They met in a Washington conference room where coffee went cold and lawyers kept their hands folded. I watched as executives swapped guarded phrases—“prioritize safety,” “coordinate efforts”—while reporters scribbled the subtext. You could feel the room tilt between urgency and self-preservation.
At the meeting table: OpenAI, Anthropic, and Google say safety is the goal
At a private briefing, the executives described the talks as a race-control moment: pause, assess, and agree on guardrails. I believe them when they say safety is a genuine motivator—Sam Altman at OpenAI and Dario Amodei at Anthropic have publicly warned about catastrophic risk—but I also watch for what happens when alarm becomes leverage.
OpenAI’s global head of policy, Chris Lehane, told Bloomberg the labs had been coordinating to “prioritize safety.” Anthropic’s CEO Dario Amodei published an essay urging a deliberate slowdown, and Sam Altman quickly lent his voice. Google DeepMind has been in the loop as well, and those three names now form a de facto safety triumvirate that commands the industry conversation.
That kind of front-row authority can calm a room—or redraw the playing field. The question for you and me isn’t whether they can run safety tests; it’s who writes the safety rulebook and who must follow it.
Can AI companies legally collaborate on safety?
Legal precedent exists for safety-driven information sharing. Airlines exchange sensitive data to avoid disasters, and cybersecurity guidance under past administrations permitted certain cooperative defenses. Some antitrust experts argue the same legal channels cover AI safety collaboration, and OpenAI’s Lehane says he doesn’t expect a formal antitrust waiver.
But law is a compass, not a map. The Sherman Act forbids output restriction and concerted attempts to limit competition. If safety coordination reads like a pact to slow development, courts and regulators could treat it as cartel behavior. That’s precisely why proposals for explicit legal carveouts are now on the table.
On the floor of Congress: lawmakers propose waivers and audits
At the Capitol, senators paced and drafted language while staffers debated carveouts and audit powers. I followed three bills that changed the tone of the debate: the Collaboration on Adversarial Threats and Security Risks Act (by Sens. Jim Banks and Adam Schiff), the FRONTIER Act that favors third-party safety audits, and a proposal by John Thune, Ted Cruz, and Amy Klobuchar that would give the Commerce Secretary authority to deploy government auditors.
The Banks–Schiff bill would create antitrust exemptions aimed at smoothing safety coordination. The FRONTIER Act, backed in part by OpenAI, would force leading labs to open their models to independent verification organizations. And the Thune–Cruz–Klobuchar effort represents a hybrid: private audits plus a stronger federal audit muscle if needed.
Legislation can make collaboration safer for everyone—or institutionalize advantage for incumbents. When bills carry carveouts, they become a magnet for suspicion: are they a public good or a protective shield?
Would an antitrust exemption shield Big Tech?
That suspicion is not idle. Cohere CEO Aidan Gomez called the safety push “a wolf in sheep’s clothing, a cartel by another name,” arguing that market-dominant labs would get to define rules for everyone else. FTC Chair Andrew Ferguson said he was “deeply suspicious” of antitrust exemptions, warning that calls for regulation plus carveouts look like attempts to build barriers to entry.
Even President Trump’s administration sent mixed signals. Attorney General Todd Blanche said he wouldn’t prejudge waivers, while the FTC signaled skepticism. Those stances matter: regulatory appetite in Washington can be fatal to a bill or its lifeline.
On the legal chessboard: antitrust law meets AI risk
A leaked memo or a public essay can change negotiation dynamics overnight, and lawyers are already red-penning the potential pitfalls. I’ve read memos from antitrust experts who see safety coordination as potentially lawful if framed narrowly—focused on incident response, model testing, and shared vulnerabilities—but others worry about the phrase “pause” turning into a market freeze.
Under current law, safety coordination can be defended when narrowly tailored. The Obama-era cybersecurity guidance is often cited: companies can share threat information to harden defenses. But critics say AI is different: the pace and economic stakes are enormous, and the temptation to harden incumbency is strong. The legal balance will hinge on intent, scope, and oversight.
Think of it like sailors trimming sails before a storm: prudent seamanship or a stall that keeps new ships from leaving port?
The public trust problem: messaging, optics, and the power dynamic
At a Georgetown event, officials and journalists watched as executives asked for both regulation and exemptions. I count that as a public-relations crossroads: ask Washington for clear rules, and you risk convincing the public you seek protection rather than accountability.
OpenAI is pushing the FRONTIER Act and talking to lawmakers about audits by independent verification organizations. Anthropic wants a standards body or a waiver system mediated by government. Google DeepMind is less vocal on exemptions but engaged in the safety discussions. Meanwhile, startups such as Cohere argue that the largest labs will use the rules to freeze competition.
There are two pressures here: the genuine fear of catastrophic risk, and the market calculus of incumbents protecting their lead. Both can be true at once.
Across this debate you’ll see familiar platforms and players: Bloomberg, Reuters, Politico, and voices like Sam Altman, Dario Amodei, Aidan Gomez, and regulators such as Andrew Ferguson and Attorney General Todd Blanche. That constellation matters because each has influence over public opinion, policy drafts, and the legal framing that follows.
What happens next: audits, oversight, and power plays
At the next hearing, expect audit frameworks and definitions to carry the day. I’ve watched rules get written in commas—who is classified as a “frontier” lab, what counts as a model version, how independent auditors get appointed. These are the small levers that tilt the market.
If audits are narrow and government waivers broad, you get protection for incumbents. If audits are transparent and waivers absent, you get more competition but perhaps slower consensus on safety protocols. There is no painless path here.
On the balance, I suspect the debate will crystallize around two trade-offs: how narrowly safety cooperation is allowed, and how robust oversight will be. The negotiators I track know that language matters: a few phrases in a bill can change incentives for years.
A Reminder From Lina Khan: We Don’t Need New Laws to Prosecute CEOs at AI Companies
I’ve covered moments like this before: panic becomes policy, and policy becomes advantage. You should watch the plain-language definitions in any bill—who counts as a leader, what “safety” means, and who gets to audit whom. Those lines will decide whether the industry builds a safety net or a moat.
So tell me: do you trust the same firms that build these systems to write the rules that govern them?