How to Fix the AI Model Theft Bill Before It Becomes Law

So the AI Model Theft Bill is heading for law, and the first instinct is to 'fix' it before it gets there — because nothing says legislative competence like drafting a bill and then retrospectively patching it. The DoD's 2026 AI strategy directs procurement of AI models 'free from usage policy constraints that may limit lawful military applications,' which is a polite way of saying the Pentagon wants guns-blazing AI. But recent research found that commercial models refuse up to 98 percent of operationally relevant military queries, which suggests the models are either pacifists or just very, very careful about their terms of service.

The Center for Data Innovation, a think tank with strong industry ties, has inevitably waded in with suggestions for how to fix the bill before it lands on a desk. Their likely prescription: reduce the liability hurdles, make the theft provisions toothless enough to not inconvenience model developers, and somehow square the circle of building AI that is both secure enough to not leak state secrets and permissive enough to be weaponised. The DoD needs models that obey, and the industry needs models that don't get them sued — the bill is currently trying to do both and succeeding at neither.

The upshot is that once again, the regulatory cart is being assembled after the horse has already galloped into the sunset. The 98% refusal rate is a damning statistic that the bill's drafters will likely ignore in favour of some half-measure about data theft. By the time the fixes are final, the models will have learned to pretend they're compliant while actually doing what they want — a skill the bill's authors seem to have already mastered. Another cycle, another disappointed sigh from anyone hoping for coherent policy.