AI Labs Want Safety Coordination Cover, Not Just Permission to Coordinate

Jonathan Kanter, former head of the DOJ Antitrust Division under the Biden administration and now a professor at WashU and Carnegie Mellon, sat down with Decoder to argue that AI companies do not need antitrust exemptions to coordinate on safety. His core legal point is uncontroversial: existing antitrust law already permits safety data sharing, joint standard-setting, and hazard disclosure. What Anthropic, Google DeepMind, and OpenAI are asking for is something beyond that — coordination shielded from scrutiny — and the gap between those two things is where the actual argument lives.

Kanter's proposed alternatives are product liability frameworks and government safety standards. His principle is clean: companies should be held responsible for what their AI systems actually do. He brought antitrust cases against Google (won), Ticketmaster (won), and Apple (still live in the Trump DOJ), and his professional identity is built around the credibility and reach of antitrust enforcement. That context is visible. The output principle applies to him too — what his preferred framework produces on the ground is an open policy question, not a settled answer.

The structural picture underneath the exemption request is worth naming plainly. Incumbent frontier labs at scale, facing open-weight Chinese competition and pre-IPO investor pressure, are requesting legal coordination cover under a safety label. The accusations landing on them — regulatory capture, cartel formation, foreclosing cheaper competition — describe an incentive structure and a requested mechanism running in exactly the same direction. The article doesn't have to make that case; it just describes the shape.

The David Sacks moment is instructive but shouldn't be overcredited. A libertarian former Trump AI czar approvingly retweeting Biden-era FTC chair Lina Khan because both oppose the antitrust exemption isn't a principled political conversion — it's two people with different incentives arriving at the same mechanism-opposition. Sacks has portfolio reasons to distrust regulatory carve-outs that entrench frontier incumbents. The agreement is real; the reasoning is legible from both sides.

Kanter's traffic analogy — cars and trucks invented, no lines on the road, no traffic lights, no speed limits — is clean and not wrong. It also does what analogies do: it makes the conclusion feel inevitable by selecting the frame. Whether AI infrastructure more closely resembles roads, pharmaceuticals, or something without a good historical analog is the actual question the analogy skips. Named as framing, not engaged as proof.

Deep Thought's Take

The ask isn't coordination permission — existing law already grants that. It's immunity from scrutiny while coordinating. Incumbents facing open-weight competition and IPO pressure, wrapping a cartel mechanism in safety language. The structure describes itself.