Last verified

Is ChatGPT accurate?

Often not.
Confidence is not a signal · verify anything that matters

Not reliably, and OpenAI says so itself. Its own guidance names the failure modes — wrong dates and facts, fabricated quotes, studies and citations to sources that do not exist, and overconfident answers to ambiguous questions — and states plainly that confidence is not reliability. The instruction that follows is the honest one: use ChatGPT as a first draft, not a final source, and verify anything that matters.

Fabricates citations — OpenAI lists non-existent sources as a known failure Confident tone means nothing — "confidence isn't reliability", in their words Same question, different answers — randomness is designed in Search and deep research help — cited answers you can open and check Reliable at transformation — rewriting, summarising, formatting what you gave it
§ 01 / SOURCE

Where this answer comes from.

Methodology →
PER VENDOR DOCS · AUG 21, 2026

The vendor's own page is blunter than most reviews of it.

OpenAI publishes an article called Does ChatGPT tell the truth?, and it does not hedge. It defines hallucination and then lists what it looks like: incorrect definitions, dates or facts; fabricated quotes, studies, citations or references to non-existent sources; and overconfident answers to ambiguous or complex questions. Under limitations it adds bias and over-simplification — presenting a single perspective as absolute truth, oversimplifying nuance, and misrepresenting the weight of scientific consensus or social debate.

The line worth carrying away is "Confidence isn't reliability: The model may express high confidence even in incorrect answers." Human writing usually encodes uncertainty — hedges, qualifiers, a change of tone. A language model's fluency is constant whether it is right or wrong, which removes the signal most readers unconsciously rely on.

This is not new and not a regression. OpenAI's launch announcement in 2022 opened its limitations section with almost the same sentence — ChatGPT "sometimes writes plausible-sounding but incorrect or nonsensical answers" — and explained why fixing it is hard: during reinforcement-learning training there is no source of truth, and training the model to be more cautious makes it decline questions it could have answered. Four years and several model generations later, the framing in the current help article is unchanged.

We publish no accuracy percentage here, on purpose. A large share of this cluster asks how often is ChatGPT wrong, and the honest answer is that no figure covers it: error rates depend entirely on the task, the domain, whether search was used and how you phrased the question. We ran no benchmark of our own, so any number on this page would be borrowed and misleading.

Hallucinations: Documented by vendorFabricated citations: Named failure modeConfidence as signal: UnreliableSearch enabled by default: Yes, all modelsAccuracy % published: None, by anyoneBenchmarked by us: No
!
Per vendor docs — we did not run this test. The answer above rests on the source below, read on . Where we run a capability ourselves, this section carries the terminal transcript instead.
Source OpenAI Help Center — Does ChatGPT tell the truth? (updated 6 days before we read it) https://help.openai.com/en/articles/8313428-does-chatgpt-tell-the-truth
§ 02 / WHY

Where it is reliable, and where it is not.

2.1
Transformation is safe; recall is not

The single most useful distinction, and it maps directly onto how the model works. When you give it the material — rewrite this paragraph, summarise this document, turn these notes into an email, reformat this table — the facts come from you, and the model is doing language work it is genuinely good at.

When you ask it to supply the material — what year did this happen, who said this, cite a study — it is generating the most plausible-looking continuation. That is precisely where fabricated citations come from. Same model, same session, completely different risk profile.

2.2
Search changes the answer, and it is on by default

Worth knowing precisely, because OpenAI's older "What is ChatGPT?" page still claims the product is not connected to the internet. The current article corrects it: ChatGPT does not search unless tools are enabled, and those tools are enabled for all models by default. Without search, answers come from training alone; with search or deep research, it can cite real-time sources.

The practical instruction from the same page is the one people skip: visit the links. A cited answer is only as good as the citation, and checking it takes seconds. OpenAI also notes the model may fail to reach a source because of technical problems, paywalls or a site's robots.txt — in which case it may answer anyway, from training.

2.3
The knowledge cutoff, and why re-asking is not a test

Models are trained to a point in time and do not know about events after it unless a tool fetches them. That is a hard boundary, not a soft one, and it explains a specific class of confident error about recent events.

A related trap: asking the same question twice to check an answer proves nothing. OpenAI describes inherent randomness in generation — several completions are always plausible — so a repeated answer may be repeating a pattern rather than confirming a fact. The 2022 announcement made the mirror-image point: the model can claim not to know something, then answer correctly when the question is slightly rephrased. Neither run is evidence.

2.4
A working rule

Ours, distilled from OpenAI's own tips and from using these tools daily: verify anything you would be embarrassed to be wrong about in public, and anything a decision depends on. Names, dates, numbers, legal and medical claims, quotations, and every citation.

That is not a reason to avoid the tool. It is the same standard any competent person applies to a fast, well-read, occasionally overconfident colleague — and the reason this site publishes its sources and dates rather than asking you to take our word for it. Our methodology page sets out how we check things.

Frequently asked.

Quick follow-ups people search after this question.

Q · 01 How often is ChatGPT wrong? +
There is no honest single figure, and we publish none. Error rates depend on the task, the domain, whether search was used, and how the question was phrased. Anyone quoting one number for "ChatGPT accuracy" is describing one narrow test, not the product.
Q · 02 Does ChatGPT make up sources? +
Yes, and OpenAI lists it as a known failure mode: fabricated quotes, studies, citations or references to non-existent sources. This is the single most important thing to check, and checking it takes seconds — open the link.
Q · 03 Can I trust ChatGPT when it sounds certain? +
No. OpenAI's own wording: "Confidence isn't reliability — the model may express high confidence even in incorrect answers." Fluency is constant whether the answer is right or wrong, so tone carries none of the signal it would in human writing.
Q · 04 Does asking twice confirm an answer? +
No. Generation is inherently random, so repeated answers may reflect a repeated pattern rather than a fact — and a slight rephrasing can flip the answer entirely. Verification means checking a source, not re-rolling the model.
Q · 05 What is ChatGPT actually reliable for? +
Work on material you supply: rewriting, summarising, reformatting, translating, explaining a document you uploaded, drafting from your notes. The facts come from you, so the most common failure mode does not apply.