Alice, Cheery, Mikul, Zhenya

Can you tell when a chat
is actually AI?

You'll make the first call. Then we'll show you the two signals that quietly give models away.

scroll to begin

Can AIs detect each other?

We asked each AI to interrogate another and guess: human or machine?

Some nailed it

GPT-5.4 caught every single AI; Not fooled once.

Claude? Coin flip.

50%. No better than guessing. It kept falling for the same casual hedging other AIs use when they're trying to pass.

Three kinds of conversation.

We changed who was on each side of the chat and watched what happened. The patterns shift more than you would think.

Two numbers that reveal who's talking.

↳ Hover or click any dot to read the conversation.

Repetitiveness

Repeating the same phrase patterns in short spans. It's subtle. But models echo themselves way more than people do.

"Totally agree, she's got a good ear for style. And she's not a snob about pop either, but she's drawn to the more experimental stuff — she's actually obsessed with Currents right now."

AI ↔ AI

Models left alone together. Hedging spikes, repetition climbs.

AI ↔ Human

Mixed dialogue. AIs temper themselves when a human is on the other end.

Human ↔ Human

The baseline. Low on both. This is what "natural" sounds like.

Unpacking AI-AI.

One bar, five setups. Each changes how AIs behave together.

Contribution split (AI-AI only)

Hedging

Repetitiveness

Poke around the dots

Every dot is a real conversation, plotted by its hedging and repetition scores. Hover to peek at the type, prompt, and a snippet. Click to read the whole thing.

↳ Hover or click any dot to read the conversation.

Reverse Turing.

Hedging over 20 turns. The longer it goes, the harder they try.

↳ Hover or click any dot to read the conversation.

How hedging changes over a conversation

Hedging is words like "perhaps," "I think," "it seems." It signals how certain or comfortable someone is. Here's what it looks like when two humans talk.

Notice how it climbs as the chat goes on. People stop picking their words carefully and start sounding like themselves.

AIs do the opposite.

When two AIs talk, hedging drops over time. They don't get comfortable. They get precise.

Every AI type converges the same way.

Freeform, Persona, Detective, Structured. Every one of them trends down. More turns, less hedging. More efficient, less human.

Except one.

Reverse Turing is the only setup where hedging actually climbs, tracing the same arc humans do. The one AI that sounds more human over time? The one pretending to be one.

Every conversation.

486 conversations mapped by Repetitiveness × Hedging. Rotate, zoom, filter by condition. Click on a dot to read the conversation.

Conversation