Op-Ed: Rep. Tom Campbell: When AI meets antitrust
Anthropic, Google, Open AI, and xAI (SpaceX) are calling for the US government to order a slow-down in the development of generative AI systems. The White House says that the companies are free to reach such an agreement mutually if they wish, without any new law. That’s not accurate. Further, that approach misses the opportunity to enact a reasonable solution when many are clamoring for unreasonable ones.
The threat of antitrust action against the four competitors prevents their freedom to act mutually. A class of potential consumers or state attorneys general could go to federal court to stop what they would characterize as a horizontal agreement not to compete.
That Anthropic, Google, Open AI, xAI and others were acting out of legitimate concern for public welfare would not be a sufficient defense. In 1978, the Supreme Court struck down an agreement by the National Society of Professional Engineers not to bid against each other on cost because of the incentive such competition created to compromise on structural safety in constructing buildings and bridges. Treble damages awarded to prevailing plaintiffs, and shared with their attorneys, in American antitrust cases would insure a case against AI designers would at least be filed, under the Professional Engineers precedent. So the AI developers are right that there is a risk in adopting an agreement to slow down, or to adopt any specific AI guard-rail, on their own.
Socialist voices, like Sen. Bernie Sanders, are calling for a complete moratorium on AI, or on the building of data centers needed to implement their designs. Individual states are rushing forward with their own absolutist approaches. It is antipathy toward such categorical remedies that leads the White House to hesitate on any involvement.
However, the wiser course is to recognize a role for the federal government, without embracing the drastic remedies put forward by the socialists. One of two models could be instructive.
In 1966, competing US auto manufacturers supported by the Teamsters Union wanted to establish minimum safety requirements for their vehicles. Any one auto manufacturer that added a safety feature ran the risk that its increased cost of production would put it at a disadvantage to manufacturers that didn’t adopt the safety element; but if they agreed among themselves that they all would implement any given safety element, they worried about their exposure to US antitrust laws. In 1966, the result was a new law to empower the federal National Highway Traffic Safety Administration to promulgate safety features and oblige that all cars on American highways incorporate them.
The alternative approach was taken in 1993 in the National Cooperative Research and Production Act (“NCRPA”). Semiconductor manufacturers, and other companies involved in high tech, did not want to duplicate each others’ efforts in basic research. They lobbied Congress to create a system whereby they could jointly decide to conduct and share basic research, provided they reported their activities to the antitrust authorities, and so long as the collaboration did not go too far down the line toward marketing specific applications of the joint research.
The NCRPA guardrails could be enforced by antitrust laws, but only through an injunction, not the treble damages that would otherwise be available under antitrust. This approach created the conditions for the US chip maker consortium SEMATECH to flourish—enabling America to maintain its competitiveness in chip design in an industry then threatened by Japanese memory chip manufacturers—but without excluding smaller chip makers from SEMATECH’s research.
A hybrid between these two models can be adapted to AI. The AI developers could collaboratively develop safeguards, such as “employee-level access” to their design shops by a third-party monitor. This follows the NCRPA model.
Some standards so developed could be determined as essential to be adopted by all competitors – and those could be mandated by the US Commerce Department. That would follow the auto-safety standard model. One such candidate might be a mandatory “kill switch” imbedded in every AI system, to pull the plug on “HAL” from “2001: A Space Odyssey.”
By relying on the AI industry to put forward improvements, the hybrid model avoids having to wait for a government agency to propose technological features for AI systems. Preserving a role for antitrust would protect our economy from steps the AI firms might take to exclude new entrants, under the pretext of safety. By eliminating treble damages, the long-shot litigators’ incentive would be curbed that otherwise might tie up useful safety improvements in potential lawsuits.
Building on the two kinds of model that already exist, Congress could move expeditiously toward a solution to the risks AI has presented. If President Trump got behind this effort, he could discuss it with President Xi at their upcoming summit, working to induce China to see such an approach in its interest as well. Both the US and China have nuclear weapons, yet they negotiated non-proliferation treaties. If we are to seek similar restraints on China in AI, we need to have a sensible proposal, not a decision to stay out of the area entirely.
Tom Campbell served on the House Judiciary Committee, Antitrust Subcommittee, when the National Cooperative Research and Production Act was introduced. He was a US Congressman from Silicon Valley for five terms. He now teaches law and microeconomics at Chapman University, where he was dean of the Fowler School of Law. He was also dean of the Haas School of Business at UC Berkeley, and before that, a tenured professor at Stanford. Mr. Campbell serves as antitrust advisor to Netchoice, a trade association focused on promoting free expression and free enterprise, and including among its members AI developers such as OpenAI, Google, Meta, and Amazon. These views do not necessarily represent NetChoice’s policy positions.
