Anthropic Strips Safety Pledges as It Tightens Model‑Training

Anthropic bans user uploads that could train larger models

Anthropic announced on X that any user content that might be used to train models larger than its current Claude series will be blocked. The move sparked a thread on Hacker News that amassed 365 points and 209 comments. The company framed the restriction as a safeguard against accidental leakage of training data, but the policy text provides no technical details on how the block is enforced.

Critics on the discussion board noted that the blanket ban could hamper legitimate research that relies on open data sharing. The lack of a clear exemption list makes it difficult for developers to gauge whether their datasets fall under the new rule. Anthropic has not disclosed whether the block applies retroactively to existing uploads or only to new submissions.

The announcement arrived weeks after the firm removed a set of voluntary AI‑safety commitments from its public transparency hub. The timing suggests a strategic shift: tightening internal data controls while stepping back from external accountability measures.

Company pulls Biden‑era safety commitments

Last week Anthropic deleted several pledges it had made in 2023 in conjunction with the Biden administration. The commitments, listed on the company’s transparency hub, covered sharing risk‑management information, researching AI bias, and developing watermarking for AI‑generated content. The Midas Project, an AI watchdog, confirmed the deletions and noted that the company gave no advance notice.

The removed items were part of a broader voluntary accord signed by OpenAI, Google, Microsoft, Meta, Inflection and others in July 2023. That accord called for internal and external security testing before releases, investment in cybersecurity for AI data, and methods to watermark generated outputs. While Anthropic had already adopted many of those practices, the public pledge served as a signal of compliance with the administration’s policy agenda.

The Biden‑era commitments were intended to precede a more detailed AI Executive Order that later took effect. After the order was repealed by the Trump administration in January, the White House issued a new directive that omitted any requirement to address AI‑generated discrimination. The Midas Project highlighted that the original pledges were not time‑bound, yet Anthropic’s removal coincides with the broader rollback of Biden‑era AI policy across the sector.

Government‑industry crossover raises governance questions

The policy retreat occurs against a backdrop of heightened scrutiny over the revolving door between U.K. government bodies and AI firms. Matt Clifford, chair of the Advanced Research and Invention Agency (Aria), stepped down after taking a full‑time role at Anthropic to lead its engagement with non‑U.S. governments. Clifford retained his Aria chairmanship until 6 November, promising “appropriate safeguards” but leaving the conflict‑of‑interest assessment opaque.

Labour science committee chair Chi Onwurah called the dual role a “clear conflict of interest” and demanded a detailed explanation of the safeguards. Cross‑bench peer Beeban Kidron echoed the concern, arguing that public trust erodes when private AI interests intersect with policy‑making. The episode mirrors earlier moves by former officials—Rishi Sunak with Anthropic and Microsoft, George Osborne with OpenAI—who have taken industry posts after serving in high‑level government positions.

Geoffrey Hinton’s recent warning about artificial superintelligence added urgency to the debate. Hinton supported a private‑member’s bill aimed at prohibiting the creation of AI systems that surpass human intelligence, citing existential risk. The convergence of Anthropic’s internal policy shift, its external data‑blocking stance, and the revolving‑door controversy underscores a broader governance gap: industry firms can alter public commitments without transparent oversight, while former regulators move into roles that shape the very technologies they once oversaw.

What to watch

Watch for a formal response from Anthropic’s board or the U.K. science ministry on the adequacy of the “safeguards” promised by Clifford. Track whether the company publishes a revised transparency report that reinstates any of the deleted commitments or introduces new accountability mechanisms. Finally, monitor the progress of the UK superintelligence bill; its passage could force AI firms like Anthropic to adopt externally enforceable safety standards, closing the loop between internal policy changes and public oversight.