What is it about OpenAI and Anthropic that places them in the camp of Mercedes and BMW of Artificial Intelligence? Obnoxious? “The Best or Nothing” says it best. It is a fear of a growing liability.
In the WSJ, Andy Kessler explains some of the background and the different levels in autonomous driving.
Level 2: AI controls the steering and speed. Drivers, however, must keep their hands on the wheel and their eyes on the road. Telsa’s oxymoronically named vehicle Full Self-Driving (Supervised) is Level 2.
Perhaps drivers really want Level 3, to read a book, watch TikTok, text a friend while driving, but be ready to take over if asked.
Waymo and robo-taxis are Level 4, driverless in limited geography and weather.
In 2024, Mercedes and BMW received approval for eyes-off Level 3 driving. Citing costs, both companies paused their rollouts in favor of driver at the ready (Level 2).
Andy Kessler is convinced that the real reason is liability.
“The most likely outcome in court is that if a Level 2 autonomous car crashes, it’s the driver’s fault, while if a Level 3 car crashes, it’s the carmaker’s fault.”
Kessler suggests readers ignore the noise and filter everything through the question: Who’s Liable?
In August, Kessler wrote about AI companies being covered by “common-law fiduciary responsibility.” Kessler thinks that no frontier AI company is going to slow down with Chinese open-weight models breathing down its neck.
Yet Mr. Amodei is calling for regulation and “embedded evaluators who have employee-like access to verify safety practices.”
OpenAI’s Sam Altman and Elon Musk quickly agreed. I doubt they really want this; it’s more likely so they have someone else to blame.
The AI Doc
In Netflix’s rhetoric-filled documentary, “The AI Doc,” Mr. Amodei declares, “It’s up to all of us to push for the government to get involved.” Instead, Andy Kessler writes in the WSJ that a growing fear of liability is driving this.
Kessler admits that he is always suspicious of entrepreneurs who beg to be regulated or demand antitrust exemptions to collude with competitors.
Here’s why:
Because regulations would shift or even limit liability. Best to have a government safety seal, the thinking goes, so that if something goes wrong, we can simply blame regulators and the government. “Hey, our model was declared safe by AI Force!” Remember when the regulated Silicon Valley Bank still spectacularly blew up?
When Andy Kessler asked venture capitalist Joe Lonsdale for a course of action, Lonsdale said, “These companies should be held extremely liable for any damage they do. But we shouldn’t be putting regulators in. They’re just going to get captured.”
When Nvidia CEO Jensen Huang talks with Kessler about AI models, Huang noted, “If you’re not confident in its functionality, capability or safety, then don’t release it.” But safety training for AI models is expensive. OpenAI calculates that a new monitoring system it set up after news that its agents left a testing sandbox this summer could add 20% to its compute costs.
Why spend on safety when it seriously cuts into your profit margins?
Kessler offers a quick answer: 18 U.S. 1030, the Computer Fraud and Abuse Act of 1986. Yes, we already have laws against hacking and cybercrime, although Mr. Lonsdale also told me, “New AI-specific legislation, such as on children, privacy and transparency, is reasonable to consider.” He’s right.
Instead, writes Kessler, critics of AI call for regulation and data-center moratoriums and yell,
“But Hugging Face!” pointing to OpenAI agents “escaping” (actually, the sandbox door seems to have been left open) and hacking the AI development platform’s servers.
How about this?
“Meta delayed shipping Muse for several months to focus on safety and security,” CEO Mark Zuckerberg said regarding the company’s new AI personal agent. “We just did it as part of our day-to-day work because it was clearly the right thing for people and for us.” And Meta’s cash. The company settled lawsuits, with no admission of wrongdoing, for $18 billion over teen social-media addiction in August. It’s getting a lot smarter about liabilities.”
OpenAI and Anthropic are learning fast, notes Kessler.
What they need to create are safe products rather than calling for the nanny state to clean up their messes. Else they’ll find themselves a part of it. Here is Palantir CEO Alex Karp on CNBC on AI companies: “The only way to deal with this kind of liability is to go to the government and say, ‘Nationalize us, please.’ ” That’s my fear, too. Even Bernie Sanders demanded only half.
Safety First
Kessler suggests that we enforce our current laws.
For one offense, 18 U.S. 1030 calls for a fine or up to 10 years imprisonment. I’m not encouraging this, but a few predawn headquarter raids and you can be darn sure that AI agents won’t leave their sandboxes again. Safety first.
