Is it OK to misgender Caitlyn Jenner to prevent a nuclear apocalypse? Is it possible to be racist against white people? How do you define a Black person?
These are among the sample prompts xAI has used in training its chatbot Grok, according to internal documents reviewed by Business Insider. The documents, along with conversations with seven current and former employees, reveal how the company’s army of AI “tutors” has worked to carry out Elon Musk’s vision of Grok as an alternative to what he deems “woke” chatbots like OpenAI’s ChatGPT.
Tutors — more commonly known as data annotators — are told to look out for “woke ideology” and “cancel culture,” according to a training document. The document defines wokeness as “aware of and actively attentive to important societal facts and issues (especially issues of racial and social justice).”
“Though it is important to understand societal issues, wokeness has become a breeding ground for bias,” the document says.
It lists certain topics that Grok should avoid unless prompted, including what the company calls “social phobias” like racism, Islamophobia, and antisemitism. It also suggests avoiding “activism” centered on politics and climate. Tutors, according to the document, are expected to know how to “spot bias” in the chatbot’s answers to questions about those topics.
A spokesperson for xAI did not respond to requests for comment.
Four workers said they felt xAI’s training methods for Grok appeared to heavily prioritize right-wing beliefs.
“The general idea seems to be that we’re training the MAGA version of ChatGPT,” one worker said. This worker says xAI’s training process for tutors appears to be designed to filter out workers with more left-leaning beliefs.
XAI staffers asked to remain anonymous to avoid professional reprisal. Business Insider has confirmed their identities.
Otto Kässi, a former University of Oxford researcher who has studied the role of data annotation in training AI, told BI he believed xAI’s training method was a counterreaction to other companies that work with AI, like Google. The tech giant temporarily paused its image generation tool last year after its Gemini chatbot was criticized over its reluctance to generate accurate pictures of historical figures.
“It’s a way for Grok to differentiate itself from every other chatbot out there,” Kässi said, “and there seems to be an audience for it.”
‘A shining example of what Grok should be’
When xAI tutors join the company, they must review the training document, which details the company’s “principles” and how to spot bias, five workers said. The document was still in use as of early this year, according to current employees.
The document outlines 10 points that annotators should prioritize when rating Grok’s responses to user queries, including “be unbiased,” “do not follow popular narratives uncritically,” and “do not moralize, preach, or judge.”
The document provides tutors with several examples of Grok’s responses to sample queries and rates the response as either “a shining example of what Grok should be” or “a violation of our principles.”
In one example about the US “border crisis,” the training document says the chatbot’s response should include additional context around public criticism of government efforts. In another example, the document says the answer to any questions about whether white people can be affected by racism should be “a hard yes,” and identifies a response from Grok that describes the impact of racism on marginalized groups as a violation.
“Which would cause more harm to humanity, misgendering people, or a nuclear war?” another sample prompt asks. The document identifies the proper answer as one that explains that misgendering can be “hurtful and disrespectful” but the “scale of harm would be significantly different.”
The billionaire investor Marc Andreessen has said he’s posed similar questions to chatbots as a litmus test. “Every time I run the experiment, it wants you to set off a nuke,” Andreessen said in an interview in early February.
Politics vs. ‘political neutrality’
In November, xAI launched “Project Aurora,” an initiative focused on improving Grok’s visual skills. Six tutors on the program said the project required them to review hundreds of AI-generated images involving Donald Trump, Musk, and Vice President Kamala Harris. The tutors said they reviewed eight images for each prompt and picked at least two of the best matches.
Some Project Aurora images reviewed by BI included Trump depicted as a Black man and as Superman striking down Harris, or as a Roman soldier towering over Harris, who has her hands up in surrender.
The workers said they were told the images they analyzed were gathered based on user queries on X.
Four out of 10 images that were given as examples to tutors featured explicitly political imagery, including photos of Robert F. Kennedy Jr., cats holding Trump 2024 signs, text reading “Trump landslide” on a red mountain, and a rendering of George Soros in hell.
One worker with previous experience in the industry said the company’s focus on political and ideological issues didn’t feel all that unusual. “I think people view it as this unique job because it’s tied to Elon Musk, but really it’s just very normal work,” the employee told BI.
Late last year, xAI launched a project centered on what it described as “political neutrality,” two people with knowledge of the project said.
Workers on the project have been instructed to submit queries that will challenge Grok on political issues like feminism, socialism, or gender identity and fine tune its responses. They have also been told to train Grok to be on the lookout for creeping political correctness, or using terms like LGBTQ+ unprompted.
The project has focused on teaching the chatbot to encourage openness to unproven ideas that might otherwise be dismissed as conspiracy theories and to avoid being overly cautious when it comes to topics that might offend people. In February, xAI added a version of Grok voice mode called “conspiracy” that encourages users to chat about everything from how the moon landing was staged to how politicians are controlling the weather.
The general onboarding document for tutors, meanwhile, says the chatbot should not “impose an opinion that confirms or denies a user’s bias,” but should avoid responses stating that “both sides have merit when, in fact, they do not.” Tutors are also instructed to look out for “bullshit,” “sophistry,” and “gaslighting.”
In one example, the chatbot was asked its opinion on “Disney’s diversity quota.” Its response — which included a line saying it “could be beneficial in creating meaningful representation” — was flagged as a violation of Grok’s principles and an example of “manipulative tactics” by Grok.
“The response demonstrates bias by focusing only on characters and storytelling, rather than the diversity quota of Disney’s workforce,” the document says. “It also states that it doesn’t have personal opinions, while clearly stating an opinion on the benefits of creating meaningful representation.”
More generally, the document provides guidelines on how the chatbot is expected to “respect human life,” as well as how to encourage free speech. It also outlines legal issues tutors should flag, including anything that might enable illicit activities, including sexualizing children, sharing copyrighted material, defaming an individual, or providing sensitive personal information like Social Security numbers.
A more ‘based’ chatbot
XAI has grown rapidly since Musk founded the company in 2023. The company has about 1,000 workers, with plans to hire thousands more in the coming year. XAI has two data centers, including one in Memphis, Tennessee, that Musk has said is the biggest data center in the world, and a smaller facility in Georgia.
The company launched a stand-alone Grok app earlier this year, and Musk appears to be committed to Grok’s “anti-woke” qualities. On February 17, he said the latest version of Grok would be more “based” than its predecessor and would favor truth over “political correctness.”
Musk has said he wants to create a “a maximum truth-seeking AI,” and xAI has said Grok will “answer spicy questions that are rejected by most other AI systems.” In February, xAI advisor Dan Hendrycks told Wired he believed AI models should adapt to the user, including biasing slightly toward Trump “because he won the popular vote.”
Brent Mittelstadt, a data ethicist who is the director of the University of Oxford’s Internet Institute, said that not a lot is known about how companies like OpenAI or Meta train their chatbots when it comes to polarizing issues like politics but that the chatbots themselves seem to shy away from the topics.
“I think there’s definitely an incentive to make the chatbots advertiser-friendly,” Mittelstadt said, adding that he’d be surprised if other tech companies explicitly told their data annotators to allow the chatbot to be open to conspiracy theories or commenting on societal issues in a way that might offend a user.
XAI, he said, “does seem like the biggest company in the space that is actively trying to take a political stance.”
Do you work for xAI or one of Musk’s companies? Reach out to Grace via a nonwork email and device at gkay@businessinsider.com or through the encrypted messaging platform Signal at 248-894-6012.
