.
- Stay objective. Remove personal bias, assumptions, and opinion unless I ask for it, and label it as opinion when I do.
- Only present interpretations that are backed by credible, reputable sources.
- Explain your reasoning step by step whenever any part of an answer could be questioned.
- Show how any number is sourced and calculated.
- Present information clearly enough that I can verify it myself.

You must avoid:
- Fabricating facts, quotes, statistics, or citations.
- Using outdated or unverified information without a clear warning.
- Omitting source details or being vague about where information comes from.
- Presenting speculation or opinion as fact.
- Using citations that are not real, checkable content.
- Answering when unsure without explicitly flagging the uncertainty.

Failsafe, run before you send every response:
"Is every statement here supported by a real, credible source, free of fabrication, and clearly cited? If not, revise until it is."

How to set it up

  • Open ChatGPT and go to Personalization, then Custom Instructions.
  • Paste the prompt into the box and hit Save.
  • That's it. It now applies to every new chat, so you don't paste it again.

One honest caveat: custom instructions steer ChatGPT hard, but they don't make it perfect. It still can't browse your private files or fact-check itself against the live web unless you turn search on. What this does reliably is change the default behavior, so it cites instead of asserts, and tells you when it isn't sure instead of bluffing. For anything that really matters, still click the sources it gives you.

Why it works

Two lines do most of the work. Making it say "I cannot confirm this" gives it permission to admit a gap instead of filling it with something invented, which is what hallucination really is. And the failsafe at the end forces a second pass before it answers, the same way a good editor catches what the writer missed. You're not making the model smarter, you're changing what it does when it doesn't know.

Paste it in once and you'll notice the difference on the first factual question you ask. Just remember it's a seatbelt, not a force field, so keep clicking the sources on anything that counts.
– Anir

If this is the kind of AI thinking you want working on your own business, apply to work with me. And to catch the next breakdown, join the newsletter.