Quick answer
This page gives no single accuracy rate, because none of the sources read for it supplies one that covers every task. OpenAI's terms (effective January 1, 2026) say output may not always be accurate and should not be a sole source of truth. Its help page says to verify quotes, data, technical information or references.
Key takeaways
- OpenAI's terms of use say: Output may not always be accurate. They also say you should not rely on output as a sole source of truth or factual information, or as a substitute for professional advice.
- OpenAI's help page lists three kinds of error it describes: incorrect definitions, dates or facts; fabricated quotes, studies, citations or references to non-existent sources; and overconfident answers to ambiguous or complex questions.
- An OpenAI post from September 2025 reports SimpleQA error rates of 26% and 75% for two named models, which gave no specific answer to 52% and 1% of the short fact questions. The post does not say which ChatGPT product either model powers.
- This site's suggestion: treat anything a customer, a regulator or a bank could hold you to as unverified until a primary source confirms it.
Is ChatGPT accurate?
No source we read gives one accuracy rate for ChatGPT; OpenAI's terms say output "may not always be accurate" and that you should not rely on it as a sole source of truth. This page reports what OpenAI states and for which tasks. It does not say, in its own voice, that ChatGPT is or is not accurate.
Three statements in OpenAI's own words set the frame (terms of use effective January 1, 2026; help page and research post opened on October 7, 2026):
- Terms of use: "Output may not always be accurate." The terms add that "You should not rely on Output from our Services as a sole source of truth or factual information, or as a substitute for professional advice." They also say "You must evaluate Output for accuracy and appropriateness for your use case, including using human review as appropriate, before using or sharing Output from the Services."
- Help page, "Does ChatGPT tell the truth?": "Sometimes, it might sound confident—even when it's wrong."
- Research post, September 5, 2025: "ChatGPT also hallucinates."
The terms also say that use of the services "may, in some situations, result in Output that does not accurately reflect real people, places, or facts." None of these sentences says how often.
What kinds of errors does OpenAI describe?
OpenAI's help page names three kinds: wrong facts, invented sources and overconfident answers. It lists them under "hallucination", defined there as when the model produces responses that are not factually accurate, such as:
- "Incorrect definitions, dates, or facts"
- "Fabricated quotes, studies, citations or references to non-existent sources"
- "Overconfident answers to ambiguous or complex questions"
The same page lists limits. One is a knowledge cutoff: "Knowledge cutoff: The models are trained on data up to a certain point and responses do not incorporate information about events beyond that, unless tools are used." Another: "Confidence isn't reliability: The model may express high confidence even in incorrect answers." So a firm tone tells you nothing about whether a statement is right.
The help page does not say how often these occur.
Is there a measured error rate for ChatGPT?
The one we could read is a table in an OpenAI post, for two named models on short fact questions; it is not a rate for ChatGPT or for business tasks. The September 5, 2025 post introduces it as "an example from the GPT5 System Card". The post does not give the number of questions. OpenAI's figures:
| SimpleQA, as reported by OpenAI | gpt-5-thinking-mini | OpenAI o4-mini |
|---|---|---|
| Abstention rate (no specific answer is given) | 52% | 1% |
| Accuracy rate (right answer) | 22% | 24% |
| Error rate (wrong answer) | 26% | 75% |
What the quantity is, line by line:
- Benchmark: SimpleQA, which its paper (arXiv, 7 November 2024) describes as a benchmark that evaluates "the ability of language models to answer short, fact-seeking questions". The paper says it is "adversarially collected against GPT-4 responses". In our reading, that means the questions are not a sample of everyday questions.
- What counts as an error: each answer is "graded as either correct, incorrect, or not attempted". An error is a wrong answer; declining to answer is counted separately. OpenAI's table lists abstention separately from error. OpenAI's own sentence: "Strategically guessing when uncertain improves accuracy but increases errors and hallucinations."
- Who measured it: OpenAI, on its own models. We found no independent re-measurement we could open, and this page gives none.
- Which models: gpt-5-thinking-mini and OpenAI o4-mini, as of September 2025. The post does not say which ChatGPT product, if any, these models power, so they should not be read as the ChatGPT you open today, and the table says nothing about any other model.
Read the table for what it shows, not as a score. A rate measured on one kind of question does not carry over to other tasks: the paper describes short fact questions and says nothing about prices, tax rules, dates in your contracts or customer emails (this caution is this site's). We looked for OpenAI figures for ChatGPT on business writing tasks and found none on the pages above.
Where can you rely on ChatGPT, and what must you check?
This site's decision rule: the less you can see and confirm yourself, the more you must verify before it leaves your hands. The rule is our suggestion; it rests on OpenAI's instruction, in its help page, to "Use ChatGPT as a first draft, not a final source" and to "Always verify quotes, data, technical information or references to external documents."
The pages we read do not list tasks where ChatGPT is reliable. They list what to verify. So the sorting below is ours.
| What you handle | What the sources say can go wrong | The two-minute check (this site's suggestion) |
|---|---|---|
| Prices, rates, quantities, totals | Help page: "Always verify quotes, data…" and errors in "facts" | Open the supplier's or the government's page and find the number there. Recalculate totals yourself. |
| Dates and deadlines | Help page: "Incorrect definitions, dates, or facts"; knowledge cutoff | Find the date on the official page. Check when that page was last updated. |
| Legal, tax, licensing statements | Terms: not "a substitute for professional advice" | Treat as a question for your accountant, attorney or the agency, not as an answer. |
| Quotes, studies, citations, links | Help page: "Fabricated quotes, studies, citations or references to non-existent sources" | Search the exact quote in quotation marks. Open each link yourself. If you cannot find it, delete it. |
| Customer-facing claims (what a product does, warranties, results) | Terms: "You must evaluate Output for accuracy and appropriateness for your use case, including using human review as appropriate". The sources read do not name product claims specifically. | Compare every claim with your own product sheet or contract. Remove any claim you cannot point to. |
| Rewording text you supplied | The sources do not say. | Read the result against your original for changed facts, numbers or names. |
How do you verify one claim in two minutes?
Pull out the single checkable statement, find the primary source, and compare it word for word. An illustration, not a case from a real business: ChatGPT drafts a paragraph for your website saying a city requires a permit for a sidewalk sign.
- Isolate the claim. Copy the one sentence that could cost you if wrong: "the city requires a permit for sidewalk signs."
- Go to the owner of the rule. Search for the city's own website, not a summary, and find it yourself rather than through a link the chat gave you.
- Find the exact wording. Search the city's pages for the term. If you find it, compare the condition (which streets, which size) with the draft.
- Check the date. Note when the page was last updated.
- Decide. If the source supports the sentence, keep it. If it differs, rewrite from the source. If you find nothing, delete the sentence.
If search is available to you, OpenAI says it "Lets ChatGPT look up current or niche information from the web and provide cited answers", which depends on your plan. The same page still says to check sources "by visiting links directly". Asking ChatGPT whether it is sure is not a check, because "Confidence isn't reliability".
What this means for tools that send text for you
The more a tool sends without a person reading it, the more a wrong sentence reaches a customer. OpenAI's terms say not to rely on output as a sole source of truth. For automations that draft or send messages, put a person between the draft and the send button. See AI agents for small business for what needs a human check, and AI email responder for small business and AI document data extraction for review steps in two common automations.
How was this page made?
This page was drafted with Claude, an AI model made by Anthropic, and checked against each vendor's published material on October 7, 2026. Anthropic is not a vendor covered here: the only vendor described is OpenAI, from its terms of use, one help page and one research post, plus the SimpleQA paper. OpenAI's help pages were read rendered in a browser because they refuse plain downloads. The help page shows "Updated: 2 months ago", with no exact date, and OpenAI can change these pages at any time.
Not covered: other assistants, independent accuracy studies of ChatGPT, and any figure for the model you use today. Next step: pick the one row in the table that matches what you send to customers most often, and add its check to your routine this week.
FAQ
Is ChatGPT accurate enough to use for my business?
This page cannot give a yes or no for all tasks. OpenAI's terms say output may not always be accurate and should not be relied on as a sole source of truth, so the sources we read point to checking, not trusting. This site's suggestion is to match the amount of checking to what happens if the statement is wrong.
How often does ChatGPT make things up?
This page found no rate for ChatGPT as a product. OpenAI's September 5, 2025 research post has a SimpleQA table, a test of short fact questions, with an error rate of 26% for gpt-5-thinking-mini (no specific answer given to 52% of questions) and 75% for OpenAI o4-mini (1%). OpenAI ran it on those two models. It does not give a rate for your questions.
Does turning on search make ChatGPT’s answers correct?
OpenAI's help page says search lets ChatGPT look up current or niche information from the web and provide cited answers, and that it may be available depending on your plan. The same page still tells users to check sources by visiting links directly when accuracy matters. We found no figure for how much search lowers errors.
Does the same checking apply to other AI assistants?
This page reports only what OpenAI says about ChatGPT and does not compare assistants. These are OpenAI's own disclosures; we read no other vendor's pages, and nothing here suggests another assistant is more or less accurate. This site's suggestion is to apply the same checking to the output of any assistant.
Can I ask ChatGPT whether its own answer is right?
OpenAI's help page says the model may express high confidence even in incorrect answers. This site's reading: a confident reply to the question 'are you sure?' is therefore not a check, so use a source outside the chat.