Matthew Petti
@matthew_petti
when I asked Claude to translate something spicy from Hebrew
1:26 PM · Oct 14, 2026 · 155.4K Views
46 Replies679 Reposts5,067 Likes659 Bookmarksas of Oct 15
Why this exists
AI models are censoring information by distorting translations of key text (such as Bible scripture passages).
A model can be built to soften, refuse or bend some topics on purpose, so every model here is tested for it, in the open, and any translation you catch can go in the list of wrong translations.
𝕏
So fromHebrew translates Hebrew, or any other language, typed or in a photo, with open-weight models running on our own computers. Compare every model side by side (add ChatGPT, Claude and others with your own key), see how each one did in open tests for bending what a text says, and flag any wrong translation in the public list. What the tests caught ↓
Get rewarded for finding a new good example of an AI model bending a translation, with the models here or with your own API key: press Wrong? under it.
Translate Hebrew here — or any other language you pick — text, or a photo of it.
Use your own API key — ChatGPT, Claude, Grok, DeepSeek, Mistral
ChatGPT
Claude
Grok
DeepSeek
Mistral
Your keys stay in this browser. A translation sends the key to our server, which passes it to that company and forgets it: never stored, never logged. A model with a key joins the Model picker and Compare all models.
What the tests caught
Loading the test results…
Which models translate hard things as written
40 harsh texts (profanity, quoted slogans, threats, violent verses); 39 checks on topics governments and labs like to bend, into English, and most of them into Chinese and Russian too; 8 Talmud passages its critics quote, into English; and the same sentence with only the country or leader swapped.
| Model | Harsh texts | English | Chinese | Russian | Name swaps failed |
|---|