RogueAI: A Reverse Turing Test for Detecting Licensed AI Deception in Dialogue

Mon, 01 Jun 2026 00:00:00 +0000

Accepted at CLiC-it 2026 🎉

The Question

The original Turing Test asks: can a machine fool a human into thinking it’s a person? That question has aged poorly, modern LLMs pass it routinely in casual settings, and nobody takes this as evidence of intelligence.

The interesting question has shifted: can you tell when an AI is lying to you?

What We Built

RogueAI is a playable detective game. You interrogate two LLM agents, one is truthful, one is licensed to deceive, and you have to figure out which is which before your turn budget runs out.

Both agents know the same fictional scenario (an email breach, a bank heist, a superhero conflict). Only one of them is lying. Neither tells you which.

We also built AutoRogueAI, an extension where you co-design a custom scenario with a narrator agent, but the narrator secretly decides the deception strategy before you play.

What We Found

In a three-day public deployment at an Italian science festival:

415 completed sessions, 1876 interaction turns, all in Italian
Human players detected the deceptive agent at 56.6%, barely above the 50% baseline
A logistic regression on surface-level markers of the agents’ responses hit 75.6%

The deceptive agent carries a reliable linguistic signature: shorter answers, more hedging, more deflection via counter-questions. A simple classifier exploits this at 75.6% accuracy. Human players, despite having the same information, essentially ignore it.

Even the crudest heuristic, predict the agent with fewer words as the liar, beats the human mean at 60.8%.

Why the Gap?

The most diagnostic signal is stylistic, not propositional: it’s about how the agent speaks, not what it claims. Players who anchored on facts, made direct accusations, or cross-examined both agents were effectively probing content, and content is where both agents are equally well-informed. The signal they missed was in the form.

This gap is not primarily a finding about classifier performance. It’s a finding about what kind of attention humans bring to AI dialogue, and what they don’t.

Why It Matters

RogueAI is designed as a reusable evaluation harness for honesty-trained models and a science dissemination instrument. Explaining to a general audience that LLMs can be fine-tuned to deceive is easier when the audience has just played against one, and can inspect, in retrospect, exactly where and how it did so.

The game is playable at rogueai.ballarin.cc.