also searched as: is chatgpt accurate · how accurate is chatgpt · is chatgpt reliable · is chatgpt always right · how reliable is chatgpt · is chatgpt trustworthy · can chatgpt be wrong
ChatGPT is right most of the time and wrong often enough that you should check anything you act on. Its errors cluster in sources, quotes, numbers and recent events, and they sound as sure as the right answers.
Trust it for explaining, drafting and brainstorming, where you would spot a wrong line.
Double-check it when the answer holds a number, a name, a date, a quote or a link.
Don't rely on it if the answer goes into a court filing, a medication decision or a citation and you will not open the source yourself.
sources checked 2026-10-13 · by dan scutaru
| who tested it | what they found |
|---|---|
| OpenAI's own test, Oct 2026 (company claim) | 5.6% of answers had an error GPT-6 Luna, the free-plan model, on hard fact-heavy prompts (4.5% on medical, legal and finance prompts); graded by an AI, and OpenAI says these are not production rates. source |
| Washington State University, Mar 2026 | 80% right, 73% consistent Free ChatGPT (GPT-5 mini) judged 719 published hypotheses true or false, ten times each. source |
| EBU and BBC, Oct 2025 | 36% of news answers flawed Journalists at 22 broadcasters rated 678 free ChatGPT (GPT-4o) answers to news questions, May to June 2025. source |
| Columbia Tow Center, Mar 2025 | 134 of 200 sources wrong ChatGPT Search had to name the headline, publisher, date and URL for 200 article excerpts. source |
| University of Oxford, Nature Medicine, Feb 2026 | under 34.5% got the condition 1,298 UK adults used GPT-4o, Llama 3 or Command R+ on ten scenarios and did no better than people using their usual sources. source |
| US District Court, SDNY, Jun 22, 2023 | $5,000 sanction, fake cases In Mata v. Avianca, lawyers filed court opinions invented by ChatGPT, then stood by them when the judge asked. source |
ChatGPT still hallucinates, and OpenAI says so. Its own research gives the reason: models graded only on right answers learn that a guess scores better than "I don't know". In OpenAI's example, a model that almost never abstained was wrong on 75% of a quiz; one that declined half the questions was wrong on 26%.
Size matters too. OpenAI says larger models tend to do better on factuality, and the free plan runs Luna, the smallest current model.
No, not as the source itself. Use it to find sources, then read them. Search mode helps because it shows what it read, but OpenAI's help center says its results and citations can still be incomplete, outdated or incorrect.
Skipping that step is public and costly. One database of court rulings on AI-invented citations lists 2,184 decisions, 864 involving lawyers, across all AI tools, not only ChatGPT.
Not for deciding what to do. In the Oxford trial the models knew the medicine; the authors blamed the conversation, meaning what people told the chatbot and what they took from it. OpenAI's terms say its output is no substitute for professional advice. Use it to prepare questions for your doctor, and give it every detail up front. Whether to hand it your health records is a privacy question, covered on our page on whether ChatGPT is safe.
It means OpenAI does not vouch for the answer. Its terms say not to rely on output as a sole source of truth, and to check it before you use or share it.
Not on purpose, but it guesses. OpenAI's research says training rewards a confident guess over admitting doubt, so a wrong answer sounds like a right one.
A little, by OpenAI's own figures. Paid plans add GPT-6 Sol, the larger model; on past user-flagged failures Sol erred in 13.5% of answers, Luna in 18.7%.