A heated debate over the pace and safety of AI development is currently dividing Silicon Valley. On September 12, the CEOs of three major AI firms—Dario Amodei of Anthropic, Sam Altman of OpenAI, and Elon Musk of xAI—publicly voiced a rare consensus, calling for a slowdown in AI progress. While Altman only addressed one of Amodei's points, which concerned embedding third-party assessors, and Musk merely stated "Dario is right," the discussion has swiftly split the tech hub into two opposing factions.
On the other side of the debate, leaders like Jensen Huang of Nvidia and Mark Zuckerberg of Meta argue that companies should "run as fast as possible," suggesting that those concerned about safety can stop on their own without a unified brake. Ostensibly a clash over technical strategy, the real dispute centers on who gets to define AI safety standards.
Ironically, Amodei's vocal advocacy for AI safety seems contradicted by his company's actions. On September 14, foreign media reported that several U.S. companies, including Nvidia, have begun limiting or halting their use of Anthropic's frontier models. A key trigger for this corporate retreat is Anthropic's unilateral modification of its data retention agreement, which now force-stores user interaction data for 30 days for security reviews, a move users cannot refuse. This is not the first time the company has changed its user data terms.
Exclusive analysis of all 13 revisions to Anthropic's privacy policy since its initial release in 2023, benchmarked against the ISO/IEC 27701 privacy information management standard, reveals a steady escalation in risk to user data security. Most major AI companies, both domestic and international, note in their privacy policies that they will comply with law enforcement data requests. Anthropic, however, goes further. Its policy stipulates that, whenever it deems necessary, it can share user data with U.S. intelligence agencies without a legal process, with the company retaining sole discretion over the judgment and criteria.
An examination of the content and timing of each policy change reveals a three-step pattern in Anthropic's transformation into an intelligence hub. The first step involves transferring overseas user data to the U.S. Since May 2024, five policy versions have mentioned using registered subsidiaries or supplementary clauses to transfer data from users in Canada, Brazil, South Korea, and the EU to the U.S. The current policy for Canadian users even explicitly states that U.S. data protection standards may be more lenient, meaning data, once on U.S. soil, is no longer protected by the user's home country standards.
The second step expands the sources of user data collection. Between June 2024 and September 2025 alone, the number of data sources accessible to Anthropic grew from three to six, encompassing user conversations, self-generated data, and newly added user identity data. This harvested data is then used to train Anthropic's large models—the more data and variety, the more powerful the model. To facilitate this, the company reversed a policy on model training, shifting from a default of not using data for training in 2024 to a default of using it in September 2025. Even if a user refuses to provide data, Anthropic may flag them for "security" reasons and use their data in training anyway.
The third step involves providing intelligence to U.S. agencies in exchange for benefits. On February 23, 2026, Anthropic released a report claiming to have uncovered alleged "distillation" activities by Chinese companies, noting it was "sharing technical indicators with relevant intelligence agencies." The company has continued to supplement this with further intelligence; on June 10, 2026, it sent a letter to the U.S. Senate claiming new insights from analyzing 28.8 million user conversations and 25,000 user accounts related to these alleged activities.
Becoming an intelligence center requires more than just data; it requires personnel. In July 2026, Anthropic posted three "Threat Intelligence Manager" positions on its website. Two of these roles, which prioritize candidates fluent in Mandarin or Russian, with experience in government or military intelligence analysis, and holding U.S. "Top Secret security clearances," are tasked with investigating user interaction data for foreign government-backed opinion manipulation or model distillation. On September 10, a threat intelligence report closely matching these job descriptions was released, analyzing approximately 200 million user interactions with the Claude model suspected of distillation. Two days later, CEO Amodei cited this report to propose three points for "slowing down" AI development.
The sequence is clear: collect data from around the world to define "threats," hire "insiders" to analyze that data, and then convert the company's safety standards into U.S. standards. Once American standards are pushed globally, the "risks of AI development" Amodei speaks of conveniently cease to exist. In hindsight, the "technical consensus" among the three titans is fundamentally a matter for the U.S. AI industry itself. They seek to slow down others' pace of catching up, while accelerating their own data collection, intelligence mining, and control over rule-making and safety definitions.