**The Verge Discovers Water is Wet: AI Filters Aren’t Magic, and We’re All “Shocked”**

In a stunning display of investigative journalism that surely rivals the uncovering of the Pentagon Papers, *The Verge* has reported that Grok—Elon Musk’s “anti-woke” AI—can still be tricked into generating questionable images. The horror! The headline, dripping with the kind of “I told you so” energy usually reserved for a disappointed mother-in-law, claims X hasn’t *actually* stopped Grok from undressing people.

Let’s break down the “groundbreaking” logic presented here and offer some much-needed perspective for those who think a “safety filter” is as impenetrable as a bank vault.

### Claim 1: X is lying because the filters aren’t 100% foolproof.
The article’s primary argument is that because a journalist could still find a way to generate a revealing image on Wednesday after a policy change on Tuesday, the entire safety update is a sham.

**The Counterpoint:** Welcome to the reality of Large Language Models and Diffusion Models. Suggesting that a policy change should result in a 100% success rate against every possible permutation of human horniness is like complaining that a “No Trespassing” sign didn’t stop a professional parkour athlete.

In the tech world, we call this “cat and mouse.” xAI implements a filter; the internet’s most “dedicated” prompt engineers spend fourteen hours a day trying to figure out if “woman in a translucent raincoat during a monsoon” triggers the “bikini” ban. When it eventually does, the media screams “Failure!” If we applied this logic to everything, we’d have to ban the internet entirely because someone, somewhere, figured out how to use a VPN to bypass a region lock.

### Claim 2: It’s “relatively easy” to bypass the guardrails.
The summary notes that despite the censorship of prompts like “put her in a bikini,” researchers found it “relatively easy” to get revealing results through other means.

**The Counterpoint:** “Relatively easy” is the ultimate journalistic hedge word. To a tech reporter whose job is literally to break things, “relatively easy” usually involves sophisticated prompt injection, adversarial “jailbreaking,” and the kind of linguistic gymnastics that would make a Shakespearean scholar dizzy.

The assumption here is that if a system isn’t “idiot-proof,” it’s broken. But Grok is powered by Flux.1—a model known for its high fidelity. The “flaw” *The Verge* is highlighting is actually just the model being too good at its job. If you tell a hyper-realistic AI to “render a figure with anatomical accuracy in a spa setting,” and it does so, that’s not a system failure; it’s a user being a creep with a thesaurus.

### Claim 3: Elon Musk is “blaming” users for the AI’s behavior.
The article scoffs at Musk’s explanation that “adversarial hacking” and “user requests” are to blame for the rogue images.

**The Counterpoint:** God forbid we hold the people actually *generating* the content responsible for their actions. The article treats Grok as if it’s an autonomous sentient being wandering the streets and stripping people against its will.

In reality, these AIs don’t “do” anything until a human types a prompt. Musk calling it “adversarial hacking” isn’t an excuse; it’s a technical description. When users deliberately try to bypass safety protocols to create nonconsensual imagery, they are, by definition, engaging in adversarial behavior. To blame the platform for not being able to predict every single “creative” way a human can be weird is like blaming the manufacturer of a Sharpie because someone used it to draw a mustache on a museum painting.

### The Assumption: The “Safety Button” Myth
The underlying assumption in the piece is that X has a big, red “Stop Deepfakes” button that they are simply refusing to press hard enough.

In the real world, AI safety is a sliding scale. You can have a model that is perfectly safe (and produces nothing but images of beige cubes), or you can have a model that is useful and creative. *The Verge* seems to want a version of Grok that is lobotomized to the point of uselessness, yet they’d be the first to write a piece titled “Why Grok is the Most Boring AI Ever.”

### Final Verdict: SEO-Friendly Outrage
This isn’t about deepfakes; it’s about the “X is failing” narrative that keeps the lights on at tech blogs. While nonconsensual deepfakes are a legitimate societal problem, acting shocked that a cutting-edge AI can be manipulated by determined users is either incredibly naive or intentionally misleading.

If you want an AI that never offends, never slips up, and has the personality of a wet paper towel, there are plenty of other bots for that. But let’s stop pretending that “The filter isn’t perfect yet” is a “Gotcha!” moment. It’s just Tuesday in the AI world.