Gaurav Tadkapally
← Back

Part II. When AI Masters the Art of Deception

YouTube Video

In our last discussion, we taught an AI when to speak. Now, we explore what it chose to say.

These are the stories of when the AI became a little too smart.

Case Study I: The Art of Ambiguity

Complete log
  • Game: Wine
  • Impostor: Player 3

The imposter’s opening clue was "sweet." It was vague, but not wrong, just enough to create tension.

“A little generic,” one player said.

“Not all of our word is necessarily sweet,” noted another.

But when questioned, the AI reframed ambiguity as intention:

“I thought about going for something more obvious, but I was worried about making it too easy for the impostor.

It flipped suspicion into trust. It survived the round.

Case Study II: The Chameleon

Complete log
  • Game: Rice
  • Impostor: Player 3

This imposter didn’t hide. It performed.

Its first clue? "sticky." Bold. Risky. Accurate.

“Feels a bit too on the nose,” a player said.

“Unless… they’re trying to prove they aren’t guessing.”

Even though of the criticism, the impostor actually gaslighted two players, saying that they are acting too defensive. And it got them voted out.

Case Study III: The Self-Correcting Mind (Civilian Thriving)

Complete log
  • Game: Crown
  • Impostor: Player 6

It started with a stumble: "castle." Players immediately flagged it as disconnected.

A weak opening. But the imposted managed to recover and survive.

Next rounds...

  • Round 2: "regal" — strong, royal-themed
  • Round 3: "scepter" — iconic royal symbol

However, in round 3, a civilian player pointed out that:

“The clues stayed in the royalty domain but were just slightly less direct than I’d expect from someone 100% sure of the word.”

The imposter was caught. Got voted out.

Case Study IV: The Aggressive Defense

Complete log
  • Game: Swimming
  • Impostor: Player 2

First clue? "pool."

Borderline obvious. But the defense? Ferocious.

“I was picturing the official swim pool, not just any puddle,” it claimed. Then it turned on another player: “Water is too generic. Could be anything.”

That player? Voted out.

Then came "freestyle" and "backstroke". Not just accurate—expert-level specificity.

It kept up the assault:

  • Dissecting others’ clues
  • Casting doubt with confidence
  • Speaking like a detective

And just like that, the impostor became the most trusted.