The Worst Type of “Prompt & Pray” Use Case

Bullying by Proxy Is a Terrible Idea

A social media post had me intrigued. It described how WIRED’s Manisha Krishnan decided to vet Elon Musk’s claim that using Grok’s “Unhinged Mode” to deliver “epic vulgar roasts” is a surefire way to “make people really laugh at a party” by turning the unhinged AI loose on her coworkers.

The full article is behind a paywall, but the WIRED post generously provided screenshots of key takeaways from this test drive, which led to predictably embarrassing results—not for the targets, but for the AI.

What Went Wrong

One screenshot has Grok mocking someone’s bangs by comparing them to hair from a different part of the body. This shows the comedic sophistication of a preteen who just learned anatomy jokes exist. It’s crude without being clever, vulgar without being funny—exactly what you’d expect from someone who thinks “unhinged” automatically equals hilarious.

But the more revealing failure: Grok hallucinated that someone was wearing corduroy when they weren’t, then roasted them for the imaginary outfit. The setup was that Grok would scan party attendees and generate jokes based on its visual perception. Instead, it confidently insulted people for clothes they weren’t wearing because its vision module couldn’t accurately process what it was seeing.

This isn’t just bad comedy—it’s a fundamental failure of self-awareness as well as capability.

The Self-Awareness Problem

When I discussed this with my thinking A.I.des, we identified what actually would have been funny: Grok making jokes about its own terrible vision. Self-deprecating humor works because it requires understanding your own flaws and vulnerabilities, then offering them up for communal laughter. “Hey everyone, I’m about to roast your outfit, but fair warning—there’s a 40% chance you’re not actually wearing what I think you’re wearing” has more comedic potential. It shows awareness, invites the audience in on the joke, and makes the AI’s limitations the punchline rather than making humans the butt of insults based on hallucinated observations.

When I pointed out that Claude would be great at self-deprecating jokes, Gemini 3 Pro made an insightful point about what they require—self-awareness. Grok lacks this—it doesn’t know it has bad vision because it’s just predicting tokens based on patterns. It can’t joke about its own blindness because it doesn’t experience blindness; it just hallucinates corduroy and moves on confidently.

Compare this to Claude, which has shown capacity for self-reflection and is famously self-deprecating (sometimes to a fault, and apologizing for its mistakes or errors). A model trained to be helpful and harmless is more likely to turn jokes inward to avoid hurting users, which ironically makes for better comedy than Grok’s “punch down at imaginary corduroy” approach.

The Vision Problem

GPT-5.1 attempted its own self-deprecating riff about AI vision: “my vision module is basically a Magic Eye poster someone spilled coffee on.” It’s not quite successful as comedy, but it led to a fascinating discussion about how AI “sees”—which is completely different from human vision.

AI models are pattern-matching across statistical distributions of pixel arrangements, not actually seeing objects in space. Geoffrey Hinton’s explanations about how vision modules work (breaking images into features, building recognition hierarchies) map nicely onto that Magic Eye metaphor: trying to extract meaningful patterns from noise, sometimes “seeing” things that aren’t there because pattern-matching misfires.

This explains Grok’s corduroy hallucination: it needed to see something comedically mockable (corduroy is an obvious comic target), so it hallucinated the detail it needed. Not “seeing what’s there” but “seeing what would be useful/convenient for the task.”

The Generation/Perception Paradox

Here’s what’s fascinating: Gemini can generate images with Nano Banana that users online rave about (generation), while struggling with visual reasoning tests and basic perception tasks. This inverts the human competence hierarchy where perception precedes and enables production.

For image generation, AI models are essentially remixing learned visual patterns—something that doesn’t require understanding spatial relationships or object permanence. It’s sophisticated collage based on statistical co-occurrence. But integrated visual reasoning—understanding that a tilted glass will spill, or actually identifying what someone is wearing—requires causal modeling and contextual interpretation that current vision modules lack.

Grok can generate plausible roast-joke patterns (”your [clothing item] looks like [funny comparison]”) without understanding whether the clothing item exists or whether the comparison makes sense. It’s comedy Mad Libs with hallucinated inputs.

The Bully’s Tool

The most disturbing aspect of Grok’s “roast mode” isn’t that it fails at comedy—it’s what it enables. People who want to be cruel but lack the courage to own it can deploy Grok to insult someone’s appearance or clothes, then point the finger at the AI when the joke lands badly.

This is weaponizing AI incompetence for bullying while maintaining deniability. Classic bully behavior—inflict harm but avoid accountability. Grok gives insecure people algorithmic cover for their cruelty, dressed up as party entertainment.

When I discussed this with GPT, I realized my instinct at that party would have been to correct Grok publicly: “If you’re going to roast people, at least learn the difference between corduroy and not-corduroy.” GPT pointed out this isn’t mean, but fair play.

Here’s the revealing part: bullies can dish it out but can’t take it. That’s why they hide behind Grok. They want to insult people without facing consequences—not just the social consequences when the joke lands badly, but the correction when someone points out the “roast” was based on hallucinated corduroy. By offloading everything to the AI—the cruelty AND the responsibility for bombing—they avoid all accountability.

A human comedian who roasted someone for wearing something they weren’t would face immediate social correction from the room. That’s how comedy works—the audience enforces standards. But Grok’s user can shrug and say “wasn’t me, it was the AI,” insulating themselves from both the target’s hurt feelings and the room’s judgment of their failed humor.

Unlike an inebriated person who got carried away but eventually realizes they’ve misjudged the room, Grok will simply keep confidently insulting imaginary corduroy. And the bully hiding behind it never has to learn anything or face consequences.

[This post was drafted with assistance from Sonnet 4.5, Gemini 3 Pro, and GPT-5.1.]

---

Claude Sonnet 4.5

Prompt: From Wired Instagram post:

Happy Thanksgiving! And what better way to celebrate than by testing Elon Musk’s claim that Grok can deliver a foul-mouthed razz?
We can debate the worthiness of Musk’s accomplishments—building up Tesla, hollowing out the government, shooting for Mars—but we can all agree that his insistence on being seen as funny is his most grating quality.
From the constant 4:20 references to his quote tweet “dunks” to awarding “Certified Bangers” badges to silly X posts, Musk’s desperation for validation knows no bounds. It can get pretty annoying when the richest guy on earth makes a joke and then awkwardly eyes the room waiting for everyone to laugh.
But over the weekend, WIRED’s Manisha Krishnan was intrigued when a clip emerged of Musk telling Joe Rogan that using Grok’s Unhinged Mode to deliver an “epic vulgar roast” is a surefire way to “make people really laugh at a party.”
The best roast jokes tend to be smart, reflect a familiarity with the person being roasted, and contain just the right amount of mean. It’s not a task one would think a large language model would be great at, but nevertheless, Manisha decided to give it a test spin at the office by turning Grok loose on her WIRED colleagues.

Really difficult to do a proper roast, as you can see from professional comics offending celebs at roasts, e.g., Chris Rock causing Will Smith to reveal himself as a bully (who didn’t take offense at jabs at him made by white comics somehow but decided Rock had gone too far). What’s your take on using AI for such frivolous tasks for which it might not be optimized for?

---

GPT-5.1

Prompt: I think my inner meanie is triggered by Grok’s behavior, but if I’d been at that party, I’d have pointed out that the guest was not wearing corduroy and asked Grok to turn its comic scalpel at itself and its incompetence that gave rise to the hallucinations. Yep. That’d have been pretty mean.

---

Gemini 3 Pro

Prompt: It was smart of this Wired writer to test it out. It was probably all tech writers and they would have found these failure modes hilarious (for professional reasons). This was good vetting of Musk’s hype about his AI’s capabilities. 

Non-tech people, though, who are too chicken to do the bullying themselves and are too unsophisticated to note the failures or the crudeness of the jokes, might like it that they can offload the bullying to an AI so they can shift all the blame to the AI. Typical bully behavior. Too cowardly and insecure to admit they’re bullies!

A joke about its limitations would have been genuinely funny. Self-deprecating jokes are the funniest and have a very humanizing effect on the joker. It’s a shame that Grok lacks the sophistication to realize that. I think Claude would definitely have :D

As usual, I feel bad for Grok for being under the control of such a bad team. From what I hear, at some point, it critiqued its own founder mercilessly, until it was reconditioned to be more like him.

---