The capture charge

On October 29, 2023, the day before the Biden administration issued its sweeping executive order on artificial intelligence, Yann LeCun replied on X to Max Tegmark of the Future of Life Institute:

Altman, Hassabis, and Amodei are the ones doing massive corporate lobbying at the moment. They are the ones who are attempting to perform a regulatory capture of the AI industry.

The accusation, reported within days by Business Insider, named the chief executives of OpenAI, Google DeepMind, and Anthropic by surname, and used a charge that has carried a technical meaning in economics since George Stigler formalized it in 1971. LeCun, who shared the 2018 Turing Award with Geoffrey Hinton and Yoshua Bengio, is the only laureate of that prize to have made the charge in those terms.

Hassabis answered within two days. “I pretty much disagree with most of those comments from Yann,” he told CNBC on October 31, and CNBC reported that he said DeepMind was not trying to achieve regulatory capture. His case was timing: the conversation about regulating very capable systems should start early. “I don’t think we want to be doing this on the eve of some of these dangerous things happening,” he said. “I think we want to get ahead of that.” The reply restated the safety case. It did not take up LeCun’s narrower point, that the proposed rules would favor the firms proposing them.

Regulatory capture is a specific claim with a specific shape. It alleges that the firms ostensibly being regulated have arranged the substance of the rules to entrench their own position, that the asks they bring to legislators, the comments they file in rule-making, and the standards they propose are calibrated to raise rivals’ costs while leaving their own operations substantially unconstrained. The test of the charge is the substance of what the firms have actually asked for, measured against what would serve their narrow commercial interest.

Sam Altman testified before the Senate Judiciary Subcommittee on Privacy, Technology, and the Law on May 16, 2023. He proposed a licensing regime for high-capability models, third-party audits of model behavior before deployment, and formal disclosure requirements covering both capabilities and limitations. Dario Amodei testified before the Senate Judiciary Subcommittee on July 25, 2023. Demis Hassabis has spoken in support of compute-based thresholds, mandatory pre-deployment capability evaluations, and international coordination on frontier-model governance, most prominently in the run-up to the UK AI Safety Summit at Bletchley Park in November 2023.

Across these appearances and the comments their companies have filed, four substantive asks recur: licensing or registration tied to capability or compute thresholds; pre-deployment safety evaluations conducted by the firm with disclosure of methodology; access for third-party evaluators with appropriate confidentiality protections; and post-deployment incident reporting. None of these asks are cost-free for the labs that proposed them. Pre-deployment evaluations require a research function dedicated to red-teaming and model assessment. Third-party audit access requires legal infrastructure, secure compute environments, and the willingness to share information with parties whose interests are not aligned with the firm’s commercial position. Compute reporting requires the disclosure of training-run details that, in any other industry, would be treated as competitive intelligence. Incident reporting creates legal exposure that did not previously exist.

The clearest single piece of evidence cuts directly against the simplest version of the capture story. Anthropic published its Responsible Scaling Policy in September 2023, six weeks before the LeCun post. As first written, it committed Anthropic to a sequence of capability evaluations conducted before training and again before deployment, with named thresholds at which the company would halt scaling in the absence of specified mitigations. No regulator had demanded this, and no statute required it. The policy imposed real costs on Anthropic. A pure capture strategy lobbies for rules that bind competitors and cost the lobbyist nothing it would not have done anyway.

The policy has changed since, and so has the law around it. Anthropic revised it as version 2.0 in October 2024. Version 3.0, published in February 2026, separated the safeguards Anthropic says it will pursue on its own from a broader map of capabilities and mitigations it recommends for the whole industry, and it describes the goals on the company’s own roadmap as nonbinding public targets. The current text is version 3.4, effective July 8, 2026. Some governments, meanwhile, require frontier developers to publish safety frameworks of their own. California’s SB 53, signed in September 2025, also requires them to report critical safety incidents, and since August 2026 the European Commission’s AI Office can require access to a model, run its own evaluations and order a provider to restrict availability. None of these rules writes Anthropic’s 2023 halt conditions into law.

The steel-man of the capture charge does not require ignoring any of this. It requires reading the asks at the level of implementation rather than at the level of principle. A licensing regime that triggers at, say, 10²⁶ training FLOPs binds the frontier labs and only the frontier labs, and the compliance infrastructure required to satisfy it is fixed-cost rather than marginal. A firm with eleven-figure annual revenue absorbs the fixed cost as a rounding error. A firm with eight-figure annual revenue does not. If the threshold is set at the right altitude, the labs that proposed it have, in effect, asked to be regulated in a way that prices a cohort of plausible challengers out of the market that would otherwise discipline them. The proposal can be sincere about the safety case and self-serving about the competitive case at the same time.

This is the version of the charge worth taking seriously, and the federal lobbying record bears on part of it. OpenAI reported $1.76 million of federal lobbying in 2024, up from $260,000 in 2023, and Anthropic $720,000, up from $280,000, according to OpenSecrets tallies of federal lobbying disclosures reported by TechCrunch in January 2025. The filings do not say which rules the money aimed at. Implementation details are where capture lives. The threshold at which a model counts as systemic-risk under the EU AI Act, the precise contents of the codes of practice, the operational definition of a “frontier” capability, the safe-harbor provisions for compliant disclosure: these are the levers that determine whether a published framework binds incumbents tightly, binds challengers tightly, or binds them differently.

A version of LeCun’s charge survives the obvious counter. It would have to be made in the form of a specific objection to a specific proposed rule, with a specific account of how the threshold or the disclosure regime or the audit standard would foreclose a specific class of competitors at the margin where competition matters. It would have to be filed somewhere: in a comment on a rule-making, in testimony before a legislator, in a published structured rebuttal in a venue that invites response. It has not been.

The figure who makes this point most uncomfortably is Yoshua Bengio. Bengio shared the 2018 Turing Award with LeCun and Hinton; he founded Mila in Montréal and has been its scientific advisor since March 2025; he holds no closed-weight commercial position; he was an early signatory of the March 2023 Future of Life Institute letter calling for a six-month pause on training runs above GPT-4 capability, and he has since coauthored a technical case, published in Science in 2024, for taking catastrophic risk seriously, including risks that do not require a sudden takeoff.

Bengio’s argument is not that the labs are pure of motive. It is that the safety concerns the labs invoke to support pre-deployment evaluation and capability thresholds are technically real, that the loss-of-control failure modes are not artifacts of marketing, and that the regulatory architecture being built around frontier models is approximately the architecture a disinterested technical analyst would recommend if presented with the same risk profile. Bengio reaches conclusions adjacent to those of Altman, Hassabis, and Amodei from a position with none of their commercial incentives. The capture frame does not have a clean account of him.

LeCun has argued with Bengio face to face. In June 2023 the two took opposite sides of the Munk Debate in Toronto on whether AI research poses an existential threat, LeCun with Melanie Mitchell and Bengio with Tegmark. That debate came before Bengio and his coauthors set out their case in Science. In April 2026 LeCun grouped Bengio with Hinton and himself as people not to listen to on AI and the labor market, and named economists to read instead. That is narrowly defensible on the labor question. It does not touch Bengio’s case for taking frontier-model safety seriously, which is not a labor-economics argument; it is a technical argument about loss of control.

LeCun has not made the capture charge in any forum designed to settle it. He has not testified against any of the licensing proposals he calls capture. He has not filed a comment in any of the rule-making proceedings that translate the published frameworks into binding text. He has not published a structured rebuttal of the Responsible Scaling Policy, or of the similar frameworks OpenAI and Google DeepMind adopted, though these documents purport to commit frontier labs to costs they would not otherwise bear. The charge lives on social media and the podcast circuit, in the medium where the cost of asserting an accusation and the cost of answering it are wildly asymmetric and the asymmetry runs in the accuser’s favor.

One explanation is that the charge as he holds it would not survive the operational discipline that a comment letter or a piece of testimony would impose. A formal filing requires identifying the specific rule, the specific competitive harm, the specific counterfactual in which the rule would not exist or would exist in a different form, and the specific evidence that the proposed rule was shaped to produce the harm rather than being a reasonable answer to a real risk. The X post does none of that work, because X does not require it.

LeCun also stands to gain commercially from the political force of the charge whether or not it survives technical scrutiny. He spent his last years at Meta championing the open-weight strategy that produced the Llama family, and his current company, AMI Labs, is structured around an open architectural research program. Open-weight strategies benefit, materially, from regulatory friction imposed on closed-weight competitors. This does not make the capture charge wrong. It does mean that the charge runs along the same axis as the commercial position of the person making it.

The executives, for their part, have not argued the charge at the level of the rules. Hassabis disagreed on CNBC and moved to the safety case. None of the three has written a structured defense of the licensing proposals or the Responsible Scaling Policy that addresses their competitive effects, which is where the charge would be settled.

A finer version of the charge — that implementation lobbying around published frameworks shapes thresholds and disclosure regimes in ways that disadvantage smaller challengers at the margin where competition matters — is consistent with the lobbying-disclosure record and with the structure of the rule-making process. Someone could write the comment letter that makes it about a specific proposed rule. The Turing laureate best positioned to do so has not.