AI safety researchers from OpenAI, Anthropic, and other organizations are speaking out publicly against the \u201creckless\u201d and \u201ccompletely irresponsible\u201d safety culture at xAI, the billion-dollar AI startup owned by Elon Musk.<\/p>\n
The criticisms follow weeks of scandals at xAI that have overshadowed the company\u2019s technological advances.<\/p>\n
Last week, the company\u2019s AI chatbot, Grok, spouted antisemitic comments<\/a> and repeatedly called itself \u201cMechaHitler.\u201d Shortly after xAI took its chatbot offline to address the problem, it launched an increasingly capable frontier AI model, Grok 4<\/a>, which TechCrunch and others found to consult Elon Musk\u2019s personal politics for help answering hot-button issues.<\/a> In the latest development, xAI launched AI companions<\/a> that take the form of a hyper-sexualized anime girl and an overly aggressive panda.<\/p>\n Friendly joshing among employees of competing AI labs is fairly normal, but these researchers seem to be calling for increased attention to xAI\u2019s safety practices, which they claim to be at odds with industry norms.<\/p>\n \u201cI didn\u2019t want to post on Grok safety since I work at a competitor, but it\u2019s not about competition,\u201d said Boaz Barak, a computer science professor currently on leave from Harvard to work on safety research at OpenAI, in a Tuesday post on X.<\/a> \u201cI appreciate the scientists and engineers @xai but the way safety was handled is completely irresponsible.\u201d<\/p>\n I didn’t want to post on Grok safety since I work at a competitor, but it’s not about competition.<\/p>\n I appreciate the scientists and engineers at @xai<\/a> but the way safety was handled is completely irresponsible. Thread below.<\/p>\n \u2014 Boaz Barak (@boazbaraktcs) July 15, 2025<\/a><\/p><\/blockquote>\n Barak particularly takes issue with xAI\u2019s decision to not publish system cards \u2014 industry standard reports that detail training methods and safety evaluations in a good faith effort to share information with the research community. As a result, Barak says it\u2019s unclear what safety training was done on Grok 4.<\/p>\n OpenAI and Google have a spotty reputation themselves when it comes to promptly sharing system cards when unveiling new AI models. OpenAI decided not to publish a system card for GPT-4.1,<\/a> claiming it was not a frontier model. Meanwhile, Google waited months after unveiling Gemini 2.5 Pro to publish a safety report<\/a>. However, these companies historically publish safety reports for all frontier AI models before they enter full production.<\/p>\n Techcrunch event<\/p>\n \n\t\t\t\t\t\t\t\t\tSan Francisco Barak also notes that Grok\u2019s AI companions \u201ctake the worst issues we currently have for emotional dependencies and tries to amplify them.\u201d In recent years, we\u2019ve seen countless stories<\/a> of unstable people developing concerning relationship with chatbots<\/a>, and how AI\u2019s over-agreeable answers can tip them over the edge of sanity.<\/p>\n Samuel Marks, an AI safety researcher with Anthropic, also took issue with xAI\u2019s decision not to publish a safety report, calling the move \u201creckless.\u201d<\/p>\n \u201cAnthropic, OpenAI, and Google\u2019s release practices have issues,\u201d Marks wrote in a post on X<\/a>. \u201cBut they at least do something, anything to assess safety pre-deployment and document findings. xAI does not.\u201d<\/p>\n xAI launched Grok 4 without any documentation of their safety testing. This is reckless and breaks with industry best practices followed by other major AI labs.<\/p>\n If xAI is going to be a frontier AI developer, they should act like one. \ud83e\uddf5<\/p>\n \u2014 Samuel Marks (@saprmarks) July 13, 2025<\/a><\/p><\/blockquote>\n The reality is that we don\u2019t really know what xAI did to test Grok 4. In a widely shared post in the online forum LessWrong, one anonymous researcher claims that Grok 4 has no meaningful safety guardrails<\/a> based on their testing.<\/p>\n Whether that\u2019s true or not, the world seems to be finding out about Grok\u2019s shortcomings in real time. Several of xAI\u2019s safety issues have since gone viral, and the company claims to have addressed them with tweaks to Grok\u2019s system prompt.<\/a><\/p>\n OpenAI, Anthropic, and xAI did not respond to TechCrunch\u2019s request for comment.<\/p>\n Dan Hendrycks, a safety adviser for xAI and director of the Center for AI Safety, posted on X<\/a> that the company did \u201cdangerous capability evaluations\u201d on Grok 4. However, the results to those evaluations have not been publicly shared.<\/p>\n \u201cIt concerns me when standard safety practices aren\u2019t upheld across the AI industry, like publishing the results of dangerous capability evaluations,\u201d said Steven Adler, an independent AI researcher who previously led safety teams at OpenAI, in a statement to TechCrunch. \u201cGovernments and the public deserve to know how AI companies are handling the risks of the very powerful systems they say they\u2019re building.\u201d<\/p>\n What\u2019s interesting about xAI\u2019s questionable safety practices is that Musk has long been one of the AI safety industry\u2019s most notable advocates<\/a>. The billionaire leader of xAI, Tesla, and SpaceX has warned many times<\/a> about the potential for advanced AI systems to cause catastrophic outcomes for humans, and he\u2019s praised an open approach to developing AI models.<\/p>\n And yet, AI researchers at competing labs claim xAI is veering from industry norms around safely releasing AI models. In doing so, Musk\u2019s startup may be inadvertently making a strong case for state and federal lawmakers to set rules around publishing AI safety reports.<\/p>\n There are several attempts at the state level to do so. California state Sen. Scott Wiener is pushing a bill <\/a>that would require leading AI labs \u2014 likely including xAI \u2014 to publish safety reports, while New York Gov. Kathy Hochul is currently considering a similar bill<\/a>. Advocates of these bills note that most AI labs publish this type of information anyway \u2014 but evidently, not all of them do it consistently.<\/p>\n AI models today have yet to exhibit real-world scenarios in which they create truly catastrophic harms, such as the death of people or billions of dollars in damages. However, many AI researchers say that this could be a problem in the near future given the rapid progress of AI models, and the billions of dollars Silicon Valley is investing to further improve AI.<\/p>\n But even for skeptics of such catastrophic scenarios, there\u2019s a strong case to suggest that Grok\u2019s misbehavior makes the products it powers today significantly worse. <\/p>\n Grok spread antisemitism around the X platform this week, just a few weeks after the chatbot repeatedly brought up \u201cwhite genocide\u201d<\/a> in conversations with users. Musk has indicated that Grok will be more ingrained in Tesla<\/a> vehicles, and xAI is trying to sell its AI models to The Pentagon<\/a> and other enterprises. It\u2019s hard to imagine that people driving Musk\u2019s cars, federal workers protecting the U.S., or enterprise employees automating tasks will be any more receptive to these misbehaviors than users on X.<\/p>\n Several researchers argue that AI safety and alignment testing not only ensures that the worst outcomes don\u2019t happen, but they also protect against near-term behavioral issues.<\/p>\n At the very least, Grok\u2019s incidents tend to overshadow xAI\u2019s rapid progress in developing frontier AI models that best OpenAI and Google\u2019s technology, just a couple years after the startup was founded. <\/p>\n\n
\n\t\t\t\t\t\t\t\t\t\t\t\t\t|
\n\t\t\t\t\t\t\t\t\t\t\t\t\tOctober 27-29, 2025\n\t\t\t\t\t\t\t<\/p>\n\n