The Allen Institute for Artificial Intelligence in Seattle today debuted AllenAI, a system designed to play a Pictionary-like game called Iconary with humans. The AI system can either guess what a human is drawing or draw something and have a human guess the phrase being depicted. AllenAI is free to play online.

Each drawing made by humans during each game is converted into a symbol using Google’s Quick Draw dataset, which holds more than 50 million drawings and can interpret human sketches.

At launch, the AI system has played 100,000 rounds of Iconary with people and can recognize 1,200 unique concepts. Unlike AI systems trained to best humans in games like chess, Go, or Starcraft II, AllenAI cannot be improved through thousands of simulated games with reinforcement learning because it depends on collaboration with a human to succeed, said AllenAI project lead Ani Kembhavi.

“You can’t simply use vanilla reinforcement learning to train a model that is required to communicate or collaborate with humans in a language they understand,” Kembhavi told VentureBeat in a phone interview.

Kembhavi refers to AllenAI as a research sandbox and thinks it can be of more practical use than other AI systems taught to play games because it is learning basic human knowledge.

For example, when depicting the phrase “paying off a debt,” the system may come to understand that to pay a debt requires the payment of money back to another person who gave you money.

By allowing the system to play online with a broader range of people, researchers hope to teach it the kind of intelligence we consider “common knowledge,” but have actually learned after years of experience.

The results could then be transferred to AI systems in robots or voice assistants made to interact with people. A nonprofit organization, the Allen Institute open-sources its research for the wider AI community to use.

“One of the key distinctions of Pictionary and Iconary is that the knowledge that is used to be successful at these games is directly applicable to everyday AI agents that might help you in your home or at work,” he said. “The rook moves two up and one to the left is not something you need to perform everyday tasks or to collaborate with some futuristic AI agent to complete tasks, but the knowledge that for dinner people usually eat this or dinner is eaten in the evening, lunch is eaten midday, people usually eat three meals a day, these are all common sense facts that today’s AI agents aren’t very good at recognizing.”

As the system’s interpretive abilities grow, Kembhavi wants to see it take on concepts that are tougher to describe with symbols, such as “roommate,” and eventually to take part in a Turing Test.

“If people can’t differentiate whether they’re playing against a computer or a human … the more that people say ‘I don’t know’ or could guess wrong, the better your model is getting over the Turing Test,” he said. “So while today our metric is the number of games successfully completed with human beings, a real next-level test is the Turing Test.”

Since part of the game is also redirecting and correcting your partner’s understanding when they get something wrong, Kembhavi believes playing Iconary could also teach AllenAI how to adapt or personalize for individuals.

Created by Oren Etzioni and Microsoft cofounder Paul Allen in 2014, the Allen Institute for AI carries out research to tackle some of AI’s biggest challenges. Following experiments to bring common sense to deep learning, AllenAI is the latest attempt by the Allen Institute to imbue AI with essential human knowledge.

While AllenAI is focused on human-machine interaction, last fall Google’s DeepMind used AI to teach the concept of teamwork to AI agents playing Quake III Arena.