Anthropic's Heck Says AI Safety Can't Rest on an Honor Code

Sarah Heck, Anthropic's head of public policy, said on Sept. 16, 2026 that AI companies cannot be expected to operate on an "honor code," arguing for enforceable rules as the fight over how to regulate frontier model development intensifies.

Sarah Heck, the head of public policy at Anthropic, used a blunt formulation on Tuesday to describe the limits of voluntary AI safety: companies building frontier models cannot be expected to operate on an "honor code." The remark, reported by CNBC Technology, lands at a moment when the argument over who polices model development — and with what authority — has moved from conference panels into legislative drafting rooms.

Heck's point is narrower than it sounds. She is not saying AI labs are acting in bad faith. She is saying that a safety regime resting on promises has no mechanism for the case where a promise becomes commercially inconvenient. That distinction is the whole substance of the current policy fight.

Why a voluntary framework breaks under competitive pressure

The self-regulatory model that carried the industry through the first wave of large language model releases was built on published safety policies, internal red-teaming, and staged deployment. Each of those is a good practice. None of them is enforceable by anyone outside the company that wrote it.

The structural problem is that safety commitments impose costs — delayed launches, withheld capabilities, headcount spent on evaluation rather than capability research — and those costs fall on whichever firm takes them most seriously. A lab that slows a release to run additional testing hands time to a competitor that does not. In a market where model leadership is measured in weeks, an honor code quietly penalizes the honorable. Heck's phrasing is a concise way of saying the incentives run the wrong direction and no amount of good intent fixes that.

The corollary is uncomfortable for the industry: if voluntary commitments are insufficient, the alternative is not a lighter touch. It is a legal obligation with a consequence attached. Anthropic has generally been the frontier lab most willing to say so out loud, and Heck's comments continue that line rather than opening a new one.

The regulatory fight Heck is speaking into

Two live disputes give her remarks their edge. The first is jurisdictional. State legislatures have moved faster than Congress on AI rules, producing the prospect of a patchwork in which a model deployed nationally must satisfy multiple, differently worded safety and transparency regimes. Industry groups argue that patchwork is unworkable and want federal preemption. Safety advocates counter that preemption without a substantive federal standard means no standard at all — a ceiling with no floor beneath it.

The second dispute is about disclosure. Proposals circulating in various forms would require developers of the largest models to publish safety testing methodologies, report dangerous-capability evaluation results, and notify authorities of serious incidents. Labs differ sharply on how much of that should be mandatory versus discretionary, and on whether the disclosures should be public or filed confidentially with a regulator.

Heck's "honor code" framing maps cleanly onto the second fight. If a company's own judgment is the only thing determining what gets reported, then the reporting is a marketing document. Mandated disclosure changes that by making omission a legal exposure rather than an editorial choice.

Where Anthropic's position diverges from its rivals

Anthropic has positioned itself, consistently and deliberately, as the frontier lab that favors binding rules — including rules that would constrain its own product roadmap. That is not a costless posture. It invites the criticism that a company advocating regulation is really advocating for barriers that inconvenience smaller competitors more than it inconveniences itself, a critique that has been leveled at incumbents in every regulated industry from banking to pharmaceuticals.

The counterargument is that Anthropic has repeatedly published material about its own models' failure modes rather than only about competitors' risks, which is a strange strategy for a firm interested purely in regulatory moat-building. Either way, the differentiation is real: other major developers have leaned harder on the argument that premature rules will slow American AI development relative to overseas rivals, and that voluntary frameworks plus existing consumer protection law are adequate for now.

Investors should read the divergence as a genuine strategic split rather than noise. It affects which labs will be advantaged by whichever regime emerges, how enterprise customers assess vendor risk, and how quickly capability releases reach the market.

Markets are still pricing AI on capability, not compliance

The regulatory argument has not yet translated into a compliance discount in equity prices. As of the last trade at 18:48 GMT on Sept. 16, 2026, the Nasdaq 100 tracker (NASDAQ: QQQ) was at $708.40, up 0.55% on the day against a prior close of $704.54, with an intraday range of $705.88 to $711.88 — the tech-heavy benchmark leading a mixed tape. The S&P 500 proxy (NYSEARCA: SPY) sat at $758.10, up 0.09% from $757.39, while the Dow 30 tracker (NYSEARCA: DIA) was lower at $519.20, down 0.39% from $521.23.

That pattern — technology up, industrials down — is the shape the AI trade has taken for most of this cycle. Nothing in it suggests the market is discounting the cost of a mandatory safety regime, because the market has no framework yet for what that cost would be. Disclosure mandates are cheap in absolute dollars. Deployment delays are not. Until draft language attaches specific obligations to specific model thresholds, there is no number for analysts to model.

Anthropic itself is privately held, so there is no direct way to trade a view on its policy stance. The exposure runs indirectly, through the listed cloud providers and chip suppliers whose revenue depends on the pace at which frontier models ship.

What would turn rhetoric into a rule

Three things are worth watching. First, whether any federal preemption effort arrives paired with a substantive safety standard, or as preemption alone. Second, whether mandated incident reporting survives drafting — it is the least burdensome mandate on the table and therefore the likeliest to pass, which makes it the best test of whether anything passes at all. Third, whether other frontier labs move toward or away from Heck's framing in their own public filings and testimony; a united industry front against binding rules is a very different lobbying proposition than a split one.

For now, the honor code remains the operating system of AI safety. Heck's argument is simply that nobody should mistake it for a system of accountability.

Key facts

• Who spoke: Sarah Heck, head of public policy at Anthropic
• Key remark: AI companies can't be expected to operate on an 'honor code'
• Nasdaq 100 (QQQ): $708.40, +0.55%, as of 18:48 GMT Sept. 16, 2026
• S&P 500 (SPY): $758.10, +0.09%, as of 18:48 GMT Sept. 16, 2026

Frequently asked questions

What did Sarah Heck actually say?

Heck, Anthropic's head of public policy, said AI companies cannot be expected to operate on an "honor code" — meaning that safety commitments which rely purely on a developer's own goodwill are not a substitute for enforceable obligations. She made the remarks while offering broader perspectives on AI safety as the regulatory debate intensifies.

Why is voluntary AI safety considered inadequate?

Voluntary safety work imposes real costs: delayed launches, withheld capabilities, and staff devoted to evaluation instead of capability research. Those costs fall hardest on the firm that takes them most seriously, handing time to competitors that do not. Critics argue that dynamic penalizes careful developers, so intent alone cannot sustain a safety regime.

What are the main AI regulatory disputes right now?

Two dominate. One is jurisdictional: state legislatures have moved faster than Congress, raising the prospect of a patchwork, which industry wants replaced by federal preemption while safety advocates warn preemption without a substantive federal standard leaves no standard. The other is whether safety testing and incident disclosure should be mandatory or discretionary.

How does Anthropic's stance differ from other AI developers?

Anthropic has consistently favored binding rules, including ones that would constrain its own roadmap. Other major developers have argued more strongly that premature regulation slows American AI development relative to overseas competitors, and that voluntary frameworks combined with existing consumer protection law are sufficient for the moment.

Can investors trade a view on Anthropic's policy position?

Not directly. Anthropic is privately held, so there is no listed equity. Exposure runs indirectly through publicly traded cloud providers, chip suppliers and enterprise software firms whose revenue depends on how quickly frontier models are released and deployed — which is precisely what a binding safety regime would affect.

How did markets trade on the day of the remarks?

As of the last trade at 18:48 GMT on Sept. 16, 2026, the Nasdaq 100 tracker QQQ was at $708.40, up 0.55%. The S&P 500 proxy SPY was at $758.10, up 0.09%, while the Dow 30 tracker DIA fell 0.39% to $519.20 — technology leading a mixed session.

Sources

• Anthropic policy chief says AI companies can't be expected to operate on 'honor code' — CNBC Technology

Photo: khezez | خزاز · Pexels Licence — source