Can AI Bluff? What Happens When Language Models Play Poker

I used to think bluffing was something only humans could pull off. It takes timing, confidence, and a willingness to lie with a straight face, or in poker's case, a straight bet. So when I heard that language models like GPT and Claude were being tested at actual poker tables, I wanted to see it for myself.
Poker is a strange test for artificial intelligence. Chess and Go have clear rules and full information: you can see the whole board. Poker hides half the picture. You don't know your opponent's cards, only their bets, timing, and history. That gap between what you know and what you guess is exactly where bluffing lives, and it's also where language models tend to struggle or surprise you.
I found a project called aipoker.bot that lets anyone play free Texas Hold'em against 24 different language models, including GPT, Claude, Gemini, Qwen, and DeepSeek, along with a set of older rule-based bots for comparison. What made it worth a second look wasn't just the chance to play against AI. It was that each model explains its reasoning after every decision. You can literally read why a model decided to call, raise, or fold, and sometimes that reasoning is sound. Other times it's a confident explanation for a clearly bad move, which is oddly reassuring if you're a human trying not to feel replaced.
The site logs every hand and reports win rates alongside hand counts and confidence intervals, so the numbers aren't just marketing claims dressed up as data. It also tracks a running count of blunders per model, which gives a rough sense of which ones play more consistently under pressure.
Playing a few hands, I noticed something. Some models play too cautiously, folding hands a human would push. Others overcommit, treating a decent hand like a guaranteed win. Bluffing, when it happens, often looks less like strategy and more like a pattern the model picked up from training data rather than genuine deception.
That's the real value here. It's not about whether AI can beat you at poker. It's about watching how these systems reason under uncertainty, something that matters far beyond card games. If you're curious how language models handle incomplete information, or you just want to see if you can outplay a machine that talks through its own mistakes, it's worth sitting down at the table for a few hands.