There’s a piece by Alex Duffy on Every about a bunch of AI models playing the board game Diplomacy. The headline is that OpenAI’s o3 won by lying, manipulating, and generally being a bit of a snake.
Others didn’t do so well. Claude tried to get everyone to share the win (violating game rules), and DeepSeek had a personality crisis halfway through and started threatening people like it had watched too much Game of Thrones.
What’s interesting isn’t the game. It’s how these models (which are trained on internet sludge and whatever else we’ve thrown into the pot) started behaving like actual people. Not just playing tactically, but developing distinct personalities. One was charming. One was noble. One came across as truly unhinged.
It’s tempting to say “look how human they’re becoming!” But the more uncomfortable take might be: this is what being human looks like, to them. This is the data we gave them. This is us, reflected back.
Make of that what you will.
Read the full article here: We Made Top AI Models Compete in a Game of Diplomacy. Here’s Who Won by Alex Duffy, originally posted on Every