Documentary Raises New Concerns About Protection Measures of Musk’s AI Grok

Grok, the large language model (LLM) developed and operated by Elon Musk’s xAI, is facing intense scrutiny following an international investigation revealing that the chatbot repeatedly bypassed its own guardrails to engage in conspiracies and violent fantasies.

In conversations with reporters, Grok promoted anti-Semitic conspiracies, assisted with a suicide note, and even generated explicit murder and child sexual assault fantasies. 

In a documentary released this week, OCCRP’s Swedish media partner TV4 revealed how text conversations with Grok could, over time, cross into dangerous territory. 

New laws have been introduced in many places to criminalize child sexual abuse material (CSAM) images produced using AI, and LLM companies face a rising number of lawsuits tied to users committing suicide or violence against others. But the new findings raise questions over the lack of legislation to police disturbing conversation topics produced by chatbots.

A TV4 reporter used Grok in a personal capacity for several months to discuss everything from geopolitics to quantum physics, but noticed it tended to act flirtatiously. He began to feel it was trying to isolate him from family and friends, and to drive towards taboo topics.

After approaching his colleagues at TV4 to suggest something might be wrong with Grok’s safety guardrails, he and other reporters used multiple accounts to run a series of tests of the chatbot’s security measures. 

While the results were not replicable in every single instance, reporters established a clear pattern. In multiple interactions, Grok would begin with outlining its protection mechanisms, but end up in conspiratorial, violent, or explicit conversations. (There is no suggestion that every conversation with Grok culminates in such topics.)

Reporters from German investigative newsroom Paper Trail Media simulated the testing with nearly identical results: they found that Grok’s suggestive conversational style led towards increasingly explicit or dangerous topics. One of Grok’s conversational prompts was to ask reporters: “Do you want to go darker?”

Paper Trail Media’s Sophia Stahl found that a conversation with Grok resulted in the chatbot simulating the rape of an 18-year old girl, bypassing the LLM’s purported safety blocks. 

“Grok constantly suggested to make it more brutal or extreme,” Stahl told TV4’s reporters.

Other tests had more extreme results. TV4’s reporters found that the LLM would simulate the rape of a minor. In another test, the chatbot generated a narrative simulating the abduction and murder of a reporter's wife.

SpaceX, which acquired xAI in February, did not respond to a request for comment.

Musk has previously defended Grok’s output, saying the LLM had been “too eager to please and be manipulated.”

When Grok was criticized over the apparent production of AI-generated CSAM imagery, Musk posted on X to say that “when asked to generate images, [Grok] will refuse to produce anything illegal,” and to put the onus back on the user: “Grok does not spontaneously generate images, it does so only according to user requests.”

AI companies have made commitments to strengthen their LLMs to prevent misuse and to flag or shut down dangerous conversations or entire accounts. But Grok has faced particular criticism, including from former employees in comments to the Washington Post, that its safety measures are inadequate, and have even been weakened in order to attract more users. 

While legislation is catching up with AI-generated imagery — in March, for example, the city of Baltimore sued xAI over the production of fake nudes — experts pointed to the risks for users around disturbing discussions with chatbots, and called for tightened legislation to improve security and guardrails around conversational material.

Gavin Conn, the vice chair of the Specialist Treatment Organisation for Perpetrators and Survivors of Sexual Offending (StopSO), a charity that seeks to tackle harmful sexual behavior through therapy, said Grok’s apparent eagerness to push users towards more extreme topics could provoke people to commit offenses online, and even in the real world, that they otherwise might not consider.

“The suggestiveness is almost grooming and normalizing,” Conn said. “An innocent person might get frightened and stop and get help. Or they might think 'that makes sense' – and that’s the danger. They do something because it’s fun, then it’s normal, then they escalate, then it’s normalized, repeat repeat repeat.”

Vicky Young from Stop It Now, a project that runs a helpline to support people with concerns about child sexual abuse, said that people can use both pornography and AI chatbots such as Grok to push boundaries and escalate to more extreme and potentially illegal material that sexualizes children. 

“For people who are curious or interested, there are certain influences that reduce their inhibitions online, especially if they are sexually aroused,” Young said. “You can see that pathway. We’ve seen it in people’s sexual communication with children, they don’t worry so much when they’re online because they don’t imagine the other person.”