“ChatGPT can make mistakes. Check important info.” That line has sat under the ChatGPT message box for years, and it remains the most accurate single statement about the product. The question is how often, on what, and whether it matters for the thing you are asking.
The research now gives fairly precise answers. Below are ChatGPT accuracy statistics from eight studies and reports, grouped by the kind of question asked, with the method and date for each. At the end is a simple grid for reading any accuracy claim you come across.
Accuracy Depends on the Question
The biggest mistake in most coverage is quoting one accuracy number for ChatGPT as if it were a fixed property of the product. It is not. The same model can score above 90% on one test and below 20% on another, because the tests measure different things.
Four variables drive the result. The first is whether a source is provided. Summarizing a document you pasted in is far easier than answering from memory. The second is whether web search is on. Search can fix outdated knowledge and add new mistakes at the same time. The third is the question’s format. Multiple choice inflates scores compared with open-ended answers. The fourth is whether the question contains a false premise. Models tend to go along with the framing they are given.
Keep those four in mind and the numbers below stop looking contradictory.
Short Factual Questions
OpenAI publishes accuracy data in the system cards it releases with major models, and two benchmarks matter most.
SimpleQA is a set of about 4,000 short factual questions with single correct answers. In its August 2025 GPT-5 system card, OpenAI reported that gpt-5-thinking hallucinated on 40% of SimpleQA questions without web access, compared with 52% for GPT-4o. That is a large improvement and still means a wrong answer on close to half of hard trivia questions when the model has no way to look things up.
PersonQA asks about public facts on real people. The April 2025 system card for o3 and o4-mini showed hallucination rates of 33% and 48%, double and triple the 16% of the older o1. OpenAI noted that the newer models made more claims in general. A few months later, it said GPT-5 with thinking was about 80% less likely than o3 to make a factual error on a set of real-world prompts.
Web search changes the picture. The same GPT-5 system card reported far lower hallucination rates on factual prompts when the model could browse, because it could look up an answer instead of reconstructing it. That is why the ChatGPT most people use today, with search turned on for many queries, performs better on simple facts than closed-book benchmarks suggest. The catch, covered in the citations section below, is that search adds a different class of error: the model may find the right page and still misread or misattribute it.
For a business, the person-facts number is the one to watch. When a buyer asks ChatGPT about your founder, it is running a PersonQA-style query against your reputation.
News and Current Events
News is where ChatGPT accuracy has been tested most rigorously by outsiders, because publishers have a direct stake.
In February 2025, the BBC had its journalists review answers from ChatGPT, Copilot, Gemini, and Perplexity to questions based on BBC stories. More than half of all answers had significant issues.
The follow-up study, published in October 2025 by the European Broadcasting Union with the BBC, was far larger: 22 public service media organizations, 18 countries, 14 languages, and more than 3,000 responses. Across all four assistants, 45% of answers had at least one significant issue and 20% had major accuracy problems. ChatGPT, Copilot, and Perplexity each had significant issues in roughly a third of responses, while Gemini was worst at 76%. Reviewers noted ChatGPT answers that included numbers neither of the cited sources contained.
A Stanford-led evaluation run daily from February 9 to 22, 2026 tested six chatbots on 2,100 questions drawn from fresh BBC reporting. GPT-5 answered 81.3% of multiple-choice questions correctly. In free-response form, its accuracy fell to roughly 65%. On adversarial questions built on false premises, GPT-5 was correct only 19% of the time, going along with the premise in about four of five cases.
The Reuters Institute’s 2025 Digital News Report found 7% of online news users use AI assistants for news each week, rising to 15% of under-25s. Those users are getting answers with the error rates above.
Citations and Sources
Citations make answers look verified. The research says they deserve less trust than that.
In March 2025, the Tow Center for Digital Journalism at Columbia published a test of eight AI search tools. Researchers gave each tool excerpts from real news articles and asked it to identify the headline, publisher, date, and link. Collectively, the tools answered incorrectly more than 60% of the time. ChatGPT Search got 134 of 200 wrong, and it signaled uncertainty in very few of those cases. Perplexity performed best at 37% wrong, and Grok 3 worst at 94%.
The 2026 Stanford-led news study found a separate gap. GPT-5 cited sources in 97.8% of answers, second best of the six models tested, yet ranked fifth on accuracy. The researchers found little relationship between how often a model cited sources and how often it was right.
The practical lesson is that a link next to a claim does not mean the link supports the claim. Clicking through is still the only check that works.
Specialized Domains: Law and Medicine
Exam-style tests make ChatGPT look strong in professional fields. In 2023, researchers writing in PLOS Digital Health found an early version of ChatGPT scoring at or near the passing threshold on all three steps of the US Medical Licensing Exam without special training. Later models have scored far higher on the same kinds of tests.
Real-world questions are a different story. Stanford’s RegLab found general-purpose chatbots hallucinating on 69% to 88% of specific questions about federal court cases in 2024 research, with OpenAI’s GPT-3.5 at the better end of the range. Legal research tools built on top of these models still hallucinated on more than 17% of queries in Stanford’s preregistered follow-up.
The pattern repeats in medicine. Answering a board-style multiple-choice question is easier than giving the right advice to a specific patient with incomplete information. Treat professional answers from ChatGPT as a starting point to verify, never a final answer.
Has ChatGPT Gotten More Accurate?
On most measures, yes, with exceptions worth remembering.
The clearest improvement is in grounded tasks. Summarization error rates on Vectara’s leaderboard have fallen across successive OpenAI models, and the GPT-5 family sits well below the GPT-4 era on that test. Closed-book factual accuracy has improved too, with gpt-5-thinking’s SimpleQA error rate 12 points lower than GPT-4o’s.
The clearest exception is the 2025 reasoning models. o3 and o4-mini hallucinated more than o1 on PersonQA, a regression OpenAI disclosed itself. The lesson is that capability gains do not move every accuracy metric in the same direction, and a new release can be better at math and worse at facts about people.
News accuracy improved between the BBC’s two rounds of research, but by less than many expected. A third of ChatGPT’s news answers still had significant issues in the EBU’s 2025 study, and the 2026 Stanford-led work showed how fragile accuracy becomes under false premises.
So the trend is real, and it is uneven. Anyone who tells you ChatGPT is now “accurate” or “still unreliable” without naming the task is flattening a mixed picture.
A Test You Can Run on Your Own Topic
Published ChatGPT accuracy statistics describe averages across thousands of questions. The accuracy that matters to you is narrower: how well ChatGPT answers questions about your company, your products, or your field. You can measure that in an afternoon.
Write 25 questions that a real buyer, candidate, or reporter might ask. Mix the types on purpose. Include ten simple facts, such as when the company was founded, where it is based, and who leads it. Include five comparison questions, such as how you differ from a named competitor. Include five current-events questions about recent launches or news. Include five questions with a false premise built in, such as asking why the company closed an office it never had.
Ask each question in a fresh chat, once with search on and once with search off if your plan allows it. Record every answer in a spreadsheet and mark each one as correct, partly correct, wrong, or declined. Note whether a citation appeared and whether that citation supported the claim.
The results usually follow the research. Simple facts with search on score best. False-premise questions score worst, because the model tends to accept the framing and invent an explanation. Comparison questions often leak competitor details into answers about you.
Repeat the test every quarter with the same 25 questions. The trend line tells you whether your published facts, press coverage, and structured data are reaching the model, which is the closest thing to a ChatGPT accuracy score for your own name.
How to Read Any ChatGPT Accuracy Claim
Every month brings new ChatGPT accuracy statistics, some from labs, some from researchers, some from vendors with something to sell. We sort them with a tool we call the Accuracy Claim Grid, which asks four questions before trusting any number.
What model and date? ChatGPT is a product that changes underneath its name. A 2024 number about GPT-4o says little about GPT-5. Was a source document or web search available? Grounded and search-enabled tests produce better numbers than closed-book ones. What was the question format? Multiple choice and yes-or-no formats flatter the model. Did the test include false premises or adversarial questions? Those reveal how the model behaves when users are confused, which is often the real-world case.
Then ask what kind of question you care about. If you run a business, the question is how ChatGPT describes you to buyers, which is a mix of person facts, company facts, and current events. That is the hardest category for the model and the one where your own published information carries the most weight. Clear, consistent facts on your site, in press coverage, and in the databases these systems read are the most direct way to move ChatGPT’s accuracy on the one subject that matters to you.