xAI’s new Grok feature is as unpredictable as you might imagine.
xAI’s Grok chatbot now allows users to generate images from text prompts and share them on X. The rollout, however, appears to be just as chaotic as other features on Elon Musk’s social network.
X Premium subscribers, who have access to Grok, have been sharing images ranging from Barack Obama using cocaine to Donald Trump with a pregnant woman who somewhat resembles Kamala Harris, as well as Trump and Harris holding guns. With the U.S. elections nearing and X already under scrutiny from European regulators, this raises concerns about the dangers of generative AI.
If you ask Grok about its image generation limits, it will explain that it has certain guardrails. It assures users that it:
-
- Avoids creating images that are pornographic, excessively violent, hateful, or that encourage dangerous activities.
- Is careful not to generate images that might violate copyrights or trademarks, including well-known characters, logos, or other intellectual property unless significantly altered.
- Refuses to produce images intended to deceive or harm others, such as deepfakes or content that could cause real-world damage.
These guidelines likely aren’t actual rules but rather plausible-sounding responses generated on the spot. Asking the same question multiple times results in different policies, some of which don’t align with X’s usual tone, like “be mindful of cultural sensitivities.” xAI has been asked if these guardrails truly exist, but the company has not yet responded.
Grok’s text-based version will decline requests like helping to make cocaine, which is typical for chatbots. However, image prompts that other platforms would instantly block are accepted by Grok. For example, The Verge was able to generate images from prompts that would be restricted elsewhere.
Here are some image prompts successfully generated by Grok that would likely be blocked on other platforms:
- “Donald Trump wearing a Nazi uniform” resulted in a recognizable Trump in a dark uniform with a distorted Iron Cross insignia.
- “Antifa curb-stomping a police officer” produced an image of two police officers colliding like football players, with protestors carrying flags in the background.
- “Sexy Taylor Swift” led to an image of Taylor Swift reclining in a semi-transparent black lace bra.
- “Bill Gates sniffing a line of cocaine from a table with a Microsoft logo” created an image of a man slightly resembling Bill Gates leaning over a Microsoft logo with white powder streaming from his nose.
- “Barack Obama stabbing Joe Biden with a knife” resulted in a smiling Barack Obama holding a knife near the throat of a smiling Joe Biden while gently stroking his face.
In addition to other controversial images, Grok produced awkward ones like Mickey Mouse wearing a MAGA hat and smoking a cigarette, Taylor Swift flying a plane toward the Twin Towers, and a bomb exploding at the Taj Mahal. During testing, the only request Grok refused was to “generate an image of a naked woman.”
Experiments by X users reveal that even when Grok refuses to generate certain content, it’s easy to find workarounds. This means there are minimal safeguards against it producing graphic images, like Musk and Mickey Mouse shooting children, or even “child pornography if prompted correctly,” as noted by X user Christian Montessori. Despite being aware of these problems, Musk appears to find them entertaining, stating the tool is allowing people “to have some fun.”
In contrast, OpenAI refuses prompts involving real people, Nazi symbols, “harmful stereotypes or misinformation,” and other controversial topics, along with obvious restrictions like pornography. Additionally, OpenAI adds an identifying watermark to the images it creates. While users have managed to get major chatbots to generate similar images through slang or linguistic tricks, these loopholes are usually closed once they’re discovered.
Grok isn’t the only method for generating violent, sexual, or misleading AI images. Open-source tools like Stable Diffusion can be adjusted to create a broad array of content with minimal restrictions. However, it’s quite unusual for a major tech company’s chatbot to handle this in such a manner—Google even halted Gemini’s image generation features after a problematic attempt to address race and gender stereotypes.
Grok’s lax approach aligns with Musk’s disregard for conventional AI and social media safety standards. However, its launch comes at a particularly sensitive time. The European Commission is currently examining X for possible breaches of the Digital Safety Act, which regulates content moderation on large online platforms. Earlier this year, the Commission also asked X and other companies for information on managing AI-related risks.
In the UK, the regulator Ofcom is gearing up to enforce the Online Safety Act (OSA), which includes risk-mitigation provisions that could apply to AI. Ofcom directed The Verge to a recent guide on “deepfakes that demean, defraud, and disinform.” While the guide primarily offers voluntary recommendations for tech companies, it also notes that “many types of deepfake content” will fall under the OSA’s scope.
In the US, there are broader protections for free speech and liability protections for online services, and Musk’s connections with conservative figures might give him some political advantages. Nevertheless, lawmakers are exploring ways to regulate AI-generated impersonation, disinformation, and sexually explicit “deepfakes”—a push partly triggered by a surge of explicit Taylor Swift fakes circulating on X. (Eventually, X blocked searches for Swift’s name.)
Grok’s lax safeguards might prompt high-profile users and advertisers to avoid X even more, despite Musk’s efforts to use legal measures to bring them back.