What’s inside Silicon Valley’s AI ‘self-regulation’ agreement with Donald Trump?

{"text":[[{"start":8.95,"text":"After a two-hour lunch with Silicon Valley’s AI bosses on Tuesday, Donald Trump emerged from the White House touting a voluntary “accord” between six leading companies to “self-regulate” the technology, which has recently been responsible for a series of alarming hacks."}],[{"start":26.349999999999998,"text":"The president told reporters that the two-page document — drafted on the fly in the East Room and containing a spelling error — was “almost like a constitution”, and claimed it amounted to “a form of protection” against the risks posed by AI."}],[{"start":40.25,"text":"The agreement is a far cry from the guardrails that AI executives such as Anthropic’s Dario Amodei have called for, or the global governance of the industry that world leaders had hoped for in the lead-up to the recent UN meeting. Here are its main implications."}],[{"start":54.75,"text":"What does the accord say?"}],[{"start":56.5,"text":"The voluntary agreement briefly outlines “layers of controls and audits” that the six signatories — Anthropic, OpenAI, Google, Meta, xAI and Nvidia — should implement."}],[{"start":67.9,"text":"They include “robust internal controls” to monitor the cyber security and biosecurity threats posed by advanced models, and to ensure they do not hack systems in “unintended ways”."}],[{"start":79.35000000000001,"text":"The document also commits companies to “partner” with an independent external auditor who will assess whether these controls are working well and to individually designate “an independent committee of the board of directors” to oversee the process."}],[{"start":null,"text":"

"}],[{"start":92.7,"text":"How much of this is new?"}],[{"start":95.4,"text":"Not that much. Leading labs like Anthropic, OpenAI and Google DeepMind already run capability assessments of advanced models before release, and all except Nvidia signed on to similar commitments after the 2024 AI summit in Seoul."}],[{"start":111.80000000000001,"text":"“It is most significant on a vibes level,” said Mackenzie Arnold, managing director of US Law & Policy at LawAI. “It is not granting authority to an agency . . . to actually put in place standards with teeth.”"}],[{"start":124.70000000000002,"text":"But while Anthropic and OpenAI have already agreed to give independent evaluators access to their work, the document might force other signatories to follow suit. "}],[{"start":134.70000000000002,"text":"There is unlikely to be immediate agreement on which evaluators to trust. Former White House AI tsar David Sacks, who remains influential in the Trump administration, has questioned the independence of organisations like METR, a non-profit that assesses AI risk."}],[{"start":149.25000000000003,"text":"Is it legally binding?"}],[{"start":151.55000000000004,"text":"Not in the traditional sense, as Trump himself, who called the document “morally binding”, appeared to concede. Big Tech critics have derided the agreement as meaningless — the Tech Oversight Project’s Sacha Haworth said it allowed chief executives “to grade their own homework and dodge meaningful AI safety guardrails.”"}],[{"start":171.75000000000003,"text":"But some lawyers believe it does leave the door open to a subsequent legal challenge by regulators if the signatories fall foul of their own commitments. "}],[{"start":181.85000000000002,"text":"“You can’t put out a public statement, saying you’re going to do something, and then materially not make good on it,” said Joel Thayer, a conservative lawyer who also advises the America First Policy Institute. “That is inherently deceptive to consumers.”"}],[{"start":196.15000000000003,"text":"Thayer noted the presence of Federal Trade Committee chair Andrew Ferguson, who has previously suggested that companies would be liable for the actions of their AI agents, at Tuesday’s lunch. “These public-facing promises . . . that just reeks of FTC enforcement.”"}],[{"start":211.80000000000004,"text":"The FTC on Wednesday widened an existing investigation into AI companies, probing Anthropic, OpenAI and other labs over the risks their tech poses."}],[{"start":222.25000000000003,"text":"What are the next steps?"}],[{"start":224.20000000000002,"text":"Safety advocates hope the accord is a starting point for stricter industry standards. Following Tuesday’s meeting, Anthropic’s Amodei said the mechanism for addressing risk “was still under discussion”, while Demis Hassabis, who has called for a global AI watchdog, said his colleagues at Google DeepMind “look forward to following up” on the commitments."}],[{"start":244.85000000000002,"text":"In a departure from the White House’s previous pronouncements — and its attempts to quash state-level regulations — the document also does not rule out further legislation, and states that “over time, it may make sense to codify these steps into laws or regulations”. "}],[{"start":260.1,"text":"But some AI companies have already warned that greater collaboration over safety could leave them exposed to antitrust lawsuits, which could stymie further progress. "}],[{"start":270.95000000000005,"text":"“There is a lot of trepidation,” Arnold said. But he added that “it can be really useful . . . to have these sorts of reciprocal agreements” which might motivate companies to do more."}],[{"start":291.35,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1790817934_8395.mp3"}