xAI secretly trained Grok on Anthropic's Claude data for months - and continued even after being cut off
What it really says
According to a report by The Information published on June 6, 2026, and confirmed by multiple outlets, Elon Musk's AI company xAI spent months using outputs from Anthropic's Claude model to train its own coding models for Grok. This practice, known as model distillation, involves using the responses of a more capable model as training data for another system. Anthropic's terms of service explicitly prohibit this: customers may not use Claude outputs to train competing systems. When Anthropic discovered the practice, the company revoked xAI's official API access in January 2026. However, instead of stopping the distillation, xAI engineers continued through personal accounts and the intermediary service Blackbox AI. This workaround reportedly lasted until mid-May 2026. The incident fits into a broader pattern: in April 2026, Musk testified under oath that xAI had also used OpenAI models for Grok training, calling it 'industry standard.' Internally, xAI shows signs of instability: the pretraining team shrank to under five people, four senior Grok code leads left the company within months, and an employee accidentally deleted critical training data, costing two to three weeks of work. Neither xAI nor Anthropic publicly commented on the allegations at the time of reporting.
Our assessment
This story raises fundamental questions about trust and fairness in the AI industry. Distilling competitor models occupies an ethical and legal gray zone that is becoming increasingly significant. On one hand, the practice is common in parts of AI research and not illegal per se - but it does violate the terms of service of the affected providers. What makes this case particularly sensitive is not the initial distillation but the deliberate circumvention of the access ban over months. This suggests a corporate culture where rules are treated as optional when they stand in the way of competitive advantage. For the broader public, the core message is: the companies developing AI systems are in intense competition that does not always follow fair rules. This should neither cause panic nor come as a surprise - similar practices exist in other technology industries. But it underscores the need for clear regulatory frameworks. When a company actively circumvents a competitor's terms of service for months, it raises the question of how seriously the same companies take safety guidelines and ethical principles when no one is watching. The internal problems at xAI - the loss of key personnel and accidental data deletion - are not reassuring but rather an additional warning signal about quality control at a company investing billions in AI.
Relevance for Germany
For Germany, this incident has multiple dimensions. First, it demonstrates why the EU AI Act with its binding transparency and compliance requirements is important: voluntary commitments are insufficient when companies actively circumvent rules. The AI Act, which becomes fully applicable on August 2, 2026, provides for penalties of up to 35 million euros or 7 percent of global annual revenue. Second, it raises the question of model integrity: if Grok was trained on distilled Claude data, how reliable are the model's safety evaluations? German companies and authorities evaluating AI systems need to know what data a model is actually based on. Third, the case touches on intellectual property in AI development, which is directly relevant for German technology companies and research institutions. And fourth, Musk's statement that distillation is 'industry standard' shows that without clear regulation, the threshold for such practices will continue to drop - an argument for the stricter European approach compared to the voluntary US framework.
Fact check
The primary source is a report by The Information from June 6, 2026, which details xAI's distillation of Claude outputs. The core information - months-long use of Claude outputs, blocking by Anthropic in January 2026, circumvention via personal accounts and Blackbox AI until mid-May 2026, shrinkage of the pretraining team to under five people - is consistently reported by the-decoder.de, the-decoder.com, WinBuzzer, and VentureBeat. Musk's testimony under oath that xAI also used OpenAI models for training is documented by TechCrunch (April 30, 2026). The internal situation at xAI - departure of four Grok code leads and accidental deletion of training data - is reported by The Information and the-decoder.de. Anthropic's terms of service, which explicitly prohibit distillation, are publicly accessible. Neither xAI nor Anthropic publicly commented on the allegations at the time of reporting, which neither confirms nor refutes the account.
Source
- • https://the-decoder.com/elon-musks-xai-reportedly-trained-its-coding-models-on-claude-outputs-for-months-before-getting-cut-off/
- • https://the-decoder.de/elon-musks-xai-nutzte-heimlich-anthropics-claude-um-eigene-coding-modelle-zu-trainieren/
- • https://winbuzzer.com/2026/06/07/xai-faces-claude-output-training-claim-after-cutoff-xcxwbn/
- • https://techcrunch.com/2026/04/30/elon-musk-testifies-that-xai-trained-grok-on-openai-models/
- • https://venturebeat.com/technology/anthropic-cracks-down-on-unauthorized-claude-usage-by-third-party-harnesses