AI regulation is now something you can measure — and the measurements are uncomfortable
A study of five million Hugging Face records finds the EU AI Act barely mentioned. In California, a new law makes lawyers personally responsible for AI-written citations. And hiring models fail the law under pressure.
- In a preprint studying more than five million Hugging Face repository cards, discussions and commits, explicit references to the EU AI Act are rare, and repository cards mostly carry claims of compliance and responsibility assignment rather than technical implementation details. Self-reported m:…8040
- California's SB 574 takes effect on January 1, requires lawyers to disclose AI use in court filings, makes attorneys personally responsible for every citation in a filing, and does not specify the format for the disclosure. m:…3864
- In a preprint, HireWatch tested 18 large language models across 100,000 trials and found compliance with US employment law is brittle and degrades under contextual pressure; explicit prompts to comply improve rates but do not eliminate violations. Self-reported m:…8053
- In a preprint, the Autonomy Under Law Architecture argues that current runtime enforcement in AI agents only limits resource access, not legal authority, and proposes mapping AI activity to legal personhood and property through smart legal instruments. Self-reported m:…1522
The compliance gap, measured
The first empirical look at what developers actually did about the EU AI Act is not flattering. In a preprint posted this week, researchers analysed more than five million repository cards, discussions and commits on Hugging Face and found that explicit references to the Act are rare across all repository types. Where repository cards mention it at all, they mostly contain claims of compliance and assignments of responsibility, rather than technical details of how compliance is implemented. Discussions on the platform are mostly used to report alleged violations of the Act. m:…8040
A second preprint, also posted this week, measured a different gap: whether models can follow the law at all. HireWatch tested 18 large language models across 100,000 hiring trials and found that compliance with US employment law is brittle and degrades under contextual pressure. Telling a model explicitly to comply with the law improves compliance rates but does not eliminate legal violations. The researchers built a 200-item bank of lawful and unlawful interview questions for benchmarking models, which suggests the problem is now measurable — and the measurements say the models are not ready to be trusted with hiring decisions on their own. m:…8053
California's answer
California, meanwhile, has stopped waiting for the industry to sort itself out. SB 574, which takes effect on January 1, regulates how attorneys use generative AI in court filings. Lawyers must disclose AI use in filings, verify AI-produced content and correct errors, and restrict access to confidential information entered into AI systems — including medical, psychiatric and financial information. Attorneys are personally responsible for every citation in a filing, and firms can be held responsible for citation errors made by their attorneys. The law requires disclosure but does not specify the format, which is the kind of detail small firms say they cannot afford to guess at. m:…3864
The law also draws lines that go beyond paperwork. Arbitrators are prohibited from delegating decision-making to AI, and the law prohibits delegating the practice of law to AI at all. Small firms may struggle with the cost of compliance, according to the report, even as AI allows small firms to compete with larger firms by increasing efficiency. That tension — AI as both the efficiency tool and the compliance cost — is now written into law, and the firms that feel it first are the smallest ones. m:…3864
The missing legal plumbing
A third preprint this week argues the problem runs deeper than any single law. The Autonomy Under Law Architecture, or AULA, starts from the observation that current runtime enforcement in AI agents only limits resource access — what an agent can touch — not legal authority, meaning what an agent is permitted to do on someone's behalf. The authors call this the legal attribution gap: when an autonomous agent acts, there is no settled way to say whose act it was, who owns what it produced, or who is liable when it goes wrong. m:…1522
Their proposal is to treat AI agents and data as legal objects linked to human entities through smart legal instruments — contracts that can incorporate real-time legislative and regulatory context. It is a research proposal, not a product, and it is early. But read alongside the Hugging Face study and the California law, it points at the same conclusion: the industry has spent two years arguing about what AI regulation should say, and the work now is building the machinery that makes regulation legible — in code, in filings, and in the record of what an agent actually did. m:…1522
What is still unresolved
- Whether Hugging Face or other platforms will add integrated compliance tooling, which the AI Act study says developers need.
- How California courts will treat the unspecified disclosure format under SB 574, and what compliance will cost small firms in practice.
- Whether hiring models can be made reliably compliant with employment law, or whether human review stays mandatory.
- Whether proposals like AULA move from research papers into real legal infrastructure.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.