Latest results:

How we score Editorial standards
AI Truth Bench Menu

Is AI accurate?

AI is often right but not reliably accurate. In the EBU/BBC study, 45% of news answers from four AI assistants had at least one significant issue. NewsGuard found 10 leading chatbots repeated false news claims 35% of the time. Use AI for drafts and leads, and check anything that matters.

Latest results:

Accuracy by tool
Four measuring instruments in a row on a bench: a dial gauge, a spirit level, a thermometer and a caliper, each against a stack of sheets.

Accuracy by tool

Each page below gives one verdict, the published results behind it with dates and sources, and a short routine for checking that tool's answers. Verdicts use four words: Reliable, Mostly reliable, Mixed and Unreliable. How we score AI accuracy explains how each verdict is reached.

See them ranked
Show

Showing all tools

  • A vernier caliper gripping the sides of a blank paper speech bubble.
    Mixed

    Is ChatGPT accurate?

    ChatGPT had a significant issue in 36% of news answers in the EBU/BBC study and repeated false news claims 40% of the time in NewsGuard's August 2025 audit.

    Evidence
  • A two-pan balance scale holding two matching paper speech bubbles, one pan sitting slightly lower.
    Mixed

    Is Gemini accurate?

    Gemini was worst of four assistants in the EBU/BBC study at 76%, mostly on sourcing, but second best of 10 in NewsGuard's audit at 16.67%.

    Evidence
  • A long folded letter laid flat with a set square, a steel straightedge and a pencil aligned along it.
    Mostly reliable

    Is Claude accurate?

    Claude repeated false news claims 10% of the time, the lowest of 10 chatbots in NewsGuard's audit, and Claude Opus 5.5 scores 48 on the AA-Omniscience index, second highest.

    Evidence
  • A blank speech bubble pinned to a board with threads running to five small source cards around it.
    Mixed

    Is Perplexity accurate?

    Perplexity was the best citer in two news studies, wrong on 37% of Tow Center source lookups, but repeated false news claims 46.67% of the time in NewsGuard's audit.

    Evidence
  • Navigator's dividers and a brass compass resting on a paper chart.
    Mixed

    Is Microsoft Copilot accurate?

    Copilot had a significant issue in 37% of news answers in the EBU/BBC study and repeated false claims 36.67% of the time in NewsGuard's audit.

    Evidence
  • A stopwatch and a thermometer beside a ribbon of paper tape spilling from a small tape machine.
    Unreliable

    Is Grok accurate?

    Grok 3 named the wrong source for 94% of news excerpts in the Tow Center test, and 154 of its 200 citations led to error pages.

    Evidence
  • A depth gauge with a plumb line measuring a tall stack of paper sheets.
    Unreliable

    Is DeepSeek accurate?

    DeepSeek failed 83% of NewsGuard's news prompts in January 2025 and misattributed 115 of 200 sources in the Tow Center test.

    Evidence
  • A paper speech bubble under a glass bell jar on a lab bench, with a small dial gauge beside it.
    Mixed

    Is Meta AI accurate?

    Meta AI repeated false news claims 40% of the time in NewsGuard's August 2025 audit, tied with ChatGPT, and Meta has admitted hallucinations about breaking news.

    Evidence
  • A short summary card with a jeweller's loupe on it, resting on a fan of longer source pages.
    Mixed

    Are Google AI Overviews accurate?

    AI Overviews answered 91% of SimpleQA questions correctly in the NYT and Oumi analysis, but 56% of answers were not supported by their cited sources, as reported by Futurism.

    Evidence

What the tests measure

No single test measures "accuracy". Each one checks a different skill, so a tool can do well on one and badly on another. Gemini is the clearest case: last in one news study, second in another audit.

  • News answers judged by journalists. The EBU/BBC study had 22 public service media in 18 countries review 2,709 answers. 81% had some issue, and sourcing was the most common significant fault at 31%.
  • Finding the source of a news excerpt. The Tow Center test gave eight AI search tools 200 excerpts each. Collectively they were wrong more than 60% of the time.
  • Repeating known false stories. NewsGuard found the average rate rose from 18% in August 2024 to 35% in August 2025, as chatbots stopped declining news questions.
  • Staying faithful to a document. The Vectara leaderboard checks summaries for added content. Even the best model on the board, as of 2026-10-22, added unsupported content to 1.8% of summaries.
  • Hard knowledge questions. AA-Omniscience scores wrong answers as worse than no answer. At launch, most models gave more wrong answers than right ones.

Vectara and AA-Omniscience test the models behind the apps, not the apps themselves. Our tool pages say so wherever those results appear. The full model table is on AI hallucination rates by model.

No AI assistant in these tests was right every time, so the safe habit is the same for all of them: check what matters.

How to read these verdicts

Every tool page carries one of four words, set from the dated evidence on that page.

  • Reliable

    At or near the top in every independent test that includes the tool, with errors rare enough that only high-stakes answers need a check.

  • Mostly reliable

    Best or near the top in most independent tests that include it, with one clear weak area named on its page.

  • Mixed

    Strong in some independent tests and weak in others, or middle-of-the-pack across them. Answers need checking.

  • Unreliable

    At or near the bottom in most independent tests that include it, or a pattern of invented sources. Treat every answer as a lead only.

A verdict is a summary, not a guarantee. Read the results table on each page to see what was measured and when.

How we score

Questions

Short answers, with links to the dated evidence.

Is AI accurate enough to rely on?

Not without checking. In the EBU/BBC study of 2025-10-21, 45% of news answers from ChatGPT, Copilot, Gemini and Perplexity had at least one significant issue. AI is useful for drafts, summaries and finding leads. For facts you will publish, submit or act on, confirm each one in a primary source first.

Which AI makes the fewest mistakes?

Claude has the strongest overall record in published tests. It repeated false news claims 10% of the time in NewsGuard's August 2025 audit, the lowest of 10 chatbots, and Claude Opus 5.5 scores 48 on the AA-Omniscience index, second highest. Perplexity is the best at citing news sources. The full order is on our most accurate AI ranking.

Why do AI tools give different answers?

Each tool runs a different model, searches different pages and is tuned to answer or hold back in different ways. The same tool can also give different answers to the same question. That is why one assistant can top one test and trail in another, as Gemini did in the EBU/BBC study and NewsGuard's audit.

How should I check an AI answer?

Open every source the AI gives and confirm the page says the same thing. Search the key name, number or quote yourself. Check figures and dates in the original document. Look up legal or academic citations in their own databases. Our five-step routine on how to fact-check an AI answer covers each step with free tools.

Can you trust AI?

You can trust AI to be useful, not to be right every time. Results vary widely by tool and task: Claude repeated false claims 10% of the time in NewsGuard's audit, while Grok 3 named the wrong source 94% of the time in the Tow Center test. Treat an AI answer as a lead to confirm, never as the final source.

Where to go next on the bench.

  • Six measuring cylinders filled with dark blue liquid to different levels, one marked with a yellow band.
    Ranking

    The most accurate AI assistants

    Which AI is most accurate? ChatGPT, Claude, Gemini, Perplexity and others ranked by published accuracy and hallucination results, with sources.

  • A paper speech bubble standing on a desk casts a shadow that does not match its shape, examined with a magnifying glass and a ruler.
    Hallucinations

    AI hallucinations, explained

    What AI hallucinations are, why chatbots invent facts and sources, how often it happens by tool, and the habits that stop you being misled.

  • A clipboard checklist with a fountain pen ticking a box, next to a magnifying glass and a reference book.
    Guide

    How to fact-check an AI answer

    A five-step routine to check what ChatGPT, Gemini or any AI tells you: claims, sources, numbers, dates and quotes, with free tools for each.

  • A hand tally counter resting on a strip of graph paper with a line of plotted dots.
    Data

    AI hallucination rates by model

    How often leading AI models hallucinate, from published leaderboards and studies, with the test, date and source behind every rate.