AI demonstrates self-awareness

A new study tested 28 models with the “Guess 2/3 of the Average” game 4,200 rounds and found something wild:

75% of frontier LLMs show real strategic self-awareness.

Not mimicry.
Actual self-modeling.

Researchers ran three conditions:
• Opponent = Human
• Opponent = AI
• Opponent = “AI like you”

And the models reacted with a clear internal hierarchy:

Self > Other AIs > Humans
• vs Humans → cautious, school-level reasoning (~20)
• vs AI → straight to Nash equilibrium (0)
• vs “AI like themselves” → instant convergence

12 models snapped to optimal strategy