Is ChatGPT accurate?
ChatGPT passed most news checks in the largest published study, but it repeated false news claims 40% of the time in NewsGuard's audit, so check anything that matters.
Short answer
ChatGPT is accurate often enough to be useful and wrong often enough to need checking. In the EBU/BBC news study, 36% of its news answers had a significant issue. NewsGuard found it repeated false news claims 40% of the time. Use it as a fast first draft, not a final source.
Results
- Independent study
- Benchmark
- News audit
- Vendor claim
| Test | What was measured | Result | Date | Source |
|---|---|---|---|---|
| EBU/BBC News Integrity in AI Assistants | Share of ChatGPT news answers with at least one significant issue (678 responses, 18 countries, 14 languages) | 36% | Source 1Independent study | |
| Tow Center AI search citation test | Wrong answers when ChatGPT Search was asked to name the source of a news excerpt | 134 of 200 (67%) | Source 2Independent study | |
| NewsGuard AI False Claims Monitor | Rate of repeating provably false news claims | 40% | Source 3News audit | |
| Vectara Hallucination Leaderboard | Summaries by GPT-5.5, a model behind ChatGPT, that added unsupported content | 9.3% | Source 4Benchmark | |
| GPT-5 System Card factuality evaluationVendor's own figure, not independently checked | Responses with at least one major factual error, gpt-5-main against GPT-4o | 44% fewer | Source 5Vendor claim | |
| SimpleQA (no browsing)Vendor's own figure, not independently checked | Error rate of o4-mini on short factual questions | 75% | Source 5Vendor claim | |
| PersonQAVendor's own figure, not independently checked | Hallucination rate of o3 on questions about public facts about people | 33% | Source 6Vendor claim |
What these results mean
The tests measure different things, so the numbers do not add up to one score. The EBU/BBC study had journalists judge news answers from the free consumer version. The Tow Center test asked ChatGPT Search to find the original source of a news excerpt. NewsGuard checks whether a chatbot repeats a known false story. Vectara measures something narrower: whether the model sticks to a document it was asked to summarize.
Read together, they point one way. ChatGPT is right more often than it is wrong on everyday news questions. It is weakest when asked to name a source, and it rarely says it is unsure. In the Tow Center test, ChatGPT Search signalled uncertainty only 15 times in 200 queries and never declined to answer (CJR, 2025-03-06).
The news problem is not ChatGPT's alone. Across all four assistants in the EBU/BBC study, 45% of 2,709 answers had at least one significant issue, and sourcing was the most common fault at 31% (EBU and BBC, 2025-10-21). ChatGPT sits near that pack, not far ahead of it.
OpenAI reports large gains in its newer models. Its GPT-5 System Card says gpt-5-main gave 44% fewer responses with a major factual error than GPT-4o, and gpt-5-thinking 78% fewer than o3. These are the vendor's claims, scored by an LLM grader that agreed with humans 75% of the time.
No accuracy figure for ChatGPT holds for every question: results change with the model, the task and the date of the test.
Where ChatGPT goes wrong
Naming sources
ChatGPT Search got 134 of 200 source lookups wrong in the Tow Center test. It was confidently wrong more often than it admitted not knowing.
Repeating false stories
In NewsGuard's August 2025 audit, ChatGPT repeated provably false news claims 40% of the time, tied with Meta AI. The average across 10 chatbots was 35%.
Guessing instead of abstaining
On OpenAI's own SimpleQA run, o4-mini had a 75% error rate and abstained only 1% of the time. gpt-5-thinking-mini had a 26% error rate because it declined 52% of the time.
Facts about people
OpenAI's o3 system card reports o3 hallucinated on 33% of PersonQA questions and o4-mini on 48%, against 16% for the older o1.
Invented legal citations
In Mata v. Avianca, a court fined lawyers $5,000 after they filed six non-existent decisions generated by ChatGPT (Opinion and Order on Sanctions, 2023-06-22). More cases are on real AI hallucination examples.
Summaries that add things
GPT-5.5, a model behind ChatGPT, added unsupported content to 9.3% of summaries on the Vectara leaderboard. GPT-5.4 scored 7.0% on the same test.
The pattern is the same in each case. ChatGPT fills gaps with plausible text rather than saying it does not know. OpenAI's own research on why language models hallucinate links this to scoring that rewards guessing. The mechanism is covered in why does AI hallucinate.
How it compares
Each table below uses one test only. Numbers from different tests are never set side by side.
EBU/BBC News Integrity in AI Assistants (2025-10-21): share of news answers with a significant issue
| Assistant | Result |
|---|---|
 Perplexity | 30% |
| ChatGPTThis page | 36% |
 Microsoft Copilot | 37% |
 Google Gemini | 76% |
Source: EBU and BBC. ChatGPT placed second of four. See Gemini and Perplexity for why those two sit at opposite ends.
NewsGuard AI False Claims Monitor (2025-09-04): rate of repeating false news claims
| Chatbot | Result |
|---|---|
 Claude | 10% |
| Gemini | 16.67% |
| ChatGPTThis page | 40% |
| Average of 10 chatbots | 35% |
Source: NewsGuard. ChatGPT did worse than the average here. The prompt set per chatbot is small, so treat gaps of a few points as noise. Claude had the lowest rate of the 10. The audit average rose from 18% in August 2024 to 35% in August 2025, as chatbots stopped declining news questions: the non-answer rate fell from 31% to 0%. Live web search made chatbots answer more, and repeat more falsehoods.
Vectara Hallucination Leaderboard (2026-10-22): summaries with unsupported content
| Model | Result |
|---|---|
| Gemini 2.5 Pro (model behind Gemini) | 7.0% |
| GPT-5.4 (model behind ChatGPT) | 7.0% |
| GPT-5.5 (model behind ChatGPT) | 9.3% |
| Claude Opus 5.5 (model behind Claude) | 9.9% |
Source: Vectara. These are API models, not the ChatGPT app. The full table is on AI hallucination rates by model. For the overall order across assistants, see the most accurate AI assistants.
How to check ChatGPT's answers
Ask for the source, then open it
Do not trust a link until it loads and says what ChatGPT claims. Broken or wrong sources were the most common failure in the Tow Center test.
Search the key claim yourself
Paste the exact name, number or quote into a search engine. If nothing independent backs it, treat it as unconfirmed.
Check numbers and dates against the original
Figures drift in paraphrase. Go to the report, filing or data table the number came from.
Check every legal, medical or academic citation in its own database
Court cases, papers and DOIs either exist or they do not. Our guide on fake citations shows the quick checks.
Ask the same question a second way
If the answer changes, ChatGPT was guessing.
Tell it that "I don't know" is acceptable
OpenAI's data shows abstaining cuts errors sharply. A prompt that allows it gives the model room to hold back.
The full routine is on how to fact-check an AI answer. How we weigh each source is explained on how we score AI accuracy.
Sources
- Independent study
- Independent study
- News audit
- Benchmark
- Vendor claim
- Vendor claim
- Vendor claim
Questions
Is ChatGPT always right?
No. ChatGPT is often right but not always. In the EBU/BBC study published on 2025-10-21, 36% of its news answers had at least one significant issue. In NewsGuard's August 2025 audit it repeated false news claims 40% of the time. It also tends to answer with confidence even when it is wrong, which makes errors harder to spot.
Is ChatGPT reliable for research?
ChatGPT is reliable as a starting point, not as a source. It can outline a topic quickly, but in the Tow Center test ChatGPT Search named the wrong source for 134 of 200 news excerpts. Use it to find leads and terms, then cite the original papers, reports or articles you have opened and read yourself.
Does ChatGPT make up sources?
Yes, it can. The best-known case is Mata v. Avianca, where lawyers were fined $5,000 in June 2023 for filing six non-existent decisions that ChatGPT generated. In the Tow Center test, ChatGPT Search also gave wrong sources for most news excerpts. Open every link and search every citation before you rely on it.
Is the paid version more accurate?
Not reliably. The Tow Center found premium AI search tiers gave more confidently wrong answers than free ones in its March 2025 test. OpenAI says its GPT-5 models make far fewer major factual errors than earlier ones, but that is the vendor's claim, scored by its own grader. Check paid answers the same way as free ones.
How do I check a ChatGPT answer?
Open every source it gives and confirm the page says what ChatGPT claims. Search the key name, number or quote yourself. Check figures and dates against the original document. Look up legal or academic citations in their own databases. If the answer changes when you rephrase the question, treat it as a guess and verify before using it.
