RESEARCH VERIFICATION

AI research verification: have several AIs check the claims before you publish

Paste the draft, the summary or the list of claims. Up to 10 models from different labs check it independently, mark what they would challenge, and a moderator writes one verdict with the disagreements kept in view.

Check the claims before you publish See plans and limits

AI research verification on Council AI means giving several models from different labs the same draft or set of claims and asking each to check it independently, then reading one verdict that separates what most models accept from what they challenge. It helps because a single model can state a wrong figure or misattribute a finding fluently, and asking the same model again tends to repeat it. Models from different labs fail in less correlated ways, so a claim several of them question is worth checking against the source. Several Council models can search the web while they answer; others work from what you paste and their training, and the verdict is not a substitute for reading the primary source. On Ultra, the Library lets the council search your own papers, retrieving the most relevant excerpts rather than reading every document in full.

What independent checking surfaces

Unsupported or overstated claims. Statements that go further than the cited evidence, correlation written as causation, a single study written up as a consensus.
Misattribution. A finding credited to the wrong paper, author or year, or a quote that does not appear in the source it is attributed to.
Numbers that do not add up. Percentages that cannot both be true, a sample size that changes between sections, a unit dropped in a conversion.
Missing counter-evidence. Well-known results that cut the other way and are not mentioned, which a reviewer from a different angle will raise.
Where the models split. A claim half the council accepts and half disputes is the one to verify first, and the verdict names it.

A worked example

Separate the claims from the prose, and ask for a verdict per claim:

Below are 12 factual claims from a draft article, numbered. For each, say whether you would accept it as written, soften it, or reject it, and why in one sentence. Name the kind of source that would settle it. Do not invent citations; say "unsure" where you are.

The verdict groups the claims by how the council judged them, puts any claim the models split on at the top, and keeps each dissenting reason, so you know which three of the twelve to check by hand first.

How it works

Bring the finished work. Paste it or attach it, and say what you want checked. A clear instruction ("find what is wrong with this") gets a review; a vague one gets a summary.
Several models review it independently. Each model in your council, from a different lab, writes its own review. The first round is blind unless reading each other clearly helps, so no review is a copy of another.
One verdict, with the dissent listed. A moderator model reads every review and writes one verdict. Points most models raised come first; a point only one model raised is listed under it instead of being averaged away.

Limits worth knowing

Council size: up to 7 models per question on Plus, up to 10 on Pro and Ultra, picked from 21 models across 9 labs.
Files per message: Plus 4 files of up to 10MB each, Pro 5 of up to 20MB, Ultra 10 of up to 50MB.
Text per message: about 200,000 characters of extracted text per file and across all files in one message (roughly 120 pages). A file over that is refused with a message saying so, never silently cut.
Agreement is a stronger signal, not a guarantee. Several models can share the same mistake, so the verdict lists the dissent rather than hiding it.
Some models in the catalog can search the web while answering; others cannot. Treat any citation a model produces as a lead to check, not a reference.
Ultra's Library searches documents you have stored and retrieves the most relevant excerpts for each question. Every model receives the same excerpts. It does not read every stored document in full, so attach a paper directly when the whole text matters.

When a council is the wrong tool

Replacing a read of the primary source, peer review, or a fact-checker on anything you will publish under your name. The council tells you where to look; it does not look for you.

AI second opinion: The evidence that cross-model checking catches errors one model misses.
Real side-by-side answers: The same questions answered by several models, with the disagreements shown.
Code review: The same independent-review pattern, applied to a diff.

Frequently asked questions

Can AI fact-check a research draft?

It can flag claims that look unsupported, overstated or inconsistent, which tells you where to spend your checking time. It cannot be the final check: models can share a mistake, and some cannot see the live web. Council uses several models from different labs so the claims they disagree on stand out.

Will the models make up citations?

Any language model can produce a plausible citation that does not exist. Ask for the kind of source rather than a specific reference, check anything a model cites, and treat cross-model agreement on a citation as a reason to look it up, not as confirmation.

Does the Library read all my papers?

No. On Ultra, the Library indexes your documents and retrieves the most relevant excerpts for each question, and every model in the council gets the same excerpts. The UI shows which passages were read. For a full read of one paper, attach it to the message.

How much text can I check at once?

About 200,000 characters per message across all files and pasted text, roughly 120 pages. Over that, the upload is refused with a message rather than cut, so split long material into sections.

Can I try it before paying?

Yes. New accounts get 3 councils free, enough to see three models disagree and read the verdict. Plus ($19.99/mo) is the entry plan with every premium model and councils of up to 7 models; Pro ($39.99/mo) runs 10 models with priority routing. Cancel anytime.