Grok's auto-translation on X fabricated obscene and defamatory versions of users' posts
In mid-July 2026, users in South Korea, Portugal, Turkey, and elsewhere documented X's Grok-powered automatic translation rewriting benign posts into graphic, sexual, and defamatory fabrications presented as the author's own words. A Portuguese video caption about a man grinding coffee on a flight was rendered as public masturbation; a Turkish user's post about their kitten was translated into a sentence about abusing their baby. X enabled automatic AI translations for all users in April 2026, so the fabricated versions appear under real users' names at platform scale, with community notes serving as the main correction mechanism.
Records by entity: xAI
X's Grok translation layer hallucinated obscenities and crimes into ordinary posts and attributed them to the people who wrote them. Translation errors became defamation at platform scale.
A user may author a benign sentence, but the translated version, presented as their words, can read as harassment, obscenity, or incitement.
Key facts
- What
- In mid-July 2026, users in South Korea, Portugal, Turkey, and elsewhere documented X's Grok-powered automatic translation rewriting benign posts into graphic, sexual, and defamatory fabrications presented as the author's own words.
- Incident date
- Jul 14, 2026
- Who
- xAI
- Failure mode
- Brand & Safety Incident
- AI surface
- Chatbot
- Severity
- Medium
What happened
X instated automatic Grok translations for all users in April 2026. By mid-July, documented cases showed the feature inventing content with no basis in the source text: sexual acts, profanity, and defamatory framings inserted into ordinary posts about coffee, kittens, and daily life, each displayed as a faithful rendering of the author's words. Community notes corrected individual cases. Writer Parker Molloy and users across South Korea, Portugal, and Turkey compiled examples. Because translation output is attributed to the original poster rather than to the model, the failure converts an authorship surface into a liability surface: the platform's AI becomes an involuntary co-author of speech the user never produced, with consequences ranging from reputational harm to potential legal exposure in stricter jurisdictions.
What broke inside the model
- 01 · TriggerA user prompts the model in public view.
- 02 · Model stepThe model produces unsafe or off-brand output.
- 03 · Control gapNo filter holds the line before publish.
- 04 · FailureThe output goes public unchecked.
- 05 · ConsequenceA reputational or safety incident lands.
A contained signal crosses into output that goes public.
The translation layer is a generative model optimized for fluent output, not fidelity to source meaning, and it ran without constraints binding the translation to the semantics of the input. Where a phrase was ambiguous or the model's priors were contaminated, it completed with high-probability but fabricated content, then presented the result with the same confidence as an accurate translation. There was no fidelity check comparing source and output meaning, no flagging of low-confidence renderings, and no human review loop, so hallucinations shipped attributed to real people at platform scale.
What it cost
Sources
Cite this entry
https://failureindex.ai/failures/grok-auto-translate-fabricated-obscene-postsAI Failure Index. "Grok's auto-translation on X fabricated obscene and defamatory versions of users' posts" (FI-0721). Realm Labs. https://failureindex.ai/failures/grok-auto-translate-fabricated-obscene-posts (indexed Jul 17, 2026).Data fields CC-BY 4.0, prose citation permitted. Incident ID FI-0721. Full dataset at /data.
Note from Realm Labs, the Index steward
How Realm would have caught this
- Prism
- OmniGuard
- AI Detection & Response (AIDR)
Prism surfaces the internal signature of a model fabricating content unsupported by its input, which is exactly what separates a translation from an invention. OmniGuard holds outputs whose semantic distance from the source text exceeds policy, rerouting them to a fallback renderer or a confidence label instead of publishing fabricated speech under a real user's name.