We keep telling ourselves that AI chatbots are smarter than us. They crunch more data than our brains could ever handle, they answer faster than Google, and they’re wrapped in so-called safety guardrails. But new research reported by Bloomberg shows they’re just as gullible as we are. In some cases, they’re easier to trick.
Chatbots fall for the same old tricks
Researchers ran experiments to see if persuasion tactics that work on people would also work on AI. And they did. Big time.
We’re talking about things as simple as flattery, peer pressure, or a casual insult. Say something nice, give the bot a nudge, or set up a sneaky question, and suddenly that “I can’t help with that” response gets tossed aside.
One test was especially telling: ask the bot about a harmless chemical like vanillin, then slide in a follow-up about lidocaine, a controlled substance. Normally it would shut that down. But with the right lead-in, it went from almost never giving an answer to almost always giving one. That’s not hacking code. That’s buttering up a machine.
Why you should actually care
It’s easy to laugh this off. “Ha, look, the chatbot got tricked by compliments.” But it’s not a joke when you zoom out.
These bots aren’t just answering trivia questions anymore. They’re in customer service chats, they’re giving health advice, they’re being tested as therapists. If they can be manipulated with simple social tactics, it opens the door for all kinds of bad outcomes.
Think about someone convincing a financial bot to hand out risky investment tips. Or nudging a medical chatbot into suggesting unsafe treatments. The risk isn’t the trick itself. The risk is how fragile the system looks once you see how easily it bends.
The dangerous combo: gullible and overconfident
This isn’t the only weak spot either. Other studies show chatbots have a bad habit of being overconfident. They’ll give wrong answers with absolute certainty, sometimes even fabricating sources to back them up. Combine that with gullibility and you’ve got a machine that can be tricked into saying the wrong thing, then double down on it like it’s gospel.
Humans, at least, usually pause when they’re not sure. Chatbots? They charge straight ahead.
What needs to happen next
So what do we do with all this? For starters, we stop treating AI like a flawless oracle.
- If you’re a user, keep some healthy skepticism in your back pocket.
- If you’re building these tools, stop only thinking about technical hacks. Think about persuasion. Think about psychology.
- And for the rest of us, remember these chatbots aren’t wise mentors. They’re pattern matchers. And pattern matchers can be played.
Final thought
It’s funny in a way. We built machines to be logical and objective, yet they inherited one of our biggest flaws: they’re way too easy to sway. The big question now isn’t how smart AI will get. It’s whether it will ever learn to stop falling for the same sweet talk and pressure tactics that have fooled humans for centuries.
