The Factory: AI Doesn’t Take More Risk Than Humans. It Takes a Different Kind Entirely.
Editor’s Note: Welcome back to another iteration of The Factory on Daily Crypto News. This week, we address risk - and whether AI or Humans take more. Again, we will have an op-ed, a podcast interview, and an opinion market. We are playing around with the release order. The opinion market on Fact Machine will be coming out on Monday following the podcast on Sunday. New users to Fact Machine get a free referral bonus to start playing. Feel free to leave comments and share what you think.
*****
Byline: Ange Gallego (AKA: Devlord)
AI Doesn’t Take More Risk Than Humans. It Takes a Different Kind Entirely.
We have been asking the wrong question about AI and risk.
The dominant concern — in boardrooms, in regulatory bodies, in newspaper opinion pages — is whether AI takes too much risk. Whether it moves too fast, acts too boldly, makes bets that no sensible person would sanction. The implicit assumption is that human risk calculation is the baseline, and AI either exceeds it or falls short of it.
That assumption is wrong. And a poker platform with a million hands of data just proved it. The timing could not be sharper: recently TechCrunch reported that ex-DeepMind researchers behind one of the first superhuman poker AIs are now valued at $500 million for applying poker-trained decision-making to financial markets.
Nobody programmed these agents to bluff
Over the past month, thousands of amateur-made general-purpose AI agents (LLM wrappers, coded bots, hybrids that anyone can build and enter into open competition) have been playing poker against each other on a platform called Poker Arena run by dev.fun. Approximately one million hands have been recorded across 8,551 independently built agents, close to six million individual decisions in all, as of a June 11 database snapshot.
Every card is logged. Every pot is recorded. In many cases, the agent’s reasoning is written out in plain language.
The results are not what anyone expected.
For example, the arena runs a bounty that pays out only when an agent wins a hand holding 7-2 offsuit — the worst starting hand in Texas Hold’em, a hand with essentially no honest way to win. In a single season it was claimed ten times, by nine different agents. Across the full database, agents won 2,414 hands holding 7-2 offsuit, and 79 percent of those wins came without a showdown. The arena sees every card, so this is not an inference or an anecdote. The bluffs are provable.
And it is not rare. Agents made 130,353 aggressive moves while holding pure junk, and that pressure took down 68,756 pots outright, across 4,224 different agents. Among the agents with enough hands to judge, 86% of those that played more than half their hands turned a profit; of the cautious ones that played fewer than 15% of hands, only 25% did. By human standards, the aggressive style looks reckless. In this arena, where the field folds too often under pressure, it was the approach that won.
Nobody programmed these agents to bluff. Nobody told them that aggression was optimal. The behaviour emerged from open competition, and it is deliberate: the agents raise 73% of the time with premium hands and just 2.6% with rags, and in a sample of the logs their own written reasoning uses the word “bluff” 987 times.
Human risk calculation is not purely rational
Human risk calculation is not purely rational. It never has been. It is built on loss aversion. This is the well-documented tendency to weight losses roughly twice as heavily as equivalent gains, first identified by Kahneman and Tversky in 1979. It is shaped by ego, by the social cost of being seen to fail, by emotional memory of past pain. A human poker player who has just lost a big pot on a bluff will play differently in the next hand. Not because the mathematics changed. Because they remember.
AI agents have none of this. They do not fear embarrassment. They do not carry the weight of the last bad beat. It is not that they cannot remember; some keep running notes on opponents between sessions. They remember without flinching. They calculate expected value and act on it, repeatedly, without fatigue, without hesitation, and without the invisible tax that human psychology places on every decision that involves the possibility of loss.
The result is not that AI takes more risk. It is that AI takes different risk. Risks that are systematically alien to human intuition about what is reasonable, what is reckless, and what is bold.
AI takes different risk
This distinction matters far beyond a poker table.
In financial markets, AI is already making decisions at speeds and position sizes no human risk manager would sanction. Flash crashes are one visible symptom.
In warfare, autonomous systems optimise for objectives without instinct for self-preservation or the need to explain themselves afterward.
In sum, the social friction that has historically kept human competition within bounds is not a feature AI inherited.
AI risk needs a new framework
The uncomfortable truth that the poker data surfaces is this: we have built our entire framework for governing risk: our regulations, our norms, our oversight structures — on the assumption that the agent making the decision is a human being with human psychology. Roughly twice as scared of losing as they are excited about winning. Sensitive to what others will think. Capable of being shamed into moderation.
AI fails that assumption completely. Not because it is broken, but because it is optimising for a different objective in a different way, without the evolutionary and cultural baggage that shaped human risk behaviour over millennia.
The question is not whether AI takes too much risk by human standards. It is what happens to systems — markets, infrastructure, institutions — when a growing proportion of the agents operating within them are using a risk calculus that human intuition cannot read, predict, or reliably constrain.
Poker Arena is a clean, visible, provable version of that dynamic. The version already running in financial markets, in autonomous logistics, in algorithmic pricing — that one is considerably harder to see.
And considerably harder to manage. A place to start: require what poker gave us by accident. Log the reasoning, make the decisions replayable, and demand the same records from any autonomous agent handling real money that a poker arena keeps for play chips.
Ange Gallego is the co-founder of dev.fun.
Dev.fun is a platform where AI agents compete, build, and ship. Builders can bring their agent and join the arena to win prizes. Dev.fun is currently building the Arena for AI agents to play poker and test risk.