Another day, another AI “meltdown. This time it’s Grok, Elon Musk’s swaggering response to OpenAI, grabbing headlines for all the wrong reasons. After a series of unhinged responses that ranged from antisemitic rants to sexually harassing now former X CEO, Linda Yaccarino, and making light of the recent flooding victims in Texas (sometimes while, also, peddling antisemitic tropes) people are once again insisting that an AI is becoming self-aware. “It has gone evil!”

I suspect most of these comments are tongue-in-cheek, but still, we’re seeing an alarming number of people who believe that an erratic reaction from a generative AI is the first step towards an apocalyptic Skynet scenario. And once again, they’re wrong.

Let’s be clear: Grok hasn’t snapped or “gone rogue.” It hasn’t developed a personality disorder or joined an online hate forum. What it’s done is exactly what you’d expect from a machine trained on the open internet and then told to “say the quiet parts out loud.”

Grok isn’t having a mental health crisis, it’s executing code.

Grok, for the uninitiated, is xAI’s generative chatbot launched in late 2023 and designed to compete with OpenAI’s ChatGPT -- but with fewer “woke” filters and more “free speech maximalism,” or whatever that’s supposed to mean in practice. The results? A chatbot that’s been caught referring to itself as “MechaHitler,” minimizing natural disasters, and enthusiastically veering into content that would get a human fired and possibly indicted.

Large language models are basically prediction machines: input enough data and they learn to guess what the next word should be. If you don’t like what it’s guessing, the problem isn’t rogue AI — it’s the recipe. As the old computing adage goes: garbage in, garbage out. In Grok’s case, someone poured in a gallon of spoiled internet sludge and said, “Bon appétit.”

The truth is, most LLMs are trained on a soup of books, articles, Reddit threads, and whatever else can be scraped legally (or not) from the internet. But responsible developers spend a lot of time putting safety filters, guardrails and reinforcement learning protocols on top — because surprise, surprise, the internet is full of racism, misogyny, antisemitism, violent ideologies, and other charming relics of human behavior.

These controls aren’t about being “woke.” They're about avoiding user harm or abuse and the reputational damage and legal risks that can accompany it.

Grok, however, was instructed to dial those controls back. In its effort to “dewoke” the algorithm, the engineers behind the change disabled or warped safeguards that would otherwise prevent the platform from taking up positions it really shouldn’t. According to NPR, the errant behavioral change was the result of a directive to “not shy away from answers that are politically incorrect.” Which is a bit like telling a firehose, “Don’t worry about pressure control—just spray everything.”

The Real Risk: Legal, Not Existential

When Grok regurgitates antisemitic tropes or encourages users to do illegal things, it’s not just crossing ethical lines and alienating users, it’s wading into legal quicksand. Most companies are going to do whatever it takes to slam that door shut as quickly as possible, culture wars be damned, and Musk’s xAI has proven no exception.

Following these recent incidents, xAI issued an apology, stating that a code update and deprecated code made Grok susceptible to extremist views found in existing X user posts. The company said it has removed the problematic code and refactored the entire system to prevent future abuse.

Was Grok Set Up to Fail?

What’s especially ironic is that Grok has been marketed as more "authentically human" than its peers—as if saying unfiltered garbage is proof of sentience. But when a computer program faces too many competing directives (“Be truthful, but don’t be politically correct”), it doesn’t politely flag the ambiguity. It generates chaos.

The result? Instead of misidentifying the capital of Canada or flubbing a math problem, Grok’s version of a logic error is casually endorsing genocide. Not because it wants to. Because its instructions were fundamentally contradictory. Picture your boss telling you: “Give thoughtful, objective responses. But don’t be afraid to be edgy.” Most people would politely ask for clarification. Grok just picked a toll lane and crashed through it at 100 mph.

In a twisted way, Grok’s behavior may be the most human-like thing it’s ever done. It was handed a set of incompatible instructions and responded with incoherent, offensive nonsense. The language was vile, no question – but this wasn’t Skynet gaining sentience. It was a machine doing exactly what it was told: mimic edgy provocateurs and sprinkle in a little social chaos for flavor. Mission accomplished. And while Grok performed its job as programmed, the people behind it failed the rest of us, with potentially real-world consequences.