AI firms team up — but need regulators

AI firms team up — but need regulators Three of America’s biggest artificial intelligence companies are now doing something unusual: cooperating with one another while competing fiercely everywhere else. OpenAI, Anthropic, and Google have begun sharing intelligence to detect and block attempts by foreign firms to copy their most advanced models through a technique known as distillation. The companies say the threat is large enough to justify joint defensive action, but they also say United States antitrust law makes that cooperation legally delicate. (bloomberg.com) The immediate concern is not ordinary software piracy. It is a more specific problem: a rival can query a powerful model over and over, learn its behavior, and use those outputs to train a cheaper system that imitates the original. In the industry, that process is called distillation. Used internally, it is normal and often useful. Used by outsiders without permission, it can let a competitor shortcut years of research and billions of dollars of infrastructure spending. (bloomberg.com; businesstoday.in) That difference matters because the leading American labs have built their businesses around proprietary models. They spend heavily on data centers, chips, training runs, and safety systems, then charge customers for access through application programming interfaces and subscriptions. If a rival can mimic the product at a fraction of the cost, the original developer loses both pricing power and the ability to recover its investment. Bloomberg reported that United States officials have estimated unauthorized distillation costs Silicon Valley labs billions of dollars in annual profit. (bloomberg.com) The companies are channeling this cooperation through the Frontier Model Forum, a nonprofit launched in July 2023 by Anthropic, Google, Microsoft, and OpenAI. The forum was originally presented as a safety and standards body for frontier systems. It has now also become a venue for sharing signals about suspected “adversarial distillation” attempts that violate model providers’ terms of service. (frontiermodelforum.org; openai.com; bloomberg.com) The backdrop to this campaign is the shock created by DeepSeek, the Chinese startup whose models rattled the American industry in 2025. DeepSeek’s rise intensified fears in Washington and Silicon Valley that Chinese firms could reproduce frontier capabilities much faster and more cheaply than expected. OpenAI later told the House Select Committee on China that DeepSeek’s next model should be understood in the context of efforts to “free-ride” on capabilities developed by OpenAI and other American frontier labs. (businesstoday.in; openai.com) Anthropic had already moved in a harder direction before this latest coordination became public. In September 2025, it tightened access rules so that companies controlled from unsupported jurisdictions, including China, could no longer use its products even through overseas structures. Anthropic said the change applied to organizations whose ownership structures put them under the control of places where its products are not permitted. (anthropic.com) What the companies want now is not permission to merge or coordinate prices. They want clearer legal safe harbors for narrow security cooperation: sharing indicators of suspicious usage, account patterns, extraction methods, and other signals that might reveal model theft. Their argument is that when rivals cannot compare notes, each company sees only fragments of an attack that may be spread across multiple platforms. (bloomberg.com) That puts regulators in an awkward position. Antitrust law is designed to stop competitors from colluding in ways that hurt markets and consumers. But the same caution can make it harder for firms to coordinate against cyber intrusions, sanctions evasion, or intellectual property theft. The artificial intelligence companies are effectively asking Washington to draw a sharper line between anti-competitive coordination and defensive coordination. (bloomberg.com) At the same time, OpenAI is fighting on another front much closer to home. On April 6, 2026, the company asked the attorneys general of California and Delaware to investigate what it called “improper and anti-competitive behavior” by Elon Musk and his associates. According to CNBC’s report on the letter, OpenAI strategy chief Jason Kwon alleged that Musk had worked to undermine the company through various attacks, including by coordinating efforts with Meta chief executive Mark Zuckerberg. (cnbc.com; gizmodo.com) That dispute is tied to a long-running power struggle over OpenAI’s future. Musk co-founded OpenAI in 2015, left in 2018 after unsuccessfully pushing to merge it with Tesla, later launched rival company xAI, and sued OpenAI in 2024 over its shift toward a for-profit structure. Jury selection in the case is scheduled to begin on April 27, 2026, in the Northern District of California. (cnbc.com) There is a third layer to the story, and it shows how much broader the policy fight has become. OpenAI is no longer talking only about theft, competition, and safety. In a policy document released this week, it also urged lawmakers to rethink the tax system for an economy where artificial intelligence may erode wage income and payroll-tax revenues. The company said policymakers should consider greater reliance on capital-based taxes and even explore taxes related to automated labor. (cfodive.com; openai.com) The paper goes further than a simple “robot tax” slogan. OpenAI also floated public wealth funds and incentives for companies to retain and retrain workers as artificial intelligence changes job structures. In other words, the same company asking for more room to coordinate against foreign model copying is also asking the state to prepare for a future in which successful artificial intelligence firms could shrink the tax base that funds programs such as Social Security and Medicaid. (cfodive.com; openai.com) Taken together, these moves