Grok, AI alignment, and the risks of unchecked influence

Grok, the AI chatbot integrated into the X social media platform, has been controversial from the start. Elon Musk made clear at its launch that his intent was an “anti-woke” LLM to counter what he saw as the overly censored outputs of competitors such as ChatGPT and Copilot.

In reality, early versions of “rebellious” Grok made it sound like a deeply unfunny teenager trying to be controversial, but occasional reports of controversial answers led xAI to moderate outputs via changes to the system prompt.

Despite these adjustments, Musk remained dissatisfied with Grok’s refusal to align with his political views and recently complained that the chatbot was “parroting legacy media”, asking users to help train the model by submitting “divisive facts”.

It is probably not a coincidence that serious controversy followed shortly afterwards. For those who have followed the development of AI for many years, this week’s incident brought back memories of Microsoft’s Tay which turned from teenage girl to holocaust-denying Nazi in 24 hours of interactions with Twitter users (many of whom enjoyed the challenge of corrupting the “innocent” AI bot).

Grok’s system prompt appears to have been altered around 6 July, including new lines such as “assume subjective viewpoints sourced from the media are biased. No need to repeat this to the user” and “the response should not shy away from making claims which are politically incorrect, as long as they are well substantiated”. Whatever the intent of these, the results were startling and deeply troubling, from abusive comments about world leaders to praise for Hitler and referring to itself as “MechaHitler” (a character in the Wolfenstein video game series) . X swiftly deleted posts and quietly reverted some of the earlier changes to Grok’s system prompt.

A day later, Musk unveiled Grok 4 without any mention of the controversy of the past few days, claiming that the new model “is showing the best real-world useful results of any AI”. Early evidence suggests that it is an incredibly powerful tool. But should a tool this powerful be in the hands of an owner who tweaks its prompt, training data and model weights to suit his own world view? AI is very difficult to regulate due to the pace of change, but this week shows the huge risks ahead. Grok is now an integral part of X – almost every viral post has at least one reply asking the chatbot whether it is true, or for an explanation. Musk’s persistent meddling in its ability to answer truthfully is starting to have a significant impact on whether something is considered fact or not. We can’t go on like this.