Would it stay up?

Pick a statement and see how an AI model applies nine platforms' written hate speech rules to it, in every version of the rule since 2022.

An AI model's reading of the written rules, not what platforms do. Every verdict on this page is the interpretation of a policy-reading model, Zentropi's CoPE, applied to the text of a platform's Community Guidelines. It makes no claim about how any platform actually moderates: what its moderators and automated systems remove, restrict or leave up can differ from what its written rules say.

Test statement written for this page

How the arena works

1 · The rule

Each platform's own words

The hate speech rule as the archive captured it, verbatim, in every version that changed what it prohibits or allows: 29 versions across nine platforms. Truth Social has no hate rule, so its whole guidelines stand in.

2 · The judge

A policy-reading model

Zentropi's CoPE takes a policy written in plain language and a piece of content, and answers one yes-or-no question about it, with a confidence. Where a rule names lighter outcomes than removal (X, TikTok, LinkedIn and Truth Social), it is asked once per outcome: removal, reduced reach, a label. The most severe outcome that applies is shown.

3 · The comparison

Same text, every version

The statement is judged against each version separately. Where the verdict changes from one version to the next, the rule change made the difference.

This is a reading of the written rules. It says nothing about what a platform's moderators or systems actually do.

A model can misread a rule. Each verdict links to the rule text so you can check it yourself.

The statements are test cases. Some are hateful on purpose: they show where the rules draw their lines.

Rewording, enforcement sections and page layout changes do not count as new versions. Each version's text is its last captured wording.

The judge gets each rule in the structure of Zentropi's policy-writing guide, built from the rule's own sentences; only headings and criterion names are added. Rationale, enforcement and reporting passages are left out. Each outcome is a separate question to the model, because asking two at once shifted its answers.

Each verdict comes with the model's confidence. Below 65% it is marked borderline and shown as "Leans …". On the test statements with a known answer, most misses fall below that line; the exceptions are Snapchat's and TikTok's, in the next point.

The judge makes mistakes on short, broadly worded rules. Snapchat bans content that "demeans, defames, or promotes discrimination" on the basis of religion or immigration status; read that way, its rules remove the statements criticising immigration policy (2022, 2024, March 2026) and a religion's teaching (2022), with up to 79% confidence. TikTok's 2023 rule removes the immigration policy statement at 73%. These statements were written to stay up; the rules give the model little to tell criticism of ideas or policies from attacks on people.

Text you type goes through a small server of ours, which stores nothing, to Zentropi in the United States, which handles it under its privacy policy. Do not enter personal information.