The best AI for research is the one you can check

In a 2026 study of 14 models doing research with tools, the best ones produced working links more than 94% of the time, and the linked pages were relevant more than 80% of the time. Yet only 39 to 77% of their claims were actually accurate (Onweller et al., Cited but Not Verified, May 2026).

That gap is the whole story of AI research tools. The visible part, a tidy list of sources that open, has got very good. The invisible part, whether each source says what the sentence next to it says, hasn't kept up.

Other studies find the same thing from different angles. Salesforce researchers auditing generative search and deep-research systems found citation accuracy between 40% and 80%, with large shares of statements unsupported by the sources cited beside them (DeepTRACE, September 2025). An earlier study of four generative search engines found only 51.5% of generated sentences were fully supported by their citations (Liu, Zhang and Liang, 2023).

So when someone asks which AI is best for research, our answer starts with a different question: which one makes it fastest to find out where it's wrong?

"Research AI" covers three quite different things, and their failure modes differ.

The 2023 citation studies used models that are now retired, and fabrication rates have fallen since. We still include them because the failure they describe, a real-sounding reference that doesn't exist, keeps reappearing in court. In the UK's Al-Haroun case, 18 of the 45 authorities cited did not exist, and the High Court said freely available tools such as ChatGPT "are not capable of conducting reliable legal research" (Ayinde v Haringey; Al-Haroun v QNB, June 2025).

The vendors are candid about this too. OpenAI's deep research launch post said it can "hallucinate facts or make incorrect inferences", may struggle to tell authoritative information from rumour, and often fails "to convey uncertainty accurately". We found no independent evaluation of Claude's Research mode's citation accuracy, which is a gap, not a clean bill of health.

Our position: for research, we'd use deep research to find material and a human to cite it. Let the tool read 200 pages; don't let it be the last reader of the five that matter.

Leaderboards like Humanity's Last Exam score closed-ended questions with one right answer. Research isn't like that, so the useful test is one you build yourself.

Pick a question you've researched properly in the last year: one where you know the key sources, the main numbers and the one paper everyone forgets. Run it through each tool you're considering, with the same wording. Then score the report on five things:

The last row is the most telling. OpenAI's own prompting guide warns that "absence of evidence shouldn't automatically become a factual 'no'" (GPT-5.5 guide). A tool that tells you something doesn't exist because it couldn't find it will mislead you on exactly the questions where you most need help.

If you're choosing between the two most common assistants, our Claude vs ChatGPT for research comparison runs through their features in detail.

Accuracy differences between flagship research modes are small next to the differences in how checkable they are. These features are worth more than a few points on a benchmark.

• An editable plan. ChatGPT's deep research proposes a research plan you can review and change before it runs, and lets you restrict or prioritise specific sites (OpenAI help centre). Narrowing it to sources you trust removes a whole class of error.
• Your own sources in scope. Claude's Research searches the web plus enabled connectors such as Google Workspace, and runs "multiple searches that build on each other" (Claude help centre). Grounding a report in your own documents beats grounding it in the open web.
• Time and depth. Claude can research for up to 45 minutes before delivering a cited report (Anthropic, May 2025). Longer runs find more, but the Cited but Not Verified study found accuracy fell about 42% as tool calls grew from 2 to 150. Depth needs more checking, not less.
• A prompt that asks for doubt. Anthropic's research guidance suggests asking Claude to "develop several competing hypotheses", "verify information across multiple sources" and "track your confidence levels" (Claude prompting best practices).

One caution from Anthropic's own help pages: during Research, connected tools can run without approval, so turn off any connector that can write to an external app. For the prompt side of this, see prompts that make AI cite sources; for the two products, Claude Research and how it compares with ChatGPT's.

A deep-research report is long, well formatted and full of footnotes, which is exactly why it gets skimmed. We've written about that effect in the trust trap.

A two-minute check

1. Find the three claims your conclusion depends on. Not the background; the ones that, if wrong, change what you'd say.
2. Open each one's citation and search the page for the claim. Read the sentence either side.
3. Check one number against its original source, not the article quoting it. Figures drift with each retelling.
4. Search for one thing the report says doesn't exist or wasn't found. Absence is where research tools are weakest.
5. Ask the tool: "Which claims in this report rest on a single source, and which have no source?"

That covers the claims that carry weight. It doesn't cover the rest, and in a 20-page report the rest is most of it. Count handles the full pass: connect it to Claude or ChatGPT and it splits the report into claims, labels each as a citation, assumption, model knowledge or hallucination, and shows the reasoning chain behind it, so you can see which sources a conclusion actually stands on.

For academic work, a wrong citation isn't a small error. It's the thing reviewers and examiners check first.

Three habits help. Use AI to map a field and suggest search terms, then run those searches in the databases your discipline trusts. Treat every AI-supplied reference as unconfirmed until you've opened the paper itself, not its abstract page. And keep a log of which claims came from which tool, because you'll be asked.

What we wouldn't do is let any tool, from any vendor, write the reference list. In 2023, only 7% of the references ChatGPT-3.5 produced for medical papers were both real and accurate (Bhattacharyya et al., Cureus). Models have improved a great deal since. The cost of the one invented reference hasn't.

What is the best AI for research?

What is the best AI for academic research?

Is Perplexity accurate for research?

Can AI do a literature review?