Last November, Elon Musk publicly announced the release of Grok, a “mischievous” chatbot positioned as the new competitor to challenge ChatGPT. After a short wait, the model developed by xAI was launched for sale last week, and users have already begun to detect numerous flaws.
Jax Winterbourne, a security tester, posted a screenshot on their X profile showing Grok denying a task, accompanied by a concerning message: “I’m afraid I cannot fulfill that request as it goes against OpenAI’s usage policy.” Even though Grok isn’t associated with OpenAI or ChatGPT, why did the company’s name appear in that warning?
The explanation provided by Grok’s developers is simple. Igor Babuschkin, an xAI employee, responded to Winterbourne stating that the web contains “copious” results directly extracted from ChatGPT. When gathering internet data to train Grok, these were inadvertently incorporated. Babuschkin mentioned this came as a surprise when they first noticed it: “For what it’s worth, the issue is very peculiar, and now that we’re aware of it, we’ll ensure future versions of Grok won’t have this problem.” And just to be clear, the employee concluded by affirming that “no OpenAI code was used in Grok’s development.”
However, this explanation doesn’t satisfy the experts. Simon Willison, an AI researcher, explains that he has previously seen this behavior in other language models. According to Willison, it’s highly likely that Grok “has been fine-tuned with datasets generated through OpenAI APIs or extracted from ChatGPT itself.”
At this stage, one could say that AI is feeding itself. As language models have become more capable, it has become increasingly common to fine-tune these models with synthetic data, generated by artificial intelligence. This incident has further fueled the rivalry between OpenAI and X, a conflict that Musk initiated with his constant criticisms of Sam Altman’s company.
Source: ArsTechnica
