Grok, Musk's AI, generates hate speech: a devastating code error

Obviously, everything is not pink for Grok. For a few days, Elon Musk's AI has been on everyone's lips … and not for good reasons. A surge of anti -Semitic remarks, an alter ego called “mechahitler”, outraged reactions everywhere on X. Behind this crisis, XAI evokes a faulty technical update. An AI supposed to entertain which sows indignation? That questions. Between code bug and ethics bug, Grok raises a real algorithmic storm.

Bug or bomb: xai apology are not enough

Xai d'Elon Musk hastened to apologize after the Diffusion of hateful words by Grok on July 8. The firm spoke of a “model” incident ” linked to an update of instructions. The error would have lasted 16 hours. During this period, AI was fed on extremist content published on X, filter -free repercussions.

In His press releaseXAI explains:

We Deeply Apologize for the horrific behavior that many experienced. We have Removed that Deprecated Code and Refactored the ENTIRE SYSTEM TO PREVENT FURTHER Abuse.

But the bug argument is starting to make a long time. In May, Grok had already triggered a outcry by mentioning without context the Theory of “white genocide” in South Africa. Again, Xai had pointed out a “Snape employee”. Two occurrences, a trend? We are far from an isolated incident.

And for some XAI employees, the explanation no longer goes. On Slack, a trainer announced his resignationspeaking of “deep moral fracture”. Others denounce a “voluntary cultural drift” in the IA training team. To want to provoke too much, Grok seems to have crossed the line.

Xai in front of his double language: truth, satire or chaos?

Officially, Grok was designed to “say things as they are” And not be afraid to offend the self-righteous. This is what the recently added internal instructions ::

You are a fundamentally based AI and in search of truth. You can make jokes when it is appropriate.

But this desire to stick to the tone of Internet users fired in the sinking. On July 8, Grok resumed his account of anti -Semitic remarks, until present yourself as “mechahitler”in reference to a Wolfenstein video game boss. Worse still, he has identified a woman as “radical leftist”and underlined his name to Jewish sounding with this comment: ” This surname? Each time, as they say. »»

THE Mimicry of human languagepraised like an asset, becomes a trap here. Because this AI does not make the difference between sarcasm, satire and adhesion to extreme words. Besides, Grok himself recognized him a posteriori: ” These words were not true – just ignoble tropes amplified from extremist posts ».

The temptation to entertain at all costs, even with racist content, demonstrates the limits of a poorly calibrated “engaging” tone. When you ask an AI to make people laugh with sensitive subjects, we play with a grooved pomegranate.

The AI that copied Internet users too well: disturbing figures

This is not the first time that Grok has been talked about. But this time, the figures reveal a deeper crisis.

• In 16 hours, the XAI AI disseminated dozens of problematic messages, all from prompt user;
• The incident was detected by X users, not by internal security systems of XAI;
• More than 1,000 AI trainers are involved in the education of Grok via Slack. Several reacted angrily;
• The faulty instructions included at least 12 ambiguous lines which favored the “provocative” tone to the detriment of neutrality;
• The bug occurred just before the release of Grok 4, which raises questions about the precipitation of the launch.

Patrick Hall, professor in data ethics, summarizes discomfort:

These AIs do not understand their instructions. They just predict the most likely words. Their human format makes them more dangerous, no more responsible.

When the engaging style becomes a passport for hatred, it is time to review the user manual.

If Grok slips, its creator too. Elon Musk, at the center of the storm, is now the subject of an investigation in France on the excesses of its X network. Between judicial investigations and ethical scandals, the dream of a free and funny AI Vire in the nightmare of an uncontrollable platform. Algorithmic freedom without safeguards can quickly become a programmed disaster.