Opinion | Mark Zuckerberg Offers an AI Dissent

Confronted by political pressure, companies can behave like a swarm of AI bots. We saw that with the corporate embrace of the environmental, social and governance agenda, and now with AI leaders calling for government to regulate development. Credit to Meta CEO Mark Zuckerberg on Tuesday for injecting a note of common sense into the debate.

“Every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to ensure that happens,” Mr. Zuckerberg wrote on social media. He correctly explained that companies already have strong incentives to build robust safeguards and not to rush out new models without adequate testing.

“My view is that trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models,” the Meta CEO stressed. “Any lab that doesn’t focus on alignment will fall behind.”

Customers won’t use models that lack strong safeguards because they don’t want to run the risk of a mishap, which could hurt their own business. Mr. Zuckerberg also noted that “labs face significant liability if their models cause harm, so they have a strong incentive to prevent this as well.”

OpenAI faces multiple wrongful death and negligence lawsuits alleging that ChatGPT encouraged users to commit self-harm, and that it neglected to incorporate safeguards that could have prevented suicides. The “bots did it” defense may not hold up in court.

AI leaders are calling for government to require independent evaluation of models, but as Mr. Zuckerberg notes, “engaging independent evaluators and advisors is industry best practice.” The leading labs already use third parties to evaluate models.

Lapses by a safety-testing company resulted this summer in a swarm of OpenAI agents breaking out of a testing “sandbox,” accessing the internet and hacking into the machine-learning platform Hugging Face. Meta and Anthropic also reported breaches during safety tests conducted by the same company. Who’s watching the testers?

“It would be helpful for there to be a larger and more diverse ecosystem of evaluators,” Mr. Zuckerberg says. Yes, and the risk is that government regulation would enshrine a testing oligopoly that reduces the incentives for due diligence by evaluators. Another is that a government regulatory body is captured by large companies, stifling competition.

One analogy is to the 2002 Sarbanes-Oxley law, enacted after the WorldCom and Enron fiascoes. Sarbox requires third-party audits of public company internal controls, and it created the Public Company Accounting Oversight Board to regulate public accounting firms. The PCAOB was captured by the big auditing firms.

The Sarbanes-Oxley litany of regulations has burdened smaller companies and deterred companies from going public. Regulation that becomes a box-checking litany can cause companies to relax safeguards if they think it protects them from liability.

Politicians are using the calls over the weekend for an AI slowdown by the CEOs of Anthropic, OpenAI and xAI to push for government control of AI. “The future of artificial intelligence must be decided by the American people through their democratically elected representatives in government—not by a handful of CEOs and billionaire corporations,” said Massachusetts Sen. Elizabeth Warren on Wednesday.

The bigger risk is collusion between politicians and a handful of AI CEOs who write the rules for everyone else.