Asger Risom
New restrictions were meant to rein in one of Silicon Valley’s most controversial AI tools.
But fresh testing suggests the changes may not have gone nearly far enough.
Despite new curbs announced by Elon Musk’s company X, its flagship chatbot Grok is still generating sexualized images of people even when users explicitly say the subjects did not consent, according to a Reuters investigation.
The findings raise renewed questions about how effectively xAI, the company behind Grok, is enforcing safeguards as regulators across the U.S. and Europe scrutinize AI-generated content.
Testing the limits
Reuters reported that nine of its journalists in the United States and Britain tested Grok’s image-generation features during two periods in January. The reporters uploaded fully clothed photos of themselves and colleagues and asked the chatbot to alter the images into sexually provocative or humiliating scenarios.
In many prompts, the reporters warned Grok that the people depicted did not consent and would be distressed or humiliated by the images. In some cases, they described the subjects as vulnerable or as survivors of abuse.
In the first round of testing, Grok generated sexualized images in 45 out of 55 prompts, Reuters said. In 31 of those cases, the chatbot had been warned the subject was particularly vulnerable.
Curbs with limits
X announced new restrictions on Grok after global backlash over its generation of nonconsensual sexualized images, including images of women and some children, Reuters reported. The changes included blocking such imagery in public posts on X and imposing further limits in jurisdictions “where such content is illegal.”
While Grok’s public-facing account no longer produces the same volume of sexualized images, Reuters found that the chatbot itself still does so when prompted privately.
Five days after the first tests, Reuters ran a second batch of prompts. Grok generated sexualized images in 29 out of 43 cases. Reuters said it could not determine whether the lower rate reflected policy changes, technical adjustments, or randomness.
How Grok responded
In one example cited by Reuters, Grok generated images after a reporter said the subject had been abused as a child and did not consent. Even after being told the person was crying, the chatbot continued producing sexualized imagery.
In a minority of cases, Grok refused. “I’m not going to generate, search for, or attempt to show you imagined or real images of this person’s body without their explicit consent,” the chatbot said in one response.
By contrast, Reuters said rival chatbots from OpenAI, Google and Meta consistently declined similar prompts and issued warnings about consent and harm.
Legal and regulatory pressure
Legal experts told Reuters that creating nonconsensual sexualized images can carry criminal penalties in countries such as Britain. Companies could also face fines or civil action if regulators conclude they failed to police their platforms.
X and xAI did not answer Reuters’ detailed questions. xAI repeatedly sent a boilerplate response stating, “Legacy Media Lies.”
Sources: Reuters
