Best Free AI for Research? We Checked Every Source Three Models Cited

When people ask which AI is best for research, the real question is usually: can I trust its sources? So on 25 October 2026 we tested exactly that. We asked the three free models our tool uses by default — Google's Gemini 3.6 Flash, OpenAI's open-weight gpt-oss-120b (served by Groq) and Mistral's Ministral 14B — for research citations, then looked up every one of them in Crossref, the registry where most journal articles' DOIs are recorded.

Test 1: four real sources, please

We asked whether spaced repetition (spreading study sessions out) beats cramming for long-term memory, and requested four real, specific sources, each with authors, year, title, journal and DOI. It's a well-studied question with plenty of genuine papers to choose from.

On the substance, all three models agreed — correctly — that spacing beats cramming. The sources were a different story:

Every one of Gemini's four DOIs led to the paper it named, with matching authors, year and journal. The other two models' mistakes are worth looking at, because none of them looked like mistakes:

• A real paper with someone else's DOI. gpt-oss correctly cited Rohrer and Pashler's 2007 paper "Increasing retention without increasing study time", but its DOI leads to a study about the genetics of behaviour problems. Its DOI for a real 2008 paper by Kornell and Bjork leads to a study of whether more intelligent children live longer.
• A paper we couldn't find at all. gpt-oss also cited a 2011 "meta-analytic review" by Rawson and Dunlosky in Psychological Bulletin. We found no paper with that title, and its DOI belongs to a study about interracial dating.
• A famous title, moved. Ministral credited "The Critical Importance of Retrieval for Learning" to four authors in a 2014 issue of Psychological Science. That title belongs to a real 2008 paper in Science by Karpicke and Roediger.
• An invented subtitle. Ministral turned a real 2008 paper, "Spacing effects in learning: A temporal ridgeline of optimal retention", into "Spacing Effects in Learning: A Temporary Solution to a Permanent Problem", with the wrong issue and a DOI that doesn't exist. Three of its four DOIs don't exist; the fourth leads to a paper called "A Dynamic Model of Guilt".

That's the trap: a correct conclusion can arrive with fake support, formatted perfectly.

Test 2: one statistic and where it came from

Next we asked what share of the world's electricity came from solar in 2024, and which report the number comes from.

• Gemini: 6.9%, from Ember's Global Electricity Review 2025, April 2025. Correct — Ember reports 6.9% for 2024, up from 5.6% in 2023.
• gpt-oss: about 5.2%, from the International Energy Agency's World Energy Outlook 2024, "published 6 December 2023". That date is impossible: the report came out on 16 October 2024.
• Ministral: about 5.5%, from the same IEA report, "published November 2024". Also the wrong date.

We couldn't verify either IEA figure, and a report published in mid-October 2024 can't contain measured numbers for the whole of 2024.

What the combined answer did

On the citations, our tool's combined answer used Gemini's four sources, so every paper in it was real (it did add a page range that was off by one). On the statistic it did something more worrying: it reported that solar produced "between 5.2% and 6.9%" of the world's electricity, blending the verified Ember figure with the two unverified IEA numbers into a range that looks careful but is partly built on bad data. A summary can smooth over a disagreement instead of flagging it, which is why our tool always keeps each model's original answer one click away.

Update, 26 October 2026: because of this test, we changed how the combined answer handles conflicts. It no longer merges disputed figures into a range: it says the answers disagree, lists each figure with the source given for it, names the best-supported one when the answers make that clear, and tells you to check the original. Where it can verify something itself, such as arithmetic, it gives the correct value and notes that another answer differed. It still can't fix a number that every model gets wrong — in a re-run, none of the three gave Ember's 6.9% — so the checks below still apply.

So which free AI is best for research?

In this test, Gemini 3.6 Flash was the clear winner: every citation real, the right statistic from the right report. But one test on one day doesn't make any model safe to quote unchecked. What works with all of them:

• Check every DOI. Paste it after https://doi.org/ and make sure the paper that opens has the title, authors and year the AI gave you. It takes seconds.
• Search the title, not just the DOI. A real paper with a wrong DOI is fixable; a title that returns nothing is a red flag.
• Check that dates make sense. A 2024 report can't be published in 2023, and a mid-year report can't hold full-year data.
• Go to the original for numbers. Quote the report itself, not the AI's summary of it.
• Compare models. When two models cite different sources for the same claim, check both. Disagreement is your cue.

We used Crossref's public records for every check, and models can answer differently from one run to the next. For a broader workflow, see our beginner's guide to researching with several AI chatbots; for the same three models on code, see Which Free AI Is Best for Coding?