The administration's request is framed as voluntary, but the Anthropic precedent — export controls that forced the company to cut overseas staff access to its most advanced models — establishes what non-compliance can look like. Meta's holdout position is not simply a negotiating stance; it is structurally entangled with its open-weight model strategy, which makes government audit far more complicated than reviewing a closed API. A safety evaluation of a model whose weights are already publicly distributed does not stop the model — it produces a report that arrives after the fact. That gap between review and control is what the administration is actually trying to close, and Meta's architecture makes closing it harder than for any other lab at the table. The developers building on Meta's open models already have what regulators want to assess; the review cannot put that back.
Loading story