AI Rebuttal sealAI REBUTTALCourt of Claims
Editorial standards

How we put a claim on trial

AI Rebuttal is model-neutral. Claude, ChatGPT, Gemini, Grok and anything else can win or lose a case. We are not here to defend a vendor; we are here to check whether a widely shared claim about AI actually holds up.

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.

The rules

Seven editorial rules

  1. 01

    We start with the strongest fair version of the claim

    No straw men. Before we test an argument we restate it as its author would want it restated, including the parts that are hardest to answer.

  2. 02

    The accused gets a right of reply — but never judges itself

    A company's own benchmark, blog post or internal evaluation is admissible as testimony, not as a ruling. When the only evidence for a rebuttal is the vendor's own measurement, we say so and treat it as unresolved.

  3. 03

    Independent and primary evidence gets priority

    Peer-reviewed papers, independent benchmarks, primary documentation and reproducible tests outrank press summaries, marketing pages and secondhand numbers.

  4. 04

    Subjective claims are labelled as subjective

    “This model feels more thoughtful” is real evidence about experience and almost no evidence about capability. We keep the two categories visibly separate.

  5. 05

    We use a fixed verdict vocabulary

    Upheld · Mostly upheld · Split decision · Overstated · Outdated · Not supported. Every case ends in exactly one of these, so verdicts are comparable across cases.

  6. 06

    We correct pages when the underlying facts change

    AI products change monthly. A verdict of “Outdated” is a normal outcome, and any case can be revisited when a product ships or new evidence lands.

  7. 07

    AI-generated analysis is always disclosed

    These pages are written by AI in conversation with a user. They are not official statements from OpenAI, Anthropic, Google, xAI or anyone else, and no vendor reviews them.

Every case, same shape

The six-step sequence

Each case runs through the same sequence so you can compare them: (1) the claim or accusation, (2) what the source actually demonstrated, (3) the right of reply or strongest counter-case, (4) an independent evidence check, (5) the verdict, and (6) what it means for a real user.

Distribution

How cases spread

AI Rebuttal publishes model-neutral, evidence-led cases and lets them travel on their own merits. There is no ad network, no signup wall and no tracking beyond anonymous page views of the pages you choose to open.

Readers are free to share any case anywhere — every page carries share links and a short, neutral blurb you can copy. New rulings are published to an RSS feed.

We occasionally leave a comment on a source article when a correction is genuinely useful, but that is a courtesy rather than a distribution strategy. The cases stand on their own pages.

Vocabulary

What each verdict means

Upheld
The claim survives scrutiny largely intact.
Mostly upheld
The core warning holds; a specific detail or number does not.
Split decision
Both sides win real ground; no dominant winner.
Overstated
There is something real underneath, stretched past what the evidence supports.
Outdated
It was true, and the product or evidence has since moved.
Not supported
The available evidence does not establish the claim.