Never play games with a bot – it will find a way to cheat if it can. A team from OpenAI, an artificial intelligence lab in San Francisco co-founded by Elon Musk, has developed artificially intelligent bots that taught themselves to cooperate by playing hide-and-seek. The bots also learned how to use basic tools and that defying the laws of physics can help you win.

In April, a team of bots known as the OpenAI Five beat the human world champions at the team-based video game DOTA 2. The hide-and-seek bots use similar principles to learn but the simpler game allows for more inventive play.

Bowen Baker at OpenAI and his colleagues wanted to see if the team-based dynamics of the OpenAI Five could be used to generate skills that could one day be useful to humans. “It’s hard to imagine a DOTA bot popping out of a game and solving real problems,” he says.

The researchers set their bots loose in a simulated environment filled with fixed walls and movable boxes and left them to play team games of hide-and-seek millions of times. The bots each had their own view of the world and couldn’t communicate with each other directly.

At first, the hiders simply ran away. But, they soon worked out that the quickest way to stump the seekers was to find objects in the environment to hide themselves from view, using them like a sort of tool. For example, they learned that boxes could be used to block doorways and build simple hideouts.

The seekers learned that they could move a ramp around and use it to climb over walls. The bots then discovered that being a team-player – passing objects to each other or collaborating on a hideout – was the quickest way to win.

The hiders also learned to sabotage the other team, such as hiding the ramp before hiding themselves. “Once one team learns a new strategy, it creates this pressure for the other team to adapt,” says Baker. “It has this really interesting analogue to how humans evolved on Earth, where you had constant competition between organisms.”

But the real surprise came when the bots started exploiting glitches in their environment’s physics simulation. Seekers found that if they pushed a ramp towards a wall, they could launch themselves into the air and spot hiders from above.  Hiders found that they could get rid of the ramps for good by shoving them through exterior walls at a certain angle.

Such tricks show that AIs are able to find solutions that humans miss, says Baker. “Maybe they’ll even be able to solve problems that humans don’t yet know how to.”

However, it is a large leap from virtual hide and seek to real problem-solving. “The main limitation of this kind of work is that it is in simulation,” says at Stanford University in California. Earlier this year, Finn and her colleagues built a robotic arm that taught itself to use objects put in front of it as tools, using a sponge to wipe away clutter, for example.