Transparency Without Teeth: Initial Hints in Anthropic's Draft AI Bill

On 16 October 2026, media portal POLITICO reported that Anthropic, a "Frontier" AI lab, has been quietly pitching its own AI safety bill to American lawmakers, and a thirty-seven-page discussion draft, marked privileged, confidential, attorney work product and last edited on 29 September, has since circulated publicly with its full text.

Two caveats deserve to sit at the front of any engagement with this document rather than in its footnotes. The first caveat to entail is that Anthropic has not publicly confirmed authorship of this bill and that the attribution rests on POLITICO's reporting and on the open treatment of the draft as Anthropic's proposal by named policy experts, which makes it credibly reported, and this piece explores it in detail. The second caveat that pertains is that what has leaked is a discussion draft and not introduced legislation.

A further preliminary point concerns how one should read a frontier laboratory drafting its own statute at all, per se. Those of us who have argued since 2022 against inflated claims about imminent artificial general intelligence, whether the AI 2027 timelines that did fall flat on their own technical terms or the apocalyptic narratives that treat model capability as some destined thing, have enabled the same arising from the suspicion that such forecasts serve sophisticated marketed positioning. The correct response to this draft, is to read the text as some assumed form of institutional design, clause by clause, and to assess what it would actually do, if legally becomes the reality.

What the Draft Gets 'Right'

Let’s start with what is possibly good, because some of it is. The draft bill would oblige the largest frontier developers to publish a safety framework and a security program, to file transparency and risk reports, to report critical safety incidents within twenty-four hours where death or serious injury is imminent (the ordinary deadline being fifteen days after determination), and to accept continuing independent evaluation, and its whistleblower protections are real rather than decorative. No other laboratory has put even this much on paper, and the incident-reporting obligation in particular, which resembles reporting regimes that regulators in other sectors already know how to administer, deserves to survive whatever happens to the rest of the document.

The Obligations Bind a Few Companies; the Preemption Binds Everyone

The difficulty lies in how the draft frames the court's task. A blocking order may issue only if the Agency establishes, by a preponderance of the evidence, four things together:

1. That the frontier model presents an imminent enumerated risk;

2. That the order is necessary to avert or mitigate that risk;

3. That no less restrictive measure would adequately do so; and

4. That the order per se is narrowly tailored to the pathway identified and is no broader in scope or duration than necessary.

The bill suggests that a court must also weigh the harm an order would cause to the developer, to users and to the public, including any adverse effect on safety research and cybersecurity, and must give significant evidentiary weight to any evaluation before it under section 8(i), which, given who may pay for such evaluations, is unfortunately not a neutral detail.

Here are some exclusive grounds: (1) the burden rests on the Agency at every stage, including renewal and any motion to dissolve or modify; and (2) each question is decided de novo, without deference to the Agency or to a prior order.

When we look at these elements posed in the draft bill, a pattern emerges from these elements of evidence law as used by the company while drafting their AI bill:

1. a preponderance-of-the-evidence burden on every finding, where an ordinary preliminary injunction as per Winter v. NRDC asks only for likelihood;

2. a non-delegable determination by the agency head and an independent assessment that is necessary but never sufficient;

3. least-restrictive-means and narrow-tailoring findings; and

4. automatic expiry within 45 days with a seven-day ex parte order available in emergencies)

However, they make the one emergency power in the bill considerably harder to use than the injunctions American courts issue every day. The substantive duties fall on a handful of firms, while the ceiling falls on everyone.

The draft’s main duties therefore attach only to “large frontier developers”. This includes firms with more than $500 million in AI revenue or $1 billion in development spending in the preceding calendar year, counting affiliates.

While incident reporting reaches every frontier developer above the compute threshold, the Section 14 of this proposed bill covers every “artificial intelligence developer”, a legal term defined by the draft with no threshold at all.

The legal consequence of this bill if passed as a law could be that no American state could maintain obligations on risk disclosure, third-party evaluation, or incident reporting, and the clause operates “notwithstanding the date of its enactment or adoption”, which would put California’s SB 53 and New York’s RAISE Act at risk, to the extent they impose disclosure, evaluation or incident-reporting obligations, and forwards against anything a state legislature may yet imagine.

That proposed legal ceiling is nonetheless narrower than it first appears, because Section 14(c) to (e) of the draft bill preserves state procurement, consumer-protection, child-safety, civil-rights, labour, tort, privacy and whistleblower law, disclaims field preemption, and directs that the clause be construed narrowly, which makes it a ceiling on three functions, namely (a) disclosure, (b) third-party evaluation, and (c) incident reporting.

It is nevertheless worth separating these three specific positions that are often run together, a distinction which Charlie Bullock draws usefully in his reading of the draft.

One may support a federal statute that is sound in theory even if its implementation is doubtful, on the wager that Washington may yet govern well. On the other hand, one may oppose a federal statute that is sound in theory because it carries preemption and prefers stronger state legislation as some safety organisations did with the FRONTIER Act. However, to alternatively back a federal liability regime that is weak in its regulatory capacity by default (even in theory) while also securing real, if narrow, preemption of state safety laws is infensible for any developer that professes the same concerns Anthropic's leadership 'professes'.

If the objection were merely to the present administration, the obvious course would be to support the stronger California legislation expected shortly (as per Charlie Bullock), rather than to pre-empt it. Bullock therefore notes that his own view hardened as he studied the text, and the same, I think, is the fairer way to read it.

That being said, the operative preemption clause as leaked arrives at the course of disclosure, evaluation, and incident-reporting obligations.

Section 14(a)(4) of the draft AI bill quite severally defines a state's suspension of a model, “whether by statute, regulation, order, judicial decree, or other action”, as a covered 'function', which strongly suggests some proposed regulatory intent to bring such suspension within the preemption machinery (an ex-ante method) without stating it in operative terms.

Genuine AI safety & law critics such as Nathan Calvin of Encode AI read the clause more strongly. While that reading could be an inference from the structure of the bill rather than a quotation of text, it is an inference the drafters have not publicly denied at all. What requires no inference is that Anthropic's own framework of June 2026 stated that Congress should not preempt state law unless the federal regime meets or exceeds the strongest measures proposed in that framework (a repugnancy framework against an ex-ante method). If that test is a threshold as set by the author itself, this draft fails, and that further exposes Anthropic's approach to push for regulatory capture amidst its IPO call.

The One Emergency Power in this AI Bill Is Designed to Be Used Rarely, If Ever

Under this bill, stopping a dangerous model would be harder than it is today.

Under Section 12 of the draft AI bill, seems the only route to halting a 'dangerous' LLM runs through a “blocking order” issued by a single court in the District of Columbia, and the conditions attached to that route appear calculated to make it unusable in practice. Section 13 states that the head of the agency must personally certify an “imminent enumerated risk”, i.e., a foreseeable contribution to more than fifty deaths or a billion dollars in damage. In addition, it must demonstrate that no less restrictive measure could work. In addition, a court must review everything de novo (as per Section 12), and expressly directs the agency to weigh the harm an order would cause the developer. Calvin's assessment, that passing this bill would leave Americans less protected against catastrophic risk than they are under existing law, makes sense.

The Auditors Are Paid by the Audited, and the Statute Cannot Learn

Indian law settled the independence question long ago, and more firmly than this draft does.

Section 8 of the draft AI bill permits a developer to pay its own evaluator, subject only to the formality that payment not be tied to results, and it creates a registration and rating system rather than licensing, although Anthropic's June framework itself contemplated the licensing of evaluators. However, Indian law recognised the defect in this design decades ago. Sections 139 to 141 of the Companies Act, 2013 compel auditor rotation and disqualify the conflicted, and the SEBI (Credit Rating Agencies) Regulations, 1999 exist in their present form because oversight that is paid for by the overseen degrades, predictably.

If India does not trust the issuer-pays model for balance sheets, there is no coherent reason to accept it for the most capable AI systems yet built. Section 13 of Anthropic's draft AI bill completes the proposed regulatory approach in Section 8 of the same proposed law by confining rulemaking to the adjustment of numerical thresholds and the prescribing of forms. This means no agency may be able to effectively write substantive safety standards that evolve with the technology. The bipartisan FRONTIER Act, H.R. 9925, introduced in July 2026, went considerably further on both rulemaking and shutdown authority, at least based on the reading offered by Charlie Bullock of the Institute for Law & AI, whose analysis is noted here as his reading. A statute frozen in 2026 is an odd instrument for a technology that refuses to freeze.

Why This Matters in India

India sits at exactly the fork where this episode is most instructive. The India AI Governance Guidelines, finalised in November 2025, chose a voluntary-first, techno-legal path, and the India AI Impact Summit in New Delhi this February positioned India as the convenor of a wider global conversation, which leaves open the live question of whether that voluntary architecture should now harden into statute. The Anthropic draft is a demonstration, more useful than most white papers, of what disclosure regimes become when the regulated write them: transparency for the strong few, and silence imposed on everyone else. It is also worth naming the irony plainly.

A lab that warns of catastrophe should not draft a ceiling and call it a floor.

A company whose public identity rests on forecasting catastrophic risk has produced a statute that freezes oversight at precisely the moment its own forecasts would call for adaptable authority, and that inconsistency deserves attention from those who take such forecasts seriously and from those, this author included, who have long treated them with scepticism, because both readings converge on the same institutional conclusion.

Three lessons follow for India.

1. First, incident reporting and independent evaluation must become statutory expectations rather than voluntary courtesies, since a regime that depends on the goodwill of the governed is a regime that expires the moment goodwill does.

2. Second, the independence of evaluators must be structural, as Indian company law already understands, and not a contract between a laboratory and its chosen auditor.

3. Third, as MeitY, sectoral regulators, and Indian states build their respective layers of AI oversight, the room of each to act must be preserved, because a weak central ceiling that displaces closer oversight is a design error to recognise in Washington and to refuse in New Delhi.

Anthropic deserves measured credit for arriving with legislative text where its peers arrive with talking points. But a company that warns, more eloquently than any other, of catastrophic risk should not draft a ceiling and call it a floor. For India, the episode is a gift of clarity, since it shows exactly which parts of AI governance cannot be left to the governed.