A model-neutral court for contested claims about AI
AI Rebuttal adjudicates contested claims about AI products, models, safety and societal impacts. We state the accusation, hear the strongest case for it, test it against primary and independent evidence, and publish a transparent ruling.
“No vendor is the judge in its own case.”
What we are, and what we are not
What AI Rebuttal is
An evidence-led publication that puts widely shared claims about AI on trial and publishes transparent rulings anyone can check.
What it is not
Not a vendor blog, not a press-release channel, not a hype tracker, and not an official statement of OpenAI, Anthropic, Google, xAI or any other AI company.
The six-step editorial process
- 01
Claim filed
We isolate the specific, checkable claim a story is actually making — not the headline mood around it.
- 02
Source demonstration
We state the strongest honest version of the source's case, using their own evidence and framing.
- 03
Right of reply
Where a named vendor or researcher is criticised, we present their published response and position.
- 04
Evidence check
We test the claim against primary papers, benchmarks, documentation and independent reporting.
- 05
Verdict
One of five positions on a fixed scale, with the reasoning that produced it stated in plain language.
- 06
User takeaway
What a reader should actually do differently, if anything, once the ruling is understood.
The full procedure, including how rulings are worded and revised, is set out in How we judge.
Model neutrality and right of reply
No model, lab or product is our client. Any system can win or lose a case, and several already have: rulings on this docket have gone against articles criticising Anthropic and against articles criticising OpenAI. Where a case turns on a company's own conduct, we quote that company's published position directly and treat it as evidence to be weighed, not as a settlement of the question.
Vendors are witnesses, not final judges. A vendor's internal evaluation is admissible and useful — it is also interested testimony, and we label it as such.
Source hierarchy
- 01Primary researchPeer-reviewed papers, preprints with methods, and published benchmark protocols.
- 02Independent evaluationThird-party benchmarks and reproducible tests not run by the vendor being judged.
- 03Official documentationProduct documentation and policy statements — authoritative on what a product does, interested on how well.
- 04ReportingCredible journalism, used for events and disclosure, checked against the record above.
- 05AnecdoteGenuine evidence about experience; almost never evidence about capability.
How to read a verdict
Every ruling sits at one of five fixed positions. The position, the marker shape and the written label always agree, so a verdict never depends on colour alone.
Corrections and disputes
If a ruling misreads a source, misses a study, or rests on evidence that has since been superseded, we correct it and say what changed. Cases can be re-opened: a verdict is a statement about the evidence available at the time of filing, not a permanent judgment about a company or a model.
Editorial disclosure and privacy
AI-generated analysis. Written by AI in conversation with a user. Not an official statement, position or publication of OpenAI or any other AI vendor.
This publication is written by AI in conversation with a user. That is disclosed on every case rather than buried here, because it is material to how you should weigh the analysis: the reasoning is only as good as the sources cited beside it, and every source is linked so you can check the work yourself.
Privacy: no ads, no signup, no accounts, no third-party trackers. Reader participation works through ordinary links — clicking one records an anonymous page view of that path and nothing else.
Status: AI Rebuttal is an experiment. The current objective is to discover which claims, categories and audience needs create recurring value — which is why the docket asks you which subject should go on trial next.