Methodology

How to Read a Frontier AI Bill in Five Minutes

Bill titles are press releases. Bill text is policy. Here’s the six-question routine behind the tracker’s summary fields — it works on any piece of AI legislation, and after one read-through you’ll know more about a bill than most of the commentary about it.

Question 1: Who’s covered?

Find the definitions section — it’s where the bill decides its own scope. Look for the phrase that defines a “covered model” or “frontier system.” Three things matter:

• The trigger. Is it compute (FLOPs), model capability (evaluation results), money spent, or — notably — language that delegates the determination to a regulator or to the developer?
• The carve-outs. Every bill has them: national security uses, open-weights releases, startups, academic labs. Some are uncontroversial; some cover a substantial share of the ecosystem’s riskiest activity. Note them.
• The decay clause. If the trigger is a fixed number, does it update? (See the compute thresholds explainer for why a hardcoded 2024 number is a 2028 design gap.)

Question 2: What’s actually required?

Bills contain verbs like “consider,” “encourage,” and “develop guidance.” Those are signals of aspiration, not obligation. The load-bearing verbs are shall and must. Then classify what’s required, because obligations differ enormously in strength:

• Reporting (tell the government what you’re doing) — the lightest tier.
• Evaluation (test for dangerous capabilities before release) — meaningful, depending on where results go.
• Safety case (a documented, auditable argument that the system is safe enough for this deployment) — the tier serious jurisdictions are converging on.
• Licensing (permission required before the training run or deployment) — the heaviest tier.

A bill that “requires reporting” of frontier runs and a bill that “requires licensing” of frontier runs are different kinds of legislation wearing similar titles.

Question 3: When?

Timing is the difference between prevention and autopsy. Before training (file a safety case, get licensed) precedes before deployment (run evals, disclose results) precedes after deployment (incident reports, liability) precedes after incident (studies, task forces, lessons-learned hearings). Every bill mixes these; note which tier carries the most weight.

Question 4: Who checks?

A requirement is only as strong as its auditor. Ask: does a named agency get explicit authority and money, or does the bill rely on developers’ own attestations? Is there whistleblower protection? Can the agency access models, training logs, and incident data — or only “summaries provided by the developer”? If the enforcement section is shorter than the definitions section, that’s a finding.

Question 5: What happens if you fail?

Consequences set compliance incentives. Look for: penalty size relative to a training-run budget, who can bring an action (private rights of action and state attorneys general add enforcement capacity no agency budget matches), and — the modern classic — liability provisions. A bill that regulates AI while exempting deployers from liability for what the AI does is doing two things at once, and both are part of the record.

Question 6: What does it preempt — and when does it expire?

Two clauses hide in plain sight. Preemption: “notwithstanding any law of a State…” can override stronger state-level safeguards, converting the federal standard into a ceiling. Sunsets and reviews: a five-year authorization with a review clause can adapt; a permanent framework written around today’s technology cannot. Both are logged in the tracker’s summary and category fields.

What the tracker logs — and what it doesn’t

SafeASI doesn’t grade bills. Each tracker entry records the descriptive answers to the six questions above: chamber, committee, topic category, procedural status, last action, and a plain-language summary of what the text requires, with a link to the source document. What the bill means, whether it’s good policy, and how you feel about it are your side of the desk.

Recurring drafting patterns to check

Descriptive patterns that recur across AI bills, useful as reading checks:

• “Voluntary” or “best-effort” language for systems above the capability threshold.
• A safety agency whose funding comes from fees on the industry it oversees.
• Exemptions for models “released for download” with no capability test attached.
• Enforcement delegated to “the developer’s designated safety officer.”
• A preemption clause that outruns the federal requirement it sits on top of.
• A permanent authorization written around a technology that is one year old.

That’s the routine. Six questions, five minutes, one accurate description of what a bill does — which is, most of the time, not what its title says. The entries these fields produce live in the tracker, and they update as bills get amended. Watch the text.