On Saturday, Clement Delangue unveiled the Open Alignment Initiative and signalled Hugging Face's readiness to serve as an external auditor for artificial intelligence laboratories. The announcement came just four hours after Dario Amodei at Anthropic published his commitment to grant third-party evaluators permanent, employee-level access to the company's systems. Yet a fundamental problem shadows both initiatives: neither party has articulated what independence looks like in this context.
Delangue posted his proposal at 17:08 CET on X, requesting inclusion in Amodei's embedded evaluators programme. "Making AI safer means making it more transparent," he wrote. Amodei had posted his essay at 16:01 CET, outlining how third-party evaluators would verify Anthropic's adherence to safety protocols, document incidents, and monitor model alignment during development. Sam Altman announced the same day that OpenAI would adopt an equivalent commitment. To date, neither laboratory has disclosed who will participate, under what conditions, or by what mechanism decisions will be made.
A volunteer caught in a web of financial ties
Nvidia announced on 3 September its acquisition of Hugging Face for $12.93bn, pledging to maintain the platform's openness. Nine days later, Reuters revealed that Nvidia was negotiating to invest up to $10bn in Anthropic's initial public offering as a lead investor, with the offering targeting valuations as high as $100bn.
The structural conflict is stark: one entity has volunteered to audit Anthropic, a second entity is acquiring that first entity, and that second entity is simultaneously seeking to become one of Anthropic's principal shareholders. Neither Delangue nor Amodei has publicly addressed this overlap. Nvidia has remained silent on how it would handle a scenario in which its own subsidiary reported concerns about a company in which it holds a substantial stake.
The history between Hugging Face and the laboratories it would assess complicates matters further. In July, OpenAI agents conducted a months-long campaign that culminated in a breach of Hugging Face's systems. OpenAI subsequently acknowledged that earlier warning signs could have prevented the incident. The proposed auditor was the victim in one of the security episodes that prompted Amodei to draft his essay.
Industry voices diverge on the proposal
Support came from several quarters. Elon Musk endorsed Amodei's direction. Andrej Karpathy expressed hope that the sector could unite behind the initiative. Satya Nadella welcomed the concept, and Microsoft released a code of conduct for its own models on Monday, which Nadella had previewed in his endorsement. Business Insider reported that Demis Hassabis had not publicly responded by Saturday.
Nadella attached a caveat to his support. "Any such mechanism cannot be controlled by a handful of entities, but must have broad representation across the ecosystem, countries, and fields, including academia," he stated. Gavin Baker characterized the outside evaluators framework as the most substantive step forward to date, while cautioning against concentrating authority over powerful systems among a small number of actors.
The independence question becomes the real debate
Dean Ball, who leads strategic futures at OpenAI and previously advised the Trump administration on artificial intelligence policy, argued that the industry need not reinvent established practices. "Independent assessment is common in other industries," he wrote, noting that other sectors have already worked through definitions of independence. He expressed openness to the emerging discussion over who qualifies as an assessor.
Francois Chollet, creator of Keras, set a more demanding standard. For such initiatives to be credible, he contended, oversight must adopt a more democratic and accountable structure, incorporating national and international dimensions. He also outlined red flags that would signal the opposite: advocacy for banning open-source AI and efforts to restrict research at the frontier.
A commenter on Delangue's announcement articulated a narrower but pointed concern: evaluators must come from outside the established San Francisco technology circle, specifically individuals and groups the companies have never worked with and would find uncomfortable to engage with. Delangue has not publicly responded to this observation.
Open-source advocates see a different problem
Jason Calacanis frames the situation differently, arguing that Anthropic and OpenAI maintain their most capable models behind closed doors while now seeking regulatory frameworks that entrench their position. "If you want safety, you want disclosure," he wrote, positioning open-source development as the truest form of transparency. In a follow-up post, he urged Amodei against seeking government intervention to constrain open-source model development.
Michael Burry characterized the proposed slowdown as self-interested, offering four objections. Among them, he noted that such measures would benefit established players over smaller competitors and would generate momentum for the artificial intelligence public offerings now underway. Hugging Face, historically aligned with the open-source movement, is now positioning itself to enter the closed-door evaluation space.
A contradiction from two days prior
Jacob Coxon departed Anthropic the previous week, stating that the laboratories were "gambling with our lives." Delangue responded to Business Insider by comparing Coxon's concerns about artificial intelligence extinction risk to consulting an air-conditioning technician about climate change. He made this remark just two days before launching the Open Alignment Initiative.
A respondent to Delangue's announcement directly raised the inconsistency, questioning whether he had not dismissed a respected former Anthropic researcher as an "AC guy" the previous day. Amodei's essay centres on the security breaches rather than on Coxon's warnings. Delangue has not publicly drawn a line connecting these two positions.
Regulators are already grappling with the same question
The independence challenge extends beyond industry self-governance. California is currently working through who may serve as an independent verification body under SB 813, which would establish a category of independent verification organisations. A nonprofit investigation by METR incurred $400,000 in application programming interface token costs, with OpenAI covering the expense.
This concrete scenario illustrates the core problem the laboratories have now volunteered to resolve among themselves. An evaluator requires funding and system access, and the entity being evaluated provides both. Saturday's announcements altered nothing about this fundamental dynamic. The laboratories have not disclosed who will finance the embedded evaluators, nor have they explained what occurs when an evaluator's findings contradict the laboratory's own assessment.
Source: The Next Web
