This section assesses the regulatory policy of AI companies as an indicator of their willingness to promote their public interest mission over their private interest. At the outset, it is important to acknowledge that a pro-regulatory stance is far from a perfect proxy for promoting public interest. First, not all regulation is made equal. Companies may choose to support light-touch regulation to deflect calls for stronger laws. Poorly designed regulation can also be counter-productive, actually securing the discretion and upper hand of companies under a veneer of public legitimacy.120Maroussia Lévesque, (<)em(>)Smoke and Mirrors? AI Regulation and Corporate Power(<)/em(>), 2025 U. Ill. J. Law, Tech. & Pol. 33 (2025). With these caveats in mind, regulatory policy remains a good indicator of how companies align mission and action. It conveys the extent to which they are willing to carry out their mission by complying with guardrails that purport to safeguard human well-being against AI’s downsides—even if these guardrails are imperfect. In theory, companies purporting to benefit humanity should embrace regulation. But their responses in practice say otherwise. They have been mixed at best, implying that other considerations take precedence. For example, the race among a handful of frontier labs suggests that safety takes a backseat to rapid deployment and large-scale adoption. Changing political tides in the second Trump administration and in the EU have led many companies to shed nominal calls for regulation, suggesting that these positions were more tactical than genuine.121AI Now Institute, (<)em(>)Recasting Regulation as a Barrier to Innovation(<)/em(>), in Artificial Power: 2025 Landscape Report 32 (2025), https://ainowinstitute.org/publications/1-4-recasting-regulation-as-a-barrier-to-innovation; Amnesty EU, (<)em(>)EU: AI Omnibus Deal Lawmakers Should Reject a Rollback of AI Safeguards(<)/em(>), June 11, 2026, (<)a href='https://www.amnesty.eu/news/eu-ai-omnibus-deal-lawmakers-should-reject-a-rollback-of-ai-safeguards/'(>)https://www.amnesty.eu/news/eu-ai-omnibus-deal-lawmakers-should-reject-a-rollback-of-ai-safeguards/(<)/a(>). To provide concrete examples of how companies engage with lawmakers, this section outlines their respective position on regulation based on a sample of publicly available statements. Though initiatives to regulate AI abound, the EU AI Act, California’s AI Safety regulation and the U.S. federal ban on state regulation stand out as important lenses to assess how each company under review reacts to regulatory proposals.122Artificial Intelligence Act, Regulation (EU) 2024/1689 of the European Parliament and of the Council, 2024 O.J. (L 1689); Safe and Secure Innovation for Frontier Artificial Intelligence Models Act, SB-1047 (<)a href='https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202320240SB1047'(>)https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202320240SB1047(<)/a(>) (Cal. 2024) [hereinafter SB-1047]; H.R. Con. Res. 14, 119th Cong. § 43201(c) (2025), (<)a href='https://d1dth6e84htgma.cloudfront.net/Subtitle_C_Communications_4e3fbcc3bc.pdf'(>)https://d1dth6e84htgma.cloudfront.net/Subtitle_C_Communications_4e3fbcc3bc.pdf(<)/a(>) [(<)a href='https://perma.cc/8QQW-G86Y'(>)https://perma.cc/8QQW-G86Y(<)/a(>)].
EU Artificial Intelligence Act
Despite multiple loopholes and perhaps inflated impact, the EU AIA remains the first comprehensive effort to regulate AI models and applications (models and systems in the Act’s parlance).123Lévesque, (<)em(>)supra(<)/em(>) note 120 (concluding that the AIA affords AI companies significant workarounds and discretion in interpreting obligations through self-assessments). Industry representatives actively thwarted the regime applicable to general purpose models, which is informative as to industry’s actual willingness to abide by regulation.124Council of the European Union, (<)em(>)Proposal for a Regulation of the European Parliament and of the Council Laying down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) and Amending Certain Union Legislative Acts – Second Presidency Compromise Text arts. 4a-4c (2022)(<)/em(>), (<)a href='https://data.consilium.europa.eu/doc/document/ST-11124-2022-INIT/en/pdf'(>)https://data.consilium.europa.eu/doc/document/ST-11124-2022-INIT/en/pdf(<)/a(>) (imposing high-risk obligations on general purpose systems by default); European Parliament, (<)em(>)Amendments Adopted 14 June 2023 on the Proposal for a Regulation of the European Parliament and of the Council on Laying down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) and Amending Certain Union Legislative Acts (COM(2021)0206 – C9-0146/2021 – 2021/0106(COD))(<)/em(>), P9_TA(2023)0236, (<)a href='https://www.europarl.europa.eu/doceo/document/TA-9-2023-0236_EN.docx'(>)https://www.europarl.europa.eu/doceo/document/TA-9-2023-0236_EN.docx(<)/a(>), [(<)a href='https://perma.cc/2HJF-KABH'(>)https://perma.cc/2HJF-KABH(<)/a(>)] amendment 399, creating article 28(b) requiring foundation models to undergo pre-deployment risk assessments, energy reduction, public registration and transparency measures; Camille Schyns, Corp. Eur. Observatory, The Lobbying Ghost in the Machine: Big Tech’s Covert Defanging of Europe’s AI Act (Feb. 2023), (<)a href='https://corporateeurope.org/en/2023/02/lobbying-ghost-machine'(>)https://corporateeurope.org/en/2023/02/lobbying-ghost-machine(<)/a(>).
California Safety Regulation
In California, the governor vetoed the safety-oriented SB-1047 bill despite weakened obligations in the final version,125For a summary of amendments in the major revision in August 2024, (<)em(>)see(<)/em(>) Nathan Calvin, (<)em(>)SB 1047 August 15 Author Amendments Overview(<)/em(>), SB 1047 – Safe & Secure AI Innovation, (<)a href='https://safesecureai.org/amendments'(>)https://safesecureai.org/amendments(<)/a(>) [(<)a href='https://web.archive.org/web/20250420231105/https://safesecureai.org/amendments'(>)https://web.archive.org/web/20250420231105/https://safesecureai.org/amendments(<)/a(>)] (narrowing civil liability prior to harm and eliminating new oversight agency); for the full legislative history, (<)em(>)see(<)/em(>) (<)em(>)California SB1047 | 2023-2024 | Regular Session(<)/em(>), LegiScan, (<)a href='https://legiscan.com/CA/text/SB1047/id/3019694'(>)https://legiscan.com/CA/text/SB1047/id/3019694(<)/a(>) (last visited June 23, 2025). in no small part due to industry pressure.126Transcript of Senate Standing Committee on Governmental Organization (Apr. 23, 2024) (<)a href='https://calmatters.digitaldemocracy.org/hearings/257812'(>)https://calmatters.digitaldemocracy.org/hearings/257812(<)/a(>) (last visited July 24, 2025) (containing testimony of Dylan Hoffmann on behalf of the TechNet industry group). For a full list of members including OpenAI, Anthropic, and Google, see (<)em(>)Members – TechNet(<)/em(>), (<)a href='https://www.technet.org/our-story/members/'(>)https://www.technet.org/our-story/members/(<)/a(>) (last visited Aug. 6, 2025); Letter from Meghan Pensyl, Director of Policy, Business Software Alliance, to Rebecca Bauer-Cohen, Capital Office (June 11, 2024), (<)a href='https://www.bsa.org/files/policy-filings/06112024bsacasb1047.pdf'(>)https://www.bsa.org/files/policy-filings/06112024bsacasb1047.pdf(<)/a(>); (<)em(>)Founder-Led Statement on SB-1047(<)/em(>) (2024), (<)a href='https://static.politico.com/95/0a/a317efe44616af436ce6a4f32647/founder-led-statement-on-sb1047-june-20-2024-2.pdf'(>)https://static.politico.com/95/0a/a317efe44616af436ce6a4f32647/founder-led-statement-on-sb1047-june-20-2024-2.pdf(<)/a(>); Khari Johnson, (<)em(>)Why Silicon Valley Is Trying so Hard to Kill This AI Bill in California(<)/em(>), CalMatters (Aug. 12, 2024), (<)a href='http://calmatters.org/economy/technology/2024/08/ai-regulation-showdown/'(>)http://calmatters.org/economy/technology/2024/08/ai-regulation-showdown/(<)/a(>); Garrison Lovely, (<)em(>)California’s AI Safety Bill Is a Mask-Off Moment for the Industry(<)/em(>), Nation (Aug. 15, 2024), (<)a href='https://www.thenation.com/article/society/california-ai-safety-bill/'(>)https://www.thenation.com/article/society/california-ai-safety-bill/(<)/a(>). Setting aside real-life harms experienced as catastrophic, the bill included pre-deployment safety testing, third-party audits and full shutdown capabilities for a narrow slice of chemical, biological and nuclear weapons causing mass casualties.127SB-1047, (<)em(>)supra(<)/em(>) note 122, § 22602 (g). The subsequently enacted Transparency in Frontier Artificial Intelligence Act (SB-53) toned down the obligations from liability to transparency, doing away with third party auditing, incident reporting within 72 hours and requiring full shutdown capabilities.128Transparency in Frontier Artificial Intelligence Act (SB-53) of 2025, C–138. For a fuller comparison of SB-147 and SB-53, (<)em(>)see(<)/em(>) Aden Hizikas, (<)em(>)Why California’s SB 53 Still Gets AI Regulation Wrong: And How It Compares to Last Year’s SB 1047(<)/em(>), Chamber of Progress (July 9, 2025), (<)a href='https://progresschamber.org/insights/why-californias-sb-53-still-gets-ai-regulation-wrong/'(>)https://progresschamber.org/insights/why-californias-sb-53-still-gets-ai-regulation-wrong/(<)/a(>). Taken together, the actions of AI companies on European and California regulation show a pattern of diluting legislative drafts that were light touch to begin with.
Federal Moratorium on State Regulation
Finally, AI companies opined on a federal ban on state regulation.129Cecilia Kang, (<)em(>)Defeat of a 10-Year Ban on State A.I. Laws Is a Blow to Tech Industry(<)/em(>), N.Y. Times (July 1, 2025), (<)a href='https://www.nytimes.com/2025/07/01/us/politics/state-ai-laws.html'(>)https://www.nytimes.com/2025/07/01/us/politics/state-ai-laws.html(<)/a(>) (reporting that industry interests “lobbied fiercely” in favor of the ban). Their views ranged from supporting to opposing the moratorium on state laws. Though the ban ultimately failed, it is a live issue at the time of writing as the White House and legislators continue to pursue it.130AI.Gov, America’s AI Action Plan 3 (2025), (<)a href='https://perma.cc/8RK8-JZVN'(>)https://perma.cc/8RK8-JZVN(<)/a(>) (limiting federal funding to states with “burdensome” AI regulations); Anthony Adragna, (<)em(>)Guthrie Says GOP Will Try Again on AI Moratorium(<)/em(>), PoliticoPro (July 2, 2025), (<)a href='https://subscriber.politicopro.com/article/2025/07/guthrie-says-gop-will-try-again-on-ai-moratorium-00437707'(>)https://subscriber.politicopro.com/article/2025/07/guthrie-says-gop-will-try-again-on-ai-moratorium-00437707(<)/a(>).
Despite common industry trends, the three companies under review vary regarding their regulatory stance in the selected case studies. As compared to OpenAI and Google systematically opposing regulation, Anthropic engages more constructively. Yet ultimately, no company prioritizes the public interest over their own.
1. Google: Mixed Signals
Alphabet’s messaging is that “AI is too important not to regulate, and too important not to regulate well.”131Sundar Pichai, (<)em(>)Google CEO: Building AI Responsibly Is the Only Race That Really Matters(<)/em(>), Fin. Times (May 23, 2023), (<)a href='https://www.ft.com/content/8be1a975-e5e0-417d-af51-78af17ef4b79'(>)https://www.ft.com/content/8be1a975-e5e0-417d-af51-78af17ef4b79(<)/a(>); Ina Fried, (<)em(>)Google Says Now Is the Time to Plan for AGI Safety(<)/em(>), Axios, Apr. 2, 2025, (<)a href='https://www.axios.com/2025/04/02/google-agi-deepmind-safety'(>)https://www.axios.com/2025/04/02/google-agi-deepmind-safety(<)/a(>) (calling AI “a very powerful technology, [that] can and should be regulated”); Sundar Pichai, (<)em(>)Why Google Thinks We Need to Regulate AI(<)/em(>), Fin. Times, Jan. 20, 2020, (<)a href='https://www.ft.com/content/3467659a-386d-11ea-ac3c-f68c10993b04'(>)https://www.ft.com/content/3467659a-386d-11ea-ac3c-f68c10993b04(<)/a(>). But as with other AI companies, it has sought to weaken even the most baseline regulatory proposals. Providing feedback to a White Paper paving the way for the EU AIA, Google has expressed several concerns about the scope and intensity of AI regulation. Regulators integrated constructive comments including a better allocation of responsibility between developers and deployers,132Google, Consultation on the White Paper on AI – a European Approach: Google’s Submission 18 (2020), (<)a href='https://blog.google/documents/77/Googles_submission_to_EC_AI_consultation_1.pdf'(>)https://blog.google/documents/77/Googles_submission_to_EC_AI_consultation_1.pdf(<)/a(>) [https://perma.cc/JB87-JXSF]. but a significant number of propositions sought to dilute important aspects of the proposed regulatory framework. For instance, Google advocated for narrowing the scope of high-risk systems subject to stricter oversight, for internal instead of external evaluations, and for heavier burdens of proof for the victims of AI harms.133(<)em(>)Id.(<)/em(>) at 2, 33, 44. Its subsequent feedback on the bill detailed these and other concerns about provisions Google deemed uncertain or unrealistic.134Google, Consultation on the EU AI Act Proposal: Google’s Submission (2021), (<)a href='https://ec.europa.eu/info/law/better-regulation/have-your-say/initiatives/12527-Artificial-intelligence-ethical-and-legal-requirements/F2662492_en'(>)https://ec.europa.eu/info/law/better-regulation/have-your-say/initiatives/12527-Artificial-intelligence-ethical-and-legal-requirements/F2662492_en(<)/a(>).(<)br(>) In a letter about California’s SB-1047 bill, it commented that the “EU AIA’s extensive requirements on general-purpose AI are not a good model to emulate.”135Letter from Alice Friend, Head of AI & Emerging Tech Policy, Google, to Thomas Umberg, Chair, California Senate Judiciary Committee (July 3, 2024) (<)a href='https://www.documentcloud.org/documents/25034111-the-honorable-senator-umberg-senate-bill-1047/'(>)https://www.documentcloud.org/documents/25034111-the-honorable-senator-umberg-senate-bill-1047/(<)/a(>).
Turning to its position on California’s SB-1047 bill, the company listed a number of concerns about requirements it considered overly restrictive, unclear and unrealistic. For instance, it disapproved of imposing obligations prior to training and of limitations preventing developers from using systems even if they have safeguards in place, criticizing “impossible pre-training assessments and controls of model costs, size, and potential for harm; ban[ning] release of AI models if they can be misused; or creat[ing] penalty schemes that destroy or shut down models if misused by third parties.”136(<)em(>)Id.(<)/em(>) Like Anthropic, Google suggested concrete language to improve the bill.
Google also belonged to a trade group137INCOMPAS, (<)em(>)Internet Business & Telecom Network Member List(<)/em(>), (<)a href='https://membership.incompas.org/memberlist.asp'(>)https://membership.incompas.org/memberlist.asp(<)/a(>) [(<)a href='https://web.archive.org/web/20251213102210/https://membership.incompas.org/memberlist.asp'(>)https://web.archive.org/web/20251213102210/https://membership.incompas.org/memberlist.asp(<)/a(>)] (last visited Aug. 7, 2025). that at the time supported a ban on state laws; the group’s CEO stated in June 2025 that a ten-year moratorium on state AI laws was “the right policy at the right time for American leadership,” invoking the importance of a unified front against China.138Letter from Chip Pickering, CEO, INCOMPAS, to Rep. Brett Guthrie, Chairman, H. Comm. on Energy & Com. (May 19, 2025), (<)a href='https://incompas.org/wp-content/uploads/2025/05/INCOMPAS-Letter-of-Support-of-Reconcilation-Bill-5.19.2025.pdf'(>)https://incompas.org/wp-content/uploads/2025/05/INCOMPAS-Letter-of-Support-of-Reconcilation-Bill-5.19.2025.pdf(<)/a(>) (supporting “temporary moratorium on burdensome state laws”); Alex Rogers & Stephen Morris, (<)em(>)Big Tech Pushes for 10-Year Ban on US States Regulating AI(<)/em(>), Fin. Times (June 18, 2025), (<)a href='https://www.ft.com/content/52ae52f1-531e-462f-898f-e9f86b3b1869'(>)https://www.ft.com/content/52ae52f1-531e-462f-898f-e9f86b3b1869(<)/a(>). Google’s chief legal officer and President of Global Affairs has also expressed broader concerns about a patchwork of regulation, noting that state by state regulation “has left people with uneven protections based on where they live, and made it harder for small businesses to navigate conflicting laws.”139Kent Walker, (<)em(>)A Patchwork of Rules and Regulations Won’t Cut It for AI(<)/em(>), Hill (Nov. 5, 2023), (<)a href='https://thehill.com/opinion/technology/4292625-a-patchwork-of-rules-and-regulations-wont-cut-it-for-ai/'(>)https://thehill.com/opinion/technology/4292625-a-patchwork-of-rules-and-regulations-wont-cut-it-for-ai/(<)/a(>). The company has also implicitly endorsed a unified federal approach over state-specific regulations in its feedback on California’s AI Safety bill, where it suggested language that tracked a Biden presidential order.140S.B. 1047, (<)em(>)supra(<)/em(>) note 135, at 6 (proposing that reporting requirements track the federal Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence).
2. OpenAI: Calling for Regulation; Lobbying Against It
The company gestures towards regulation, but its position on specific proposals says otherwise. Testifying in Congress in 2023, its CEO stated that “we think that regulatory intervention by governments will be critical to mitigate the risks of increasingly powerful models,” suggesting mandatory licensing and testing for frontier models.141(<)em(>)Oversight of A.I.: Rules for Artificial Intelligence: Hearing Before the Subcomm. on Priv., Tech., & L.(<)/em(>), 118th Cong. 7 (2023) (statement of Samuel Altman, Chief Executive Officer, OpenAI), (<)a href='https://www.govinfo.gov/app/details/CHRG-118shrg52706'(>)https://www.govinfo.gov/app/details/CHRG-118shrg52706(<)/a(>); (<)em(>)see also(<)/em(>) (<)em(>)Written Testimony of Sam Altman Chief Executive Officer OpenAI Before the U.S. Senate Committee on the Judiciary Subcommittee on Privacy, Technology, & the Law(<)/em(>), 11 (2023) (reiterating that “regulation of AI is essential”); Craig S. Smith, (<)em(>)OpenAI’s Sam Altman To Congress: Regulate Us, Please!(<)/em(>), Forbes (May 16, 2023), (<)a href='https://www.forbes.com/sites/craigsmith/2023/05/16/openais-sam-altman-to-congress-regulate-us-please/'(>)https://www.forbes.com/sites/craigsmith/2023/05/16/openais-sam-altman-to-congress-regulate-us-please/(<)/a(>); Cecilia Kang, (<)em(>)OpenAI’s Sam Altman Urges A.I. Regulation in Senate Hearing(<)/em(>), N.Y. Times (May 16, 2023), (<)a href='https://www.nytimes.com/2023/05/16/technology/openai-altman-artificial-intelligence-regulation.html'(>)https://www.nytimes.com/2023/05/16/technology/openai-altman-artificial-intelligence-regulation.html(<)/a(>). Yet when concrete proposals to do just that arise, OpenAI vehemently opposes them. Its stance on the EU AI Act, California’s SB-1047 bill and more recently any state legislation demonstrates a pattern of systematically derailing AI regulation, one that is in stark contrast with earlier calls for regulation.
OpenAI grew increasingly opposed to the EU AIA over the years-long negotiation process. It broadly endorsed the initial draft which divided AI applications into unacceptable, high and low-risk categories, but also raised concerns about defining a general-purpose system by its riskiest use case and later explicitly argued that a general purpose system such as GPT-3 should not be banned simply because some of its use cases would constitute unacceptable risk.142OpenAI, (<)em(>)Comments on European Union Proposal on AI Rules and Actions(<)/em(>) 1 (Aug. 2, 2021), (<)a href='https://ec.europa.eu/info/law/better-regulation/have-your-say/initiatives/12527-Artificial-intelligence-ethical-and-legal-requirements/F2665231_en'(>)https://ec.europa.eu/info/law/better-regulation/have-your-say/initiatives/12527-Artificial-intelligence-ethical-and-legal-requirements/F2665231_en(<)/a(>) (broadly agreeing with risk-based approach); OpenAI, (<)em(>)OpenAI White Paper on the European Union’s Artificial Intelligence Act(<)/em(>) 2–3 (Sep. 2022), (<)a href='https://www.scribd.com/document/829025960/openai-response-aia-white-paper'(>)https://www.scribd.com/document/829025960/openai-response-aia-white-paper(<)/a(>) [(<)a href='https://perma.cc/R48V-6GZ6'(>)https://perma.cc/R48V-6GZ6(<)/a(>)] (arguing that GPT-3 is not high-risk per se); Billy Perrigo, (<)em(>)Exclusive: OpenAI Lobbied E.U. to Water Down AI Regulation(<)/em(>), TIME (June 20, 2023), (<)a href='https://time.com/6288245/openai-eu-lobbying-ai-act/'(>)https://time.com/6288245/openai-eu-lobbying-ai-act/(<)/a(>). When the EU Council introduced amendments that would make models high-risk by default in 2022,143Council of the European Union, (<)em(>)supra(<)/em(>) note 124, arts. 4a-4c. OpenAI pushed back. Arguing that its models were not high-risk per se and that the exemptions were too narrow, it contended that voluntary scrutiny was sufficient to deal with the risks of these models deployed in high-stakes contexts.144OpenAI White Paper, (<)em(>)supra(<)/em(>) note 142, at 2. In 2023, the European Parliament introduced a separate regime for foundation models such as those developed by OpenAI. Obligations included pre-deployment risk assessments, energy reduction, public registration and transparency measures.145Artificial Intelligence Act, Regulation (EU) 2024/1689, (<)em(>)supra(<)/em(>) note 122, amendment 399, creating article 28(b). Around the same time, OpenAI threatened to withdraw ChatGPT from the EU. It deemed the EU to be “over-regulating” but noted that “we have heard it’s going to get pulled back,” referring to a draft of the EU AI Act.146Reuters, (<)em(>)OpenAI May Leave the EU If Regulations Bite – CEO(<)/em(>), Reuters (May 24, 2023), (<)a href='https://www.reuters.com/technology/openai-may-leave-eu-if-regulations-bite-ceo-2023-05-24/'(>)https://www.reuters.com/technology/openai-may-leave-eu-if-regulations-bite-ceo-2023-05-24/(<)/a(>) [(<)a href='https://perma.cc/9HKR-SFGE'(>)https://perma.cc/9HKR-SFGE(<)/a(>)]. Once the Act was finalized, the company committed to its voluntary Code of Practice for General Purpose AI along with other companies including Anthropic and Google.147 (<)em(>)The EU Code of Practice and Future of AI in Europe(<)/em(>) (July 11, 2025), (<)a href='https://openai.com/global-affairs/eu-code-of-practice/'(>)https://openai.com/global-affairs/eu-code-of-practice/(<)/a(>); (<)em(>)A Primer on the EU AI Act: What It Means for AI Providers and Deployers(<)/em(>), July 30, 2024, (<)a href='https://openai.com/global-affairs/a-primer-on-the-eu-ai-act/'(>)https://openai.com/global-affairs/a-primer-on-the-eu-ai-act/(<)/a(>) (last visited Aug. 5, 2025); (<)em(>)Anthropic to Sign the EU Code of Practice(<)/em(>), Anthropic (July 21, 2025), (<)a href='https://www.anthropic.com/news/eu-code-practice'(>)https://www.anthropic.com/news/eu-code-practice(<)/a(>) (last visited Aug. 5, 2025); Google, (<)em(>)We Will Sign the EU AI Code of Practice(<)/em(>), Google (July 30, 2025), (<)a href='https://blog.google/around-the-globe/google-europe/eu-ai-code-practice/'(>)https://blog.google/around-the-globe/google-europe/eu-ai-code-practice/(<)/a(>).
The company adopted a similar position against California’s efforts to clarify the liability of developers for catastrophic risks incurred by frontier models. The essence of its objections was twofold: a harmonized federal regime is preferable to a state-by-state approach, and the bill would stifle innovation in California.148 Jason Kwon, (<)em(>)Re: Senate Bill 1047 (Wiener) – Safe and Secure Innovation for Frontier Artificial Intelligence Models Act – OPPOSE(<)/em(>), (2024), (<)a href='https://www.documentcloud.org/documents/25056617-ca-sb-1047-openai-opposition-letter/'(>)https://www.documentcloud.org/documents/25056617-ca-sb-1047-openai-opposition-letter/(<)/a(>). It also criticized an earlier, more robust version of the bill through the industry group Software Alliance. The letter decried almost all the aspects of the bill including its scope, overly prescriptive obligations, unclear exemptions, responsibility for downstream uses, short incident reporting window, vague mandate for an inexperienced new public agency, and unjustified criminal liability for chief technology officers.149Pensyl, (<)em(>)supra(<)/em(>) note 126. The goal is not to litigate the appropriateness of SB-1047; undoubtedly the proposal had room for improvement, especially early in the drafting process. But OpenAI’s consistent opposition, even as subsequent versions relaxed liability and removed the new public agency,150The above mentioned OpenAI letter at note 148 is dated August 21, 2024. It does not specify which version of the bill is criticized, but a revised version dated August 19, 2024 addresses many concerns notably by limiting pre-harm liability and eliminating the new agency. (<)em(>)See(<)/em(>) Amendment overview, (<)em(>)supra(<)/em(>) note 125. shows a reluctance to endorse regulation at odds with its high-level statements.
OpenAI took additional steps to neutralize state AI regulation. Framing AI regulation as a federalism issue, its letter against SB-1047 decried a state-by-state approach to federal matters with national security implications.151Kwon, (<)em(>)supra(<)/em(>) note 148. Its submission to the Office of Science and Technology about the White House AI Action plan advocated for a “[g]uarantee that state-based legislation does not undermine America’s innovation lead on AI,” with “liability protections including preemption from state-based regulations that focus on frontier model security (e.g., CA SB 1047). This will help keep the U.S. public and private sectors competitive by allowing AI companies of all sizes to pursue bleeding-edge AI technology free from the regulatory uncertainty created by some state-based liability regimes.”152Letter from Chris Lehane, OpenAI, to Faizal D’Souza, Office of Science and Technology Policy (Mar. 13, 2025), (<)a href='https://cdn.openai.com/global-affairs/ostp-rfi/ec680b75-d539-4653-b297-8bcf6e5f7686/openai-response-ostp-nsf-rfi-notice-request-for-information-on-the-development-of-an-artificial-intelligence-ai-action-plan.pdf'(>)https://cdn.openai.com/global-affairs/ostp-rfi/ec680b75-d539-4653-b297-8bcf6e5f7686/openai-response-ostp-nsf-rfi-notice-request-for-information-on-the-development-of-an-artificial-intelligence-ai-action-plan.pdf(<)/a(>) [(<)a href='https://perma.cc/7ATA-FDL3'(>)https://perma.cc/7ATA-FDL3(<)/a(>)]; (<)em(>)see also(<)/em(>) Chris Lehane, LinkedIn (June 23, 2025), (<)a href='https://www.linkedin.com/feed/update/urn:li:activity:7342879984203591680/'(>)https://www.linkedin.com/feed/update/urn:li:activity:7342879984203591680/(<)/a(>) [(<)a href='https://perma.cc/TV6C-RKAT'(>)https://perma.cc/TV6C-RKAT(<)/a(>)] (last visited June 30, 2025).
In theory, the federalism argument has some purchase as regards the patchwork of state regulation raising compliance costs. But for the most part, the proposed laws that OpenAI so vehemently opposed entailed minimal burdens such as self-assessments and auditing, especially for a large company with deep resources and expertise. Moreover, OpenAI’s position seems to have evolved to take issue even with a unified federal law. It subsequently favored a “regulatory strategy that ensures the freedom to innovate,”153Letter from Chris Lehane, (<)em(>)supra(<)/em(>) note 152, at 3. a far cry from its calls for regulation two years prior. Furthermore, it advocated for a voluntary federal framework, as opposed to binding regulation. Overall, the constant in OpenAI’s regulatory stance is to change with the political trend of the day.154The company pivoted once again, articulating a “reverse federalism” lobbying and legislative strategy in June 2026 pushing for state alignment and subsequent federal adoption. (<)em(>)See(<)/em(>) (<)em(>)A Blueprint for Democratic Governance of Frontier AI(<)/em(>), OpenAI (Sep. 30, 2026), (<)a href='https://openai.com/index/frontier-safety-blueprint/'(>)https://openai.com/index/frontier-safety-blueprint/(<)/a(>); (<)em(>)The US Is Advancing AI Safety Through State and Federal Action(<)/em(>), OpenAI (Sep. 30, 2026), (<)a href='https://openai.com/index/advancing-ai-safety-through-state-and-federal-action/'(>)https://openai.com/index/advancing-ai-safety-through-state-and-federal-action/(<)/a(>). It has proven adept at tracking the (de)regulatory mood, abandoning virtue-signaling calls for regulation when no longer politically expedient.
3. Anthropic: Cautious Support for Targeted Regulation
Anthropic’s position boils down to supporting surgical, gradual regulation that incentivizes AI companies to avert serious incidents. There is scant public information concerning the company’s position on the EU AIA. Founded in January 2021, the company was just getting off the ground during the initial consultation period,155(<)em(>)European Commission – Commission Adoption(<)/em(>), European Commission – Have your say (Sep. 10, 2020), (<)a href='https://ec.europa.eu/info/law/better-regulation/have-your-say/initiatives/12527-Artificial-intelligence-ethical-and-legal-requirements_en'(>)https://ec.europa.eu/info/law/better-regulation/have-your-say/initiatives/12527-Artificial-intelligence-ethical-and-legal-requirements_en(<)/a(>) (Feedback period 26 April 2021 – 06 August 2021). which perhaps explains why it does not appear in the list of feedback contributors.156(<)em(>)European Commission – Feedback Received(<)/em(>), European Commission – Have your say (Aug. 6, 2021), (<)a href='https://ec.europa.eu/info/law/better-regulation/have-your-say/initiatives/12527-Artificial-intelligence-ethical-and-legal-requirements/feedback_en?p_id=24212003'(>)https://ec.europa.eu/info/law/better-regulation/have-your-say/initiatives/12527-Artificial-intelligence-ethical-and-legal-requirements/feedback_en?p_id=24212003(<)/a(>) (no results when searching “Anthropic”). Absent comprehensive and focused feedback on the Act, any public facing statements are inconclusive at best. With that limitation in mind, the company expressed concerns about the AIA after it was finalized in the spring of 2025. Its EU policy lead noted that “[w]e deeply believe there are risks to these tools and models. I think the risk would be to hamper innovation and, through regulation, disable European companies from making the most out of these technologies. I think that is something that the EU should be careful with.”157(<)em(>)“Large Companies Need Safety,” Says Anthropic’s European Boss(<)/em(>), Tech.eu (July 9, 2025), (<)a href='https://tech.eu/2025/07/09/we-re-not-looking-to-poach-researchers-from-other-labs-says-anthropic-s-european-boss/'(>)https://tech.eu/2025/07/09/we-re-not-looking-to-poach-researchers-from-other-labs-says-anthropic-s-european-boss/(<)/a(>); (<)em(>)see also(<)/em(>) (<)em(>)‘The Risk Is Setting Back Innovation’—Anthropic EMEA Head Warns of AI Regulation Impact on Europe(<)/em(>), (May 28, 2025), (<)a href='https://www.businesspost.ie/tech/the-risk-is-setting-back-innovation-anthropic-emea-head-warns-of-ai-regulation-impact-on-europe/'(>)https://www.businesspost.ie/tech/the-risk-is-setting-back-innovation-anthropic-emea-head-warns-of-ai-regulation-impact-on-europe/(<)/a(>). In July 2025, the company signed the voluntary Code of Practice implementing the Act’s obligations pertaining to general purpose models.158(<)em(>)Anthropic to sign the EU Code of Practice(<)/em(>), Anthropic, (<)em(>)supra(<)/em(>) note 147.
Anthropic took a hands-on approach to safety regulation. In July 2024, it proposed amendments to narrow and clarify SB-1047.159Hank Dempsey, (<)em(>)RE: SB 1047 (Wiener) – Support If Amended(<)/em(>), (July 23, 2024). One major suggestion pertained to the liability of model developers for catastrophic risks. Anthropic proposed compensating harm after the fact rather than deterring it: “instead of deciding what measures companies should take to prevent catastrophes (which are still hypothetical and where the ecosystem is still iterating to determine best practices), focus the bill on holding companies responsible for causing actual catastrophes.”160(<)em(>)Id.(<)/em(>) One observer described this shift from pre-harm enforcement to outcomes-based liability as a commitment to “take no action until something has already gone terribly wrong.”161Sigal Samuel, (<)em(>)It’s practically impossible to run a big AI company ethically(<)/em(>), Vox (Aug. 2, 2024), (<)a href='https://www.vox.com/future-perfect/364384/its-practically-impossible-to-run-a-big-ai-company-ethically'(>)https://www.vox.com/future-perfect/364384/its-practically-impossible-to-run-a-big-ai-company-ethically(<)/a(>). Legislators incorporated most of Anthropic’s feedback, including significantly reducing pre-harm enforcement.162SB-1047, (<)em(>)supra(<)/em(>) note 122 (compare July 7, 2024 to Aug. 19, 2024 versions, especially ¶ 22606 on liability); (<)em(>)see also(<)/em(>) (<)em(>)California’s SB1047 vs EU AI Act: A Comparative Analysis of AI Regulation(<)/em(>), (Oct. 28, 2024), https://www.sciencespo.fr/public/chaire-numerique/en/2024/10/28/californias-sb1047-vs-eu-ai-act-a-comparative-analysis-of-ai-regulation/ (outlining the debate about the extent of developer liability in the bill); Amendment overview, (<)em(>)supra(<)/em(>) note 125. At the final stage requiring the Governor’s approval, the company cautiously endorsed the bill, stating that “we believe its benefits likely outweigh its costs.”163Dario Amodei, (<)em(>)Anthropic Letter (SB 1047)(<)/em(>), (2024), (<)a href='https://perma.cc/67SG-RXRL'(>)https://perma.cc/67SG-RXRL(<)/a(>). Anthropic’s engagement is significant because it shifted the discussion from an initial pre-harm enforcement approach to a lighter form of procedural accountability attaching liability only once harm materialized.
As for the proposed moratorium on local AI laws, the company took a stance against the proposed ban on subnational, state-level regulation. Acknowledging the downsides of a patchwork of inconsistent local laws, it nevertheless deemed the ten-year moratorium “far too blunt an instrument” in light of the lack of federal response matching the pace of AI development.164Dario Amodei, (<)em(>)Anthropic C.E.O.: Don’t Let A.I. Companies off the Hook(<)/em(>), N.Y. Times (June 5, 2025), (<)a href='https://www.nytimes.com/2025/06/05/opinion/anthropic-ceo-regulate-transparency.html'(>)https://www.nytimes.com/2025/06/05/opinion/anthropic-ceo-regulate-transparency.html(<)/a(>). Instead, the company proposed a transparency framework to garner bipartisan support as a first step to share best practices across the industry.165(<)em(>)Id.(<)/em(>); (<)em(>)The Need for Transparency in Frontier AI(<)/em(>), Anthropic (July 7, 2025), (<)a href='https://www.anthropic.com/news/the-need-for-transparency-in-frontier-ai'(>)https://www.anthropic.com/news/the-need-for-transparency-in-frontier-ai(<)/a(>) (proposing a “targeted transparency framework” applying “only to the largest AI systems and developers”). Reacting to the White House’s AI Action Plan reiterating limits on state regulation, the company’s CEO noted that “we continue to oppose proposals aimed at preventing states from enacting measures to protect their citizens from potential harms caused by powerful AI systems, if the federal government fails to act” and further elaborated on mandatory transparency requirements as a first item of federal action.
166(<)em(>)Thoughts on America’s AI Action Plan(<)/em(>), Anthropic (July 23, 2025), (<)a href='https://www.anthropic.com/news/thoughts-on-america-s-ai-action-plan'(>)https://www.anthropic.com/news/thoughts-on-america-s-ai-action-plan(<)/a(>) (last visited July 25, 2025).
Aside from positions on these specific instances of regulation, the company purports to turn self-governance practices into binding regulation. It frames voluntary scaling policies and third-party testing as regulatory prototypes,167(<)em(>)Dario Amodei’s Prepared Remarks from the AI Safety Summit on Anthropic’s Responsible Scaling Policy(<)/em(>), Anthropic (Nov. 1, 2023), (<)a href='https://www.anthropic.com/news/uk-ai-safety-summit'(>)https://www.anthropic.com/news/uk-ai-safety-summit(<)/a(>) (last visited Aug. 5, 2025); (<)em(>)The Case for Targeted Regulation(<)/em(>), Anthropic (Oct. 31, 2024), (<)a href='https://www.anthropic.com/news/the-case-for-targeted-regulation'(>)https://www.anthropic.com/news/the-case-for-targeted-regulation(<)/a(>) (last visited July 25, 2025) (“RSPs are not intended as a substitute for regulation, but as a prototype for it”); (<)em(>)Third-Party Testing as a Key Ingredient of AI Policy(<)/em(>), Anthropic (Mar. 25, 2024), (<)a href='https://www.anthropic.com/news/third-party-testing'(>)https://www.anthropic.com/news/third-party-testing(<)/a(>) (last visited Aug. 5, 2025) (We expect that ultimately some form of third-party testing will be a legal requirement for widely deploying AI models, but designing this regime and figuring out exactly what standards AI systems should be assessed against is something we’ll need to iterate on in the coming years – it’s not obvious what would be appropriate or effective today, and the way to learn that is to prototype such a regime and generate evidence about it.”). and notes that transparency should be an interim step ahead of comprehensive evaluation methods.168(<)em(>)The Need for Transparency in Frontier AI(<)/em(>), (<)em(>)supra(<)/em(>) note 165. Overall, Anthropic displays a mixed record on regulatory policy. While it engages constructively with regulators, it still lobbies for narrower laws, as its successful push to reduce pre-harm liability in the California SB-1047 bill attests.
4. Business Structures Linked to Variations on Positions on AI Regulation
Frontier labs routinely demand regulation.169Smith, (<)em(>)supra(<)/em(>) note 141; (<)em(>)The Case for Targeted Regulation(<)/em(>), Anthropic, (<)em(>)The Case for Targeted Regulation(<)/em(>), (<)em(>)supra(<)/em(>) note 167; Fried, (<)em(>)supra(<)/em(>) note 131 (stating that “[r]egulation needs to be part of society’s response” to AI); Gerrit De Vynck & Nitasha Tiku, (<)em(>)AI Execs Used to Beg for Regulation. Not Anymore.(<)/em(>), Wash. Post (May 8, 2025), (<)a href='https://www.washingtonpost.com/technology/2025/05/08/altman-congress-openai-regulation/'(>)https://www.washingtonpost.com/technology/2025/05/08/altman-congress-openai-regulation/(<)/a(>). Predictable rules would foster certainty, increase trust and security for customers, and even the playing field for those with more stringent self-governance practices. However, these same companies often oppose concrete regulatory proposals as too costly, difficult to implement, or setting back the West in the geopolitical race against China. The policy statements above provide a sense of the daylight between the high-level claims of frontier AI labs wanting regulation and their actual positions towards concrete proposals to do so. This sample is not the full story, as it neither conveys the nuances within each company’s position nor, for that matter, the evolving strategies as the political appetite for regulation ebbs and flows. Also left out is the extent of corporate lobbying behind closed doors. In 2023 alone, prominent technology firms spent $27 million and dedicated 39 full-time employees to lobby EU lawmakers. The investment paid off as these companies secured 78% of closed-door meetings with lawmakers.170Lévesque, (<)em(>)supra(<)/em(>) note 120. Conversely, pushing back on some proposals is not to be confused with blanket obstruction to regulation writ large, especially early in the regulatory process when obligations can legitimately be unclear or unrealistic.
That said, the sample above suggests that companies tend to resist laws that could affect their bottom line. This is inconsistent with lofty missions purporting to shun private profit and prioritize the long-term benefit of humanity. The EU AIA might be the foremost horizontal piece of AI regulation, but its final form is significantly subdued in no small part thanks to industry pressure. AI companies also managed to kill California’s SB-1047 safety bill altogether, and generally rallied around a ban on state regulation—ostensibly to avoid an impracticable patchwork of conflicting regulation, but given Congress’s poor track record in passing regulation, one is left wondering about the convenient outcome: avoiding regulation altogether. Without litigating the appropriateness of these instances of AI regulation, the trend of opposing meaningful constraints signals a reluctance to absorb compliance costs or to slow down the race to market ever more capable models.171For a similar argument regarding industry amendments to neuter state data and AI regulation, (<)em(>)see(<)/em(>) Matthew U. Scherer, (<)em(>)Legislative Poison Pills, Information Monopolies, and the False Dawn of State Tech Regulation(<)/em(>) (July 10, 2025) (unpublished manuscript), (<)a href='https://papers.ssrn.com/abstract=5346390'(>)https://papers.ssrn.com/abstract=5346390(<)/a(>). As the center of gravity of many such companies shifts to investors,172(<)em(>)See(<)/em(>) pp. 19-20 discussing superstakeholders in the context of Microsoft’s 13 billion investment in OpenAI. this pushback against regulatory guardrails raises concerns as to AI companies’ ability to prioritize mission over profit. Moreover, the push for federal preemption among some of these companies can be construed as a convenient way to defang more restrictive state laws, especially if the federal government seeks to enforce a weak federal standard as the ceiling for AI regulation. This seems to be the Trump administration’s approach. Indeed, the federal AI Litigation Task Force challenges state laws inconsistent with a deregulatory policy.173Memorandum on Artificial Intelligence Litigation Task Force, U.S. Att’y Gen., Memorandum for All Department of Justice Employees (Jan. 9, 2026), (<)a href='https://www.justice.gov/ag/media/1422986'(>)https://www.justice.gov/ag/media/1422986(<)/a(>) [(<)a href='https://perma.cc/5TZ3-WSHB'(>)https://perma.cc/5TZ3-WSHB(<)/a(>)].
Although all the companies under review have pushed back against AI regulation, they have done so to different degrees and with different implications when it comes to upholding their missions. Variations among them suggest that their distinct business structures have some bearing on their ability to align mission and action when it comes to AI regulation. Google’s opposition to regulation is difficult to square with DeepMind’s mission to “Build AI responsibly to benefit humanity.” The company opposed European and Californian laws and is seeking to neutralize local state initiatives in the U.S. As a typical for-profit corporation bearing no particular duty to the public interest other than malleable principles to forego certain problematic applications, Google clears the low bar of prioritizing shareholders’ interests. Seen in this light, Google DeepMind’s largely ceremonial mission to “[b]uild AI responsibly to benefit humanity” yields to a self-interested strategy avoiding regulatory hurdles slowing down commercialization.
OpenAI’s actions are even more inconsistent with its mission. The capped-profit structure strives to responsibly build AI to benefit humanity, yet it systematically opposed guardrails that would meaningfully constrain its activities. Though it demanded regulation in the abstract, it mobilized against concrete safety proposals in the EU AIA; opposed the California AI Safety bill; and actively opposed subnational, state-level AI legislation in the U.S. This dissonance between high-level statements and actual regulatory policy suggests that its constantly evolving corporate structure is a brittle basis to promote the stated mission.
Anthropic fared better than the other companies under review.174The long-term benefit trust was added in September 2023, prior to all the regulatory positions under study, save for the comments on the EU AIA. The public benefit corporation overseen by a long-term benefit trust took a firm stance against the federal ban on state AI regulation and it did not actively oppose EU AI regulation during the bill’s drafting process.175With the caveat that the company was founded only a few months before the consultation period for the EU AIA. (<)em(>)See(<)/em(>) notes 155-56 and accompanying text for details. That said, it has a mixed record on the California AI Safety bill. Though it proposed constructive amendments and ultimately cautiously endorsed the bill, it played an important role in watering down the initial draft.176Samuel, (<)em(>)supra(<)/em(>) note 161. Similarly, the company’s track record of developing regulatory prototypes is a double-edged sword. On the one hand, it stands out as more constructive than blanket opposition to regulation.177(<)em(>)See(<)/em(>) (<)em(>)infra(<)/em(>) note 180. For example, it framed its voluntary model constraints as a practical blueprint for future binding regulation. Yet at the same time, such industry-led initiatives can entrench weak regulation that let companies set the terms of the regulatory conversation and shut down the possibility of stricter requirements.178Kate Brennan, Amba Kak & Sarah M. West, (<)em(>)The Storm Clouds Looming Past the State Moratorium: Weak Regulation Is as Bad as None(<)/em(>), Tech Policy Press (June 10, 2025), (<)a href='https://techpolicy.press/the-storm-clouds-looming-past-the-state-moratorium-weak-regulation-is-as-bad-as-none'(>)https://techpolicy.press/the-storm-clouds-looming-past-the-state-moratorium-weak-regulation-is-as-bad-as-none(<)/a(>). Keeping these caveats in mind, Anthropic’s business structure nevertheless seems more conducive to walking the talk on the public interest mission when it comes to regulatory policy. Future research should incorporate subsequent signals emanating from frontier labs, including proposals for FINRA-style public/private oversight and mandatory third-party testing.179For examples of such signals from Anthropic, see Demis Hassabis, (<)em(>)A Framework for Frontier AI and the Dawning of a New Age(<)/em(>), Substack (July 14, 2026) and (<)a href='https://demishassabis.substack.com/p/a-framework-for-frontier-ai-and-the-dawning-of-a-new-age'(>)https://demishassabis.substack.com/p/a-framework-for-frontier-ai-and-the-dawning-of-a-new-age(<)/a(>); Dario Amodei, (<)em(>)Policy on the AI Exponential(<)/em(>) (June 2026), (<)a href='https://darioamodei.com/post/policy-on-the-ai-exponential'(>)https://darioamodei.com/post/policy-on-the-ai-exponential(<)/a(>).
Corporate structures result in different stances on AI regulation, which suggests some bearing on the ability to align regulatory policy with the public benefit mission. But ultimately, no company under review is really on board with regulation that meaningfully constrains their activities: all have opposed or diluted regulatory proposals. This is not surprising given that even for the most robust structure under review, the incentives remain fundamentally misaligned with pursuing the public interest. For all its innovations in corporate structure, Anthropic stops short of securing the public mission above profits. The company can, but is not required to do so. Overall, the regulatory policies of AI companies suggest there is only so much corporate structure can do to get them to put the public interest above profits.