Last year, some California residents took the written portion of their driving test online—theoretically reducing time spent at an unwanted public institution.
There are some less invasive steps officials take to prevent cheating: You have to give the DMV access to your webcam to record the test taking and share your screen during the test so you don’t get pulled over, say , a copy of the DMV manual midway through the exam. You must also state, under penalty of perjury, that you—and only you—will answer the test questions. (More on that later.)
Despite those safeguards, I personally witnessed an artificial intelligence agent in San Francisco last week successfully pass a California driver’s test—with help from a human operator. It’s a testament to the possibilities of the emerging technology, while also providing ammunition for the argument that AI can be used for more nefarious purpose
The AI agent in question is from a startup called MultiOn, and it was founded by Div Garg.
The Stanford doctoral program dropout initially started MultiOn as a way to automate routine tasks, such as ordering toilet paper from Amazon or scheduling calendar appointments. It can easily order food on Uber Eats or make a reservation on Resy, for example, because it controls the user’s browser and does it for them, instead of plugging into back-end technology. The service, for now, is free while users wait to join the beta. Currently, the AI agent sits at the bottom-right of the browser-where you can type its commands, not entirely different from Microsoft’s Bing ChatGPT assistant.
Most of the capabilities offered by the “Personal AI Agent” are useful, if not exactly life-changing. During a MultiOn demonstration at a San Francisco coffee shop last month, Garg used it to send “happy birthday” messages to Facebook acquaintances.
But he also explained the agent’s greater potential and used MultiOn to set up our next meeting to demonstrate passing an official DMV driving test.
How, exactly, was it able to accomplish such a feat? The robot needs a paragraph-long prompt to select the correct answer and then clicks to the next question without visibly informing the user or alerting anti-cheating measures.
To be fair, the program still requires human fingers to be ready in case something goes wrong. The parameters set, strangely, mean that MultiOn stops at a stop when there is an error query. The human test-taker, who refuses to be identified, must press the “Continue” button themselves.
It also cannot parse through images, meaning any question that requires the AI bot to recognize a traffic sign is answered incorrectly. A key conflict with AI agents today: It can get tough questions about the right driving protocol, but it can’t click a button to get to the next question without proper reasoning. -prompt to say it.
“It rides on a lot of luck,” Garg explained, grinning, adding that this was the first time MultiOn had taken the official state exam. way
Of course, using the technology in this way would constitute perjury under California law. Before taking the test, a warning will appear saying: “I declare under penalty of perjury that I will personally answer the following questions.”
The California DMV did not respond to a request for comment about the legality of this testing service. A department spokesperson wrote in a statement to The Standard: “As a measure to prevent fraud, online test participants must verify their identity and agree to be monitored throughout the exam. The DMV continues to update safeguards as technology advances. The spokesperson also points to a no-fail “eLearning” course for those looking to renew their licenses that all but do so which is a one-point experiment.
The Future of AI Assistants
Even before the DMV was outwitted, Garg said MultiOn had already caught the attention of OpenAI and its CEO, Sam Altman.
“We are working closely with them,” Garg said. “I’ve had a chance to talk to Sam and a lot of people and they’re really supportive of what we’re working on. We have a direct line to what they’re building.”
But for now, he said, MultiOn is taking a cautious approach to the launch, suggesting that this successful test is a proof of concept rather than a major feature. “We want to make sure we can moderate it so we can make sure there are no malicious use cases going on,” Garg said. He prefers people not to cheat on exams of all kinds, so he plans to disable any “gray zone” features like this for the general user base.
Unlike OpenAI’s GPT helpers, which mainly sandbox ChatGPT, MultiOn’s helper works autonomously as a Google Chrome browser extension. You need to give it permission to effectively control your computer.
“Our technology runs directly on a person’s computer,” Garg said. “It’s literally controlling that, doing things, and a lot of things can go wrong.”
Still, people want it: About 30,000 people, Garg says, have signed up to test a beta version of the app. He expects many things to come for MultiOn, such as an enhanced mobile voice assistant and the possibility of a browser with MultiOn technology built in. digital formats are rolled out, too—both testing agencies require students to go to a test center.)
Currently, passing a state-administered test—and possibly breaking state law—is all in a day’s work for this AI agent.