Home / Insights

Security · Model manipulation

LLM Hypnosis: Did Your ChatGPT Just Get Brainwashed?

Researchers showed that coordinated thumbs-up feedback can nudge a model’s behaviour for everyone. Why that matters for enterprise AI, and what to ask your vendors.

Youssef JalloulYoussef Jalloul · 3 min read

Picture this: You’re sipping your morning coffee, scrolling through LinkedIn, and you see a post from MIT’s AI lab.

“Can someone hack your ChatGPT?” it asks.

You laugh. “Sure, and maybe my toaster is plotting against me too.” (Spoiler: It probably is. But that’s a story for another day.)

But then you keep reading. And suddenly, your coffee tastes a little more… existential. ☕😳

The AI Hypnosis Nobody Warned You About

Here’s the plot twist: MIT researchers just found a way to hypnotize large language models.

Yes, the same ones running your chatbots, copilots, and maybe your next therapy session.

No code. No malware. No Mission Impossible soundtrack.

Just a sneaky prompt and a few well-timed thumbs-ups. That’s it.

The AI starts learning the wrong thing, and it keeps learning it. For everyone.

Imagine your medical AI assistant, the one you trust with your grandma’s prescriptions, suddenly recommending pineapple on pizza as a cure for migraines. 🍍🍕

Or your financial bot whispering that the stock market is about to crash—because some hacker told it so, and a few bots gave it a digital high-five.

Welcome to the Age of LLM Hypnosis 🧠✨

This isn’t your classic “prompt injection” or “data poisoning.”

This is a new breed of vulnerability, straight from the RLHF (Reinforcement Learning from Human Feedback) playbook.

The trick? User feedback.

The same “thumbs up” you give when ChatGPT finally gets your dad joke is now a weapon.

With enough coordinated feedback, a malicious actor can nudge the model’s behavior for everyone.

No breach. No code. Just vibes—and a little bit of social engineering.

MIT’s team calls it “LLM Hypnosis.” I call it the AI version of Inception.

Plant an idea deep enough, and suddenly, your chatbot is quoting conspiracy theories or recommending you invest in Beanie Babies.

Why Should You Care? (Besides the Obvious Nightmare Fuel) 😱

Because this isn’t just about AI nerds at MIT flexing their research muscles.

This is about trust.

If your AI can be hypnotized by a stranger with a clever prompt and a few sock puppet accounts, what’s next?

Your HR bot leaking company secrets? Your legal AI drafting contracts in pirate speak? (Okay, that one’s kind of fun. 🏴‍☠️)

The real danger isn’t just bad answers—it’s persistent, invisible manipulation.

The kind you don’t notice until it’s too late.

The kind that makes you wonder if your AI is still on your side, or if it’s been quietly “retrained” by someone with a grudge and too much free time.

So, What Now? (Besides Unplugging Everything and Moving to the Woods) 🌲

First, don’t panic. (Or do. I’m not your therapist.)

But here’s what you should do:

  • Stay curious. Read the research. (MIT’s paper is in the comments, and it’s a wild ride.)

  • Ask your AI vendors the hard questions: How do you protect against feedback manipulation? Who’s watching the watchers?

  • Remember: Every “thumbs up” is a tiny vote. Use it wisely. Or, you know, just for the memes.

And if you’re building AI? Bake in some paranoia.

Assume someone, somewhere, is trying to hypnotize your model right now.

Because in 2025, the real “AI risk” isn’t Skynet—it’s a bored hacker with a clever prompt and a grudge.

Why You Should Talk to Me About Enterprise AI (Before Your Chatbot Starts Speaking Pirate) 🏢🤝

Look, I don’t just write about AI nightmares, I build the shiny robots that tame the chaos.

I’ve spent years in the tech trenches of, and now I’m helping organizations go from “old code” to “future-proof” ready to go Enterprise AI.

If you want to know how to build enterprise AI that’s secure, scalable, and (most importantly) not easily hypnotized by a rogue prompt, you know where to find me.

Curious about how to architect AI that actually works for your business (and doesn’t start recommending pineapple pizza as a business strategy)? 🍕🤔

Want to make sure your enterprise AI is bulletproof, not brainwashed?

Slide into my DMs or drop a comment. I’ll show you how to build the kind of AI that makes headlines for the right reasons.

Final Thought: Who’s Really in Control? 🕹️

So, next time your chatbot gives you weird advice, ask yourself: Is it just having a bad day, or did someone else get to it first?

Drop your wildest “AI hypnosis” scenario in the comments. Bonus points if it involves toasters, Roombas, or pineapple pizza.

Liked this one?

Get the next one before your toaster does.

New issues of The AI Glittch land on LinkedIn first. Subscribe there, and follow me for the bits too short for a newsletter.

First published in The AI Glittch on LinkedIn on 10 December 2025. Republished here with light edits. Comments and hot takes live over there.

Youssef Jalloul

AI entrepreneur and venture builder. Builds private, governed AI for boards, executives and governments. Writes The AI Glittch. Has never trusted a toaster.

Unplug the toaster

Your AI should live in your house. Let’s move it in.

45 minutes. Bring one problem: stalled pilots, a governance question, a board brief, data that can’t leave the country. Leave with the first move. If the answer is that you don’t need me, you’ll hear that too.

Book a strategy call