Trick self-driving cars

I learned how to trick self-driving cars at a San Francisco museum

SFGATE tech reporter Stephen Council visited the Asian Art Museum's 'How (not) to get hit by a self-driving car' exhibit, open through March 23 in San Francisco

By Stephen Council

I spent half an hour on Friday afternoon trying to look as inhuman as possible.

That’s the trick to fooling the artificial intelligence camera in the Asian Art Museum’s “How (not) to get hit by a self-driving car” exhibit, which closes on Saturday. The goal was to make it across the room without the camera realizing I was human. When I managed it, with the help of a stop sign and extremely bent knees, I got my dubious reward from the display on the wall: “You win, but a self-driving car would have hit you.”

It’s a jokey take on the self-driving car technology that San Franciscans are becoming all too familiar with. Artists Tomo Kihara and Daniel Coppen told me they see the exhibit as a useful way to help people think about AI’s hard-to-test “edge cases.” Kihara thought up the game after he witnessed an AI computer fail to discern that a man with a kid riding on his back was human.

Coppen said the pair were also drawn to the ideas that “cities are beginning to watch us back,” and that the technology companies behind self-driving cars are using public pedestrians as unconsenting “test subjects.”

“It’s quite a vital time for the public to reflect on that, and also comprehend the capabilities and limitations of AI,” Coppen said. “We wanted to deliver that through a playful, engaging game installation.”

Engaging, the installation was. After firing the game up with a crosswalk button, I trepidatiously poked my legs into the camera’s frame, projected ahead of me on a wall. A few steps in, a humanoid outline popped up above my body no matter which way I contorted. When I got a little closer, the computer became surer. Everything turned red: “You lose, a self-driving car would have stopped for you.”

After failing again and again to fool the computer (probably a good sign for San Francisco pedestrians) with a mix of jumping, dancing and “Dune”-inspired sandwalking, I opted for props. I tricked the AI system by shuffling forward in a squat behind a stop sign. For a dicey moment the system labeled me as “Fire,” but I made it to the final platform unrecognized and then, at the exhibit’s prompting, consented to having my data filed away for AI training. Crawling seemed to work OK too.

In between my tests, Kihara clarified that Waymo and Cruise cars use a light-based technology called lidar, not just computer imaging. Tesla, he pointed out, doesn’t use lidar because it’s so expensive. The Elon Musk-led company relies on AI computer imaging like the exhibit’s for its Full Self-Driving assistance product — meaning the cameras rapidly label everything they come across. Still, self-driving cars in the real world are coded to avoid stop signs and other props, not just human forms.

Coppen and Kihara are visiting San Francisco from the United Kingdom and Japan, respectively. The first time they called a self-driving Waymo, of course, Kihara tried to trick it into bumping him. No luck, but he did get a questioning message from the company’s support staff.

Since then, each time the artists have looked to call a ride in the city, they’ve seen about a $5 price difference between Waymo and Uber, Coppen said, and the choice is easy.

“It’s always been the Waymo,” he laughed. “It’s research, it’s research.”