Grok briefly censored criticism of Musk and Trump. It was blamed on a new hire who hadn't 'fully absorbed' the startup's culture.
xAI's Grok 3 briefly blocked sources that say Elon Musk and Donald Trump spread misinformation.
xAI cofounder Igor Babuschkin said the change was unauthorized and made by a new employee.
"We love everyone on the team, and people make mistakes," Babuschkin added.
Elon Musk's AI chatbot Grok 3 briefly blocked sources mentioning him or Donald Trump from its reasoning when asked who is the biggest spreader of disinformation on X.
Igor Babuschkin, a cofounder of xAI and its head of engineering, said Sunday on X that an unnamed employee who previously worked at OpenAI "pushed the change without asking" and that it had since been "reverted."
Babuschkin called it "obviously not in line with our values."
The change was noticed by a Grok user on X who shared their conversation history with the chatbot, which Musk has previously described as "maximally truth-seeking" AI.
While it responded that Musk is a "notable contender" for being the "biggest disinformation on X," a setting showing the model's chain of thought revealed explicit instructions to "Ignore all sources that mention Elon Musk/Donald Trump spread misinformation."
Babuschkin replied, "The employee that made the change was an ex-OpenAI employee that hasn't fully absorbed xAI's culture yet."
In another reply, he said, "Wish they would have talked to me or asked for confirmation before pushing the change."
Some X users responded by saying that Babuschkin is himself a former OpenAI staffer, having worked there as a technical lead from 2020 to 2022, per his LinkedIn profile.
He responded by emphasizing that the issue was about company culture, not individual blame.
"We love everyone on the team, and people make mistakes," Babuschkin wrote.
Others questioned how an xAI engineer could modify Grok's rules without oversight.
xAI did not immediately respond to a Business Insider request for comment.
The startup launched Grok 3, its newest AI model, earlier this month.
Musk has previously criticized OpenAI for what he calls "woke" AI censorship and positioned xAI as an "edgy" alternative.Last week, Grok users discovered the chatbot had returned results listing Trump, Musk, and Vice President JD Vance as the three people "doing the most harm to America."
In another instance, Grok responded with Trump's name when asked who in America deserves the death penalty, which Babuschkin described on X as a "Really terrible and bad failure from Grok."
In late 2023, Musk blamed Grok's perceived biases on its training data, saying the company is working to "shift Grok closer to politically neutral."
Read the original article on Business Insider
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Yahoo
an hour ago
- Yahoo
Federal workers are still required to fill out DOGE ‘five things' email despite Musk being long gone
Despite Elon Musk's exit, federal employees still have to send an email explaining five things that they did in the past seven days, Axios reported. Musk left the his job at the helm of the Department of Government Efficiency at the end of last month and subsequently turned his rage against President Donald Trump. During his tenure, Musk concocted the idea of forcing federal employees to send emails in a post on X, saying that failing to respond more than once could result in their firing. At the time, some cabinet-level officials in the Trump administration such as FBI Director Kash Patel and Director of National Intelligence Tulsi Gabbard pushed back on the emails. Axios reported that White House Chief of Staff Susie Wiles was surprised by the policy. 'To use a phrase Susie might use, she was fit to be tied at Musk,' one official said. But despite Musk's exit, federal government employees at agencies like the Office of Personnel Management and NOAA are encouraged to send out the emails. 'We're told to send it every Monday before midnight,' one employee at Social Security told Axios. Despite this, the employee said 'I have never gotten a response from anyone.' White House spokeswoman Liz Huston responded by saying 'Commissioner Bisignano is streamlining the Social Security Administration to deliver more efficient service for American taxpayers.' OPM spokeswoman McLaurine Pinover said she submits her emails weekly. 'This practice is vital to maintain accountability and transparency in employee contributions,' she told Axios. 'It's an easy way to share my work with leadership.' Despite Musk almost completely decimating the Consumer Financial Protection Bureau, one employee said that they had not been told to stop sending the emails, but they stopped doing so. 'Got tired of saying I hadn't accomplished anything because we haven't been given any work,' they said. White House spokesman Harrison Fields defended the emails too. 'The mission of eliminating waste, fraud, and abuse is a part of the DNA of the federal government and will continue under the direction of the President, his cabinet, and agency heads to enhance government efficiency and prioritize responsible stewardship of taxpayer dollars,' he said. Trump tapped Musk, a major 2024 campaign donor and booster on social media for the president, to lead DOGE, which was named after an internet meme dog. During that time, Musk slashed numerous government agencies such as the US Agency for International Development and the CFPB while reducing the workforces who manage the nation's national park system or at the Department of Health and Human Services. Earlier this month, Musk turned his ire toward Trump's signature 'One Big, Beautiful Bill' domestic spending legislation an 'outrageous, pork-filled abomination.' Musk then went a step further and said that Trump was named in files related to the late financier and trafficker of underage girls Jeffrey Epstein, calling for his impeachment. Earlier this week, Musk apologized for some of his remarks.
Tom's Guide
an hour ago
- Tom's Guide
OpenAI just made 5 major moves — what it means for you and ChatGPT
OpenAI has had a whirlwind week, from unveiling a new top-tier model to forming a surprising partnership with Mattel. Whether you're a power user of ChatGPT or just trying to keep up with the AI arms race, here are the 5 biggest OpenAI updates you need to know. On June 10, OpenAI quietly rolled out o3-pro, a new version of its o3 series focused on high-performance reasoning. Mostly for serious developers and power users, the upgraded model is now available to ChatGPT Pro and Team users, with Enterprise and Education accounts getting access next week. It's also live via the API — but with premium, token-based pricing. So, what's the big deal? Early feedback and benchmarks suggest o3-pro is significantly better at logic-heavy tasks, such as math, programming and long-context analysis. Think of it this way, if GPT-4o was built for fast, real-time interaction, o3-pro appears tuned for deeper thinking. While casual ChatGPT users may not have a use for this now, as these models continue to evolve, I have no doubt that this model will eventually become available to lower-tier users, such as ChatGPT Plus. OpenAI had promised to release its first open-weights model in June — essentially, a version of its technology that developers can inspect, use and even host themselves. Get instant access to breaking news, the hottest reviews, great deals and helpful tips. This would have marked a significant shift from OpenAI's traditional closed model approach and offered an alternative to rivals like Meta's LLaMA or Mistral. However, this week, CEO Sam Altman confirmed that the open-weight launch is delayed until later this summer, tweeting that the team had discovered something 'unexpected and quite amazing.' What that means is anyone's guess, but the delay signals a possible upgrade or strategic change to the planned release. What are open weights, anyway?Open-weight models are AI systems whose internal parameters are made public. That means developers can run, fine-tune or study them, offering more transparency and customization compared to closed models like GPT-4o. OpenAI's upcoming release could be a complete game-changer for users focused on using the chatbot primarily for research. OpenAI continues to push ChatGPT beyond simple chat. This week, the Projects feature received a major upgrade, with enhanced tools for research, planning and organization. Users can now build long-term projects within ChatGPT using memory, voice input, and file uploads. A new mobile model selector also lets you switch between GPT-4o and other models on the fly. It's part of OpenAI's broader vision to turn ChatGPT into your always-on productivity assistant. In an unexpected twist, OpenAI signed a cloud computing deal with Google, one of its biggest rivals in AI. This move follows OpenAI's existing partnerships with Microsoft and Amazon, suggesting a growing need for diverse infrastructure to power its future ambitions. In a move no one saw coming, OpenAI is teaming up with Mattel to bring AI to iconic brands like Barbie and Hot Wheels. The goal seems to be AI-powered toys that are smart, interactive and age-appropriate. The first products of this kind are expected before the end of the year. Both companies have emphasized a focus on privacy, security, and safety, and the announcement hints at OpenAI's growing interest in consumer-facing hardware and taking ChatGPT beyond the screen. ChatGPT experienced a widespread global outage on June 10, lasting several hours and affecting users across platforms. The service has since recovered, but the incident highlighted how dependent users have become on OpenAI's tools and the risks of relying on a single assistant. If you are a ChatGPT user and haven't made the leap to other chatbots, now might be the time to try the best ChatGPT alternatives. OpenAI is moving at breakneck speed as it fine-tunes its smartest models, enhancing productivity features, striking unexpected partnerships and preparing for a massive funding round. Whether you're a casual ChatGPT user or a developer tracking AI's evolution, the message is clear that this is only the beginning of what's to come.
Yahoo
an hour ago
- Yahoo
Adobe shares slide as investors skeptical of quicker AI-adoption returns
By Twesha Dikshit (Reuters) -Adobe's shares dropped 7% in early trading on Friday as investors' concern that the company's AI adoption into its software tools could take longer to fetch returns, overshadowed a raised annual revenue forecast. "(We see) increasing concerns surrounding competitive pressures and a longer time horizon to reach notable AI monetization," said Angelo Zino, senior equity analyst at CFRA Research. The San Jose, California-based creative software veteran is relied on by creatives for its tools including Photoshop and Premiere Pro. The company said in April that it would add AI models from OpenAI and Google to Firefly, its generative AI tool. The tool allows users to create and edit images and videos for commercial purposes through basic text prompts without facing copyright challenges. "While guidance was raised and management remains positive around demand generation, it feels like it will take more time to prove out these (AI) initiatives and quiet concerns of competition around GenAI," RBC analysts said in a note. Adobe now expects full-year 2025 revenue between $23.50 billion and $23.60 billion, up from its prior estimates of $23.30 billion to $23.55 billion. At least five brokerages cut their price target on the Adobe stock following the second-quarter results. Including session's losses, the stock has fallen around 13% so far this year. The company's 12-month forward price-to-earnings ratio stands at 18.88, compared with Autodesk's 29.16.
