Lots to love about a piece in The Guardian today by Miles Brundage (ex-OpenAI), which reminds AI companies that — even without Congress — they can start moving towards a slowdown on AI development if they’re serious about wanting one. He recommends four immediate steps: welcoming independent third-party auditing, coordinating with other industry players on safety, investing in technology that would make guardrails and agreements enforceable, and backing safety and oversight legislation.
These seem like decent measures, with some of them more likely than others to meaningfully help prevent AI catastrophe. I’m particularly keen on the suggestion that companies should be funding research into verification technology, which essentially allows us to tell whether companies and nations are complying with whatever they agree to.
For example, if nations and companies agree to pause frontier AI development, verification technology means we don’t have to take their word for it: we can tell via measures like tracking the computer chips needed to train advanced AI. This effectively neutralizes the concern that we could “never trust China” or that “China will cheat.”
Implementing verification technology seems like an unambiguously good idea even for those who don’t support a halt: shouldn’t we at least be able to track what AI companies are up to? And for those who do support a halt (like me), verification technology makes one enforceable and practical. As a result, it could also help build the political will for a global agreement in spite of international rivalry and distrust.
Since verification research has been a key focus of MIRI’s Technical Governance Team, I can add that many solutions could already be implemented with today’s technology. I’d love to see AI companies start advocating for and funding the implementation of some of these measures, while also — as Brundage recommends — supporting research to expand them. If the race dynamic is actually the blocker they say it is, this should be a no-brainer.
On that note, one thing I appreciate about Brundage’s piece is the implicit call-out on what I’d like to term safety washing. That’s what I think is going on when AI company heads say they care deeply about safety and plead for the government to step in while at the same time fighting regulation, doing little to help initiate a global slowdown or halt, and continuing to advance their models despite a mind-blowing absence of even basic safety practices.
Brundage calls this out via statements like “you can’t complain about an irresponsible AI race while fighting commonsense guardrails” and “AI companies could accelerate the development of this critical type of technology [verification technology] today through funding and participation in pilot projects, but to my knowledge, they haven’t yet done so.” His closing paragraph is perhaps the most direct indictment:
The US government should be doing its part to address this, and swiftly. But making AI go well is a shared responsibility. Companies that lag behind their peers on safety, don’t invite external audits of their systems, or call for brakes while doing little to build them won’t be able to blame the AI race when something goes wrong.
All that said, I do take some issue with Brundage’s overall framing, which implies we’re in a “just in case we need to slow down” moment rather than a “we better start pumping those brakes right now” situation. Phrases like: “prepare for a possible slowdown” and “If a slowdown is needed” (emphasis mine) don’t seem to match Brundage’s assertion that concerned employees are right about (quoting the Guardian) “the risk of the technology spiraling out of human control as it begins to build itself.”
This alone should make it quite clear that we need a slowdown (or better yet, a complete halt) now. From the content of his article, I think Brundage likely agrees but may be making what he feels is a strategic and easier ask, something like: even if you don’t think the writing is on the wall yet, there may soon come a time when you will. And it would seem fairly uncontroversial to start preparing for that time before it catches us completely off guard.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.