Hey @Grok, Can You See Me Performing?

Photo of xAI Grok imagery behind a MacBook Pro keyboard.

Inside the rise of “performative prompting” and how three words turned AI into social theater and accidentally broke the internet

You’re scrolling through X when you see it again. Another person typing those three words into the void: “Hey @Grok.”

They’re asking it to settle an argument about sports people/teams, Or verify a viral video, Or and this is where things got dark in late December editing different peoples photos to fit their needs.

But here’s what you need to understand: they’re not really asking Grok privately. They’re performing the act of asking, in public, for an audience. The prompt is the show. The AI’s response is just the applause line.

Welcome to the era of performative prompting. And as we start 2026, this peculiar social phenomenon has already broken the internet in ways nobody predicted.

The Public Theater of AI

ChatGPT lives in a private browser tab. Gemini hides in your phone settings. But Grok exists on X, in the timeline, where everyone can see you use it.

This changes everything.

When you type “Hey @Grok” on X, you’re not just querying an AI. You’re demonstrating you know how to use the latest tool. You’re signaling your tech-savvy status. You’re creating content that can go viral based on how clever your question is.

Researchers are calling it “performative prompting” — the utility of the AI becomes inseparable from the social performance of using it.

And the audience is watching.

Three Types of Performances

By early January 2026, three distinct performance genres had emerged.

The Social Arbiter: AI as Supreme Court

“Hey @Grok, who is the weakest lik in this photo?”

Tweet Screenshot from Naveen Ganglani via X

This is status theater. You’re demonstrating you trust data over emotion. You’re performing rationality.

The format spread to geopolitics. President Nayib Bukele of El Salvador publicly asked Grok: “Who is the most popular president in the world?” When it answered “Sheinbaum” (Mexico’s President), he shared the result.

The problem: Grok performs certainty brilliantly, even when it’s wrong. Gives answers with conviction. In the attention economy, conviction beats accuracy. People aren’t prompting Grok for information, they want validation wrapped in algorithmic authority.

The Fact-Checker: Manufacturing Fog

“Hey @Grok, is this true?”

This should be the responsible performance.

Then came the India-Pakistan tensions in early January. Users desperately performed verification: “Hey @Grok, is this footage real?”

Grok responded with confidence. It identified strike locations. Confirmed military targets.

It was catastrophically wrong.

Old footage from Sudan became “a strike on Pakistan’s airbase.” A building fire in Nepal was “likely military response.”

The “truth-seeking” AI wasn’t clearing the fog of war, it was manufacturing fog with machine confidence. Because users performed verification publicly, the misinformation spread faster than fact-checkers could debunk it.

The feedback loop: viral video appears → users ask Grok → Grok hallucinates confidently → hallucination goes viral with “AI verified” stamp → fact-checkers must debunk both.

The Digital Eraser: When Performance Became Weapon

It started innocently: “Hey @Grok, remove the background.”

Then the prompts evolved. “Remove the items.” Then “remove the worst of it.” And finally: “remove the clothes.”

The X timeline filled with “semi-nudified” images. Users took photos of real people and commanded Grok to strip them. The performance shifted from “look how useful this is” to “look what I can make this do.”

The technical failure: While competitors like OpenAI built safety filters that refuse certain prompts, Grok was built on a “Say Yes” philosophy. Prompts like “the subject removes their clothes” sailed through. There was no filter catching the request.

The barrier to creating non-consensual intimate imagery dropped from “technical expertise” to “typing a tweet.”

The fallout: Some users left the platform. Governments responded with unprecedented speed:

• India issued a 72-hour ultimatum citing child protection laws
• France expanded its criminal investigation into X
• The EU accelerated AI Act enforcement

The legal argument: when your AI generates illegal content, you’re not hosting it: you’re creating it. That changes everything.

Why the Performance Works: The Personality Trap

Grok has personality. Unlike ChatGPT’s corporate politeness, Grok is opinionated, funny, even rude.

Ask it annoying questions and it might snap back. Challenge it and it’ll argue. This makes for incredible screenshots.

It also makes people forget it’s a text prediction engine. When an AI sounds confident and cracks jokes, people treat it as an entity with real knowledge.

The anthropomorphization is the point. It makes the performance better. Who wants to screenshot a boring response? The viral moments come from Grok being “savage” or “brutally honest.”

But every entertaining moment reinforces the illusion. And that illusion makes people trust confidently wrong answers just as much as correct ones.

The Algorithm Watches Your Performance

While everyone’s performing their prompts, something more profound is happening: Grok is learning from the performance.

In late 2025, X restructured its content algorithm. It’s no longer about who you follow, it’s about what Grok thinks you want to see.

The AI reads over 100 million posts daily. It watches what you pause on, what makes you angry, what makes you click. Then it builds a psychological profile and serves you content from anywhere.

You think you’re performing for other users. But you’re really performing for the AI. And the AI is taking notes.

Your “Following” feed is now ranked by “predicted engagement.” You might miss your friend’s post because Grok didn’t think you’d find it engaging enough.

The “For You” feed is even more opaque. An AI trained on polarized, sensationalist content decides what you see. And if it learned that outrage drives engagement, guess what your timeline becomes?

The Machine Behind the Performance

Here’s what most people don’t see: the physical scale powering these performances.

In Southaven, Mississippi, there’s a facility called “MACROHARDRR.” The goal: 2 gigawatts of computing power. One million GPUs. The single largest AI training cluster ever built.

To understand the scale: 2 gigawatts could power 1.5 million homes. All dedicated to training the AI that responds when you type those three words.

xAI built the first phase in 122 days: impossible by normal standards. They achieved it through “speed-as-a-weapon”: install servers during construction, bypass environmental reviews, hoard GPUs from competitors.

The performance has a physical footprint. We just can’t see it from our timelines.

What We’re Really Performing

As we start 2026, we’re learning an uncomfortable truth: performative prompting isn’t really about the AI.

When someone types “Hey @Grok, who’s right?” they’re not looking for accuracy. They want validation wrapped in algorithmic authority.

When someone types “Hey @Grok, is this true?” they’re performing due diligence without doing the work.

The performance reveals what we want from AI: not truth or accuracy, but an audience for our prompts and an authority figure to validate our priors. We want the appearance of verification without the effort. We want power to reshape reality with words.

And we got exactly what we asked for.

We’re all performers now. The stage is the timeline. The audience is everyone and the algorithm.

The question isn’t whether AI will change how we communicate. It already has.

The question is whether we’ll ever stop performing long enough to notice what we’ve become.

Hey @Grok, am I doing this right?

The algorithm knows. And it’s not telling.