OpenAI calls for stricter California AI law

OpenAI repositions itself in California’s AI debate

OpenAI is now calling on California to strengthen Senate Bill 53, which is devoted to the safety and transparency of the most advanced artificial intelligence systems. The news, reported by TechCrunch in its article “OpenAI says California should strengthen its AI safety bill”, marks a significant shift in the political positioning of the company behind ChatGPT.

The central point is not simply that OpenAI accepts the idea of a regulatory framework. According to TechCrunch, the company believes the California measure should go further on certain obligations applying to developers of frontier models. This stance comes as AI companies, elected officials, and U.S. federal authorities have clashed for several years over a fundamental question: who should set safety rules for models capable of producing text, code, images, audio, or carrying out increasingly complex tasks?

California occupies a particular place in this confrontation. The state is home to a large share of the sector’s most influential labs, including OpenAI, Google, Anthropic, Meta, and numerous specialized startups. Choices made in Sacramento can therefore have an impact far beyond California’s borders. A publication, documentation, or reporting requirement imposed on developers operating in the state can quickly become a de facto benchmark for the entire U.S. market, or even internationally.

SB 53 fits precisely within this logic. The measure targets developers of so-called frontier AI systems—that is, the most advanced models, whose capabilities and potential uses raise specific questions regarding safety, cybersecurity, risk control, and transparency. The debate therefore does not concern all software using AI, nor every application incorporating a conversational assistant. It primarily concerns the players that train and deploy the most powerful general-purpose models.

In TechCrunch’s account, OpenAI’s support for strengthening SB 53 represents a notable political reversal. The company had previously expressed opposition to a more restrictive California measure on AI safety. This precedent matters: it is a reminder that OpenAI long defended the view that overly prescriptive regulation decided at the level of a single state could fragment the U.S. legal framework and complicate innovation.

The change in tone does not necessarily mean that OpenAI is abandoning this concern. Rather, it shows that the discussion has shifted. The question is no longer only whether frontier models should be regulated, but rather what form of regulation is appropriate, what information must be disclosed, what safety procedures must be formalized, and which authority should be responsible for monitoring their implementation.

This nuance is decisive. For major labs, a legal framework can represent a cost: producing technical documents, internal procedures, assessments, notification obligations, litigation risks, and greater exposure to public scrutiny. But it can also create a more predictable environment. A company that already has legal teams, safety officers, public policy specialists, and substantial resources can absorb these obligations more easily than a smaller competitor.

The growing role of private actors in drafting the rules is therefore at the heart of the matter. OpenAI is not merely reacting to an already-written law: the company is seeking to influence the contours of the future regime applying to the most powerful models. This is precisely the arena of the current battle for influence around AI: labs are calling for a common framework while seeking to avoid having that framework defined without them or against their development methods.

SB 53: safety, transparency, and obligations for frontier models

SB 53 is part of a series of California attempts to address the risks associated with advanced AI systems. It rests on an apparently simple idea: companies that design frontier models should spell out the safety measures they put in place and show greater transparency when significant incidents occur.

This principle differs from an approach that would ban a technology or subject every AI-based product to prior administrative approval. SB 53 instead seeks to establish process obligations. In other words, the legislature is interested in how developers organize their safety policies, assess certain risks, and account for their practices.

Transparency is one of the pillars of the measure. In the field of frontier models, information available to the public often remains limited. Companies sometimes publish technical reports, safety notes, or fact sheets describing certain limitations of their models. But these documents are generally produced according to each lab’s own standards, with widely varying levels of detail. Assessment methodologies, complete test results, incidents encountered, or internal trade-offs are not always made public.

A legislative mechanism may seek to reduce this heterogeneity. It may require developers to formalize their commitments, publish them, or submit them to the relevant authorities. It may also govern the reporting of safety incidents. The issue is not merely to produce more documents: it is to make practices comparable, verifiable, and, to some extent, enforceable.

For supporters of such an approach, this type of mechanism addresses a challenge specific to the most advanced systems. Users, client companies, independent researchers, and public authorities do not necessarily have the information needed to assess a model’s actual limitations. The company training it knows its data, its internal procedures, its capabilities observed during testing, and its deployment choices. According to advocates of regulation, this information asymmetry justifies documentation and disclosure obligations.

For opponents of overly rigid rules, the risk is different. They believe a poorly calibrated obligation can become bureaucratic, expose sensitive information, or create legal uncertainty. Models evolve quickly, testing techniques are constantly changing, and the risk categories themselves are subject to debate. A law drafted with too much precision could be overtaken by the technologies it seeks to regulate; conversely, a law that is too vague could leave considerable room for interpretation to companies and regulators alike.

This is the space in which OpenAI’s intervention falls. According to TechCrunch, the company is not calling for SB 53 to be abandoned: on the contrary, it believes California should strengthen it. The political signal is significant because it validates the idea that a state measure can play a role in governing frontier models, despite the criticisms regularly directed at the local level.

The exact scope of such support will nevertheless depend on the amendments being advocated, their wording, and how they are received by California lawmakers. A company can support the principle of transparency while challenging the level of detail required. It can approve the existence of safety reports while preferring certain elements to be submitted confidentially. It can accept a reporting obligation while seeking to define precisely what constitutes an incident serious enough to trigger that procedure.

These distinctions are not merely technical debates. They determine the public authorities’ effective ability to monitor risks, the amount of information available to researchers and customers, and the administrative burden imposed on labs. They can also create an industry standard. When a major player adopts a procedure to comply with California requirements, its partners, suppliers, and competitors may be led to align with it, even in the absence of an identical obligation in other states.

SB 53 is therefore less an isolated measure than a regulatory laboratory. It tests the possibility of subjecting frontier model developers to specific requirements without treating AI as a uniform whole. This distinction between ordinary applications and the most advanced models is now central to most international discussions on AI governance.

A change of course after clashes over California regulation

To understand why OpenAI’s position is attracting attention, it must be placed in the recent history of California regulation. The state has become one of the main arenas of confrontation between the desire to impose guardrails on advanced systems and the concerns expressed by part of the technology industry.

The most visible precedent is SB 1047, a California bill devoted to the safety of high-powered AI models. The measure sparked intense debate among companies, researchers, investors, and civil society organizations. Its sponsor, Senator Scott Wiener, argued for the need to introduce safety obligations suited to systems likely to pose significant risks.

SB 1047 was adopted by the California Legislature before ultimately being rejected by Governor Gavin Newsom. This veto became a landmark moment in the U.S. debate. It showed that ambitious legislation on the safety of advanced models could secure substantial legislative support while still failing in the face of concerns about its potential impact, implementation, or suitability for a rapidly evolving technology.

OpenAI was among the companies that expressed reservations about this type of framework. The discussion notably revealed many players’ preference for a federal approach rather than a proliferation of state laws. Their argument is well known: a company that develops and distributes digital services nationwide could have to navigate different requirements depending on the jurisdiction, increasing compliance burdens and reducing the system’s coherence.

But the federal argument has a practical limitation: at this stage, the United States does not have a unified and binding equivalent to European AI regulation. Washington has encouraged voluntary commitments and risk-management initiatives, but the federal framework remains fragmented. Federal agencies can act within their areas of competence, while Congress regularly debates measures whose adoption remains uncertain.

OpenAI’s position on SB 53 must be read in light of this reality. When the federal framework does not produce a single rule, states retain the capacity to take initiative. California, due to its economic weight and concentration of sector companies, is particularly well placed to exercise this influence. Opposition to any California intervention then becomes harder to maintain if federal rules remain largely voluntary or incomplete.

The reversal reported by TechCrunch should therefore not be reduced to a simple conversion by OpenAI to regulation. It may also reflect a strategy of participation. Once a measure has a chance of structuring the market, companies have an interest in making their expertise, operational constraints, and vision of effective obligations heard. The debate is less about the existence of standards than about their concrete content.

This dynamic is common in heavily regulated industries. Established players may fear new legislation, then seek to contribute to drafting it when it becomes clear that it is likely to succeed. In AI, this sequence is particularly rapid because the technology is advancing quickly and because frontier models are developed by a limited number of organizations with technical knowledge that is difficult for public authorities to replicate.

Nevertheless, the announced support should not be interpreted as a guarantee of total alignment between OpenAI and California elected officials. Calling to “strengthen” a measure does not automatically mean seeking more constraints on every issue. A company may believe that a law should be clearer, more harmonized, more targeted, or more consistent with its own procedures. The legal force of an obligation, its scope, the information covered, and the oversight arrangements remain points for negotiation.

This ambivalence explains the attention paid to the matter. Major AI companies are both the subjects of regulation and major participants in its development. They are best placed to describe their methods, but they also have a direct economic interest in ensuring that rules do not hinder their activity or expose certain strategic information. The role of elected officials and oversight authorities is precisely to arbitrate between this indispensable expertise and the risk of seeing regulation largely written by the companies it is meant to govern.

Competition between governance models in the United States and Europe

The California case highlights a profound difference between U.S. and European regulatory trajectories. In the United States, the debate remains marked by the tension between state initiatives, federal ambitions, and voluntary corporate commitments. In Europe, by contrast, the AI Act has created a common framework applicable across the European Union, with graduated obligations according to risk categories and specific provisions for general-purpose AI models.

This comparison does not mean that SB 53 is a Californian replica of the AI Act. The two measures have neither the same scope, nor the same legal architecture, nor the same institutional place. The European Union legislates for a market of several member states. California acts within the U.S. federal framework. But they address a similar question: how can transparency and safety obligations be imposed on systems that can be integrated into a multitude of products and deployed in very different countries?

For French and European companies, this partial convergence is far from abstract. Many organizations already use models designed in the United States or depend on cloud services, APIs, and tools offered by major labs. Compliance choices made by OpenAI and its competitors in the United States may influence the documentation available to European customers, assessment practices, contractual terms, and the way providers describe the limitations of their systems.

In the case of the AI Act, obligations related to general-purpose AI models are based in particular on documentation and transparency requirements. The European regulation also provides for enhanced treatment of models presenting systemic risks. The logic is therefore comparable on one essential point: regulation cannot focus solely on end applications. It must also examine the players that train general models upstream, because their technical choices have consequences for the entire value chain.

With SB 53, California is seeking to intervene in this same strategic space: that of the lab that develops the model before it is made available to software publishers, client companies, and users. This is where safeguards, testing methods, publication policies, and incident-response arrangements are decided. Once a model has been released at a very large scale, it becomes much more difficult to reconstruct those decisions or correct certain structural flaws.

OpenAI’s competitors are directly affected by this development. Google, Anthropic, Meta, and other major developers are closely following regulatory debates, because any obligation applicable to frontier systems could affect their go-to-market methods. The issue goes beyond commercial rivalry between conversational assistants. It also concerns compliance costs, the speed at which new generations of models are deployed, and companies’ ability to demonstrate that they have taken security risks into account.

California discussions are also reviving a broader debate over whether obligations should be based on model capabilities, the resources used to train them, their actual uses, or harms observed after deployment. Each approach produces different effects. A capability-based rule seeks to anticipate risks before use. A harm-based rule prioritizes response after an incident. A rule based on the size or resources of the developer more easily targets major labs, but may leave aside less costly systems or systems distributed differently.

The choice of transparency, which lies at the heart of SB 53 as presented by TechCrunch, is an attempt to address this complexity without claiming to solve every problem through a general ban. Publishing safety frameworks and reporting certain incidents can enable authorities to accumulate information, identify trends, and eventually adapt the rules. This method nevertheless assumes that disclosures are sufficiently precise, legal concepts are clear, and authorities have the means to use them.

For the French-speaking market, the issue is also competitive. European companies want to avoid excessive dependence on foreign platforms while using the best tools available. A proliferation of national or regional standards can complicate their relationships with providers. But the absence of coherent standards can also increase uncertainty: a company that integrates a model into a critical product wants to know what assessments have been carried out, what limitations have been identified, and what procedures exist in the event of a problem.

OpenAI’s support for a strengthened SB 53 could thus be seen in Europe as a sign that transparency and risk management are becoming lasting elements of competition. A model’s raw capabilities are no longer enough to convince institutional clients and large companies. The quality of documentation, the robustness of safety processes, and the ability to engage with regulators are becoming increasingly important.

A battle for influence set to structure the advanced-model market

In the short term, OpenAI’s stance may weigh on the discussion surrounding SB 53. The company has considerable visibility and direct experience in developing frontier models. Its intervention lends weight to the idea that safety and transparency obligations are no longer only demanded by NGOs, academics, or political leaders: they are also discussed, and even supported under certain conditions, by the companies themselves.

But this influence does not erase the contradictions in the debate. Labs are encouraged to show safety commitments because the trust of the public, customers, partners, and authorities is essential to their growth. They are also encouraged to preserve their technical autonomy, protect sensitive information, and avoid constraints likely to slow their development cycles. The resulting regulatory compromise will necessarily be fragile.

The question of oversight will be decisive. An obligation to publish a safety policy can have real value if it makes it possible to compare commitments, identify gaps, and demand accountability in the event of an incident. It can become essentially declarative if documents remain too general, criteria are undefined, or no authority is able to verify its implementation. The final text of SB 53, its possible amendments, and the conditions of its implementation will therefore matter as much as the very principle of its adoption.

Regulators’ ability to keep pace with a shifting technology will be another test. AI models are regularly updated, and their behavior can vary depending on versions, connected tools, and deployment methods. A law designed around occasional reports will need to be sufficiently adaptable to account for these developments. At the same time, it will need to provide sufficient stability for companies seeking to plan their investments, compliance teams, and internal procedures.

For OpenAI, its current positioning also offers a potential political advantage. By calling for SB 53 to be strengthened, the company can present itself as an interlocutor supportive of safety, without necessarily abandoning its preference for rules defined with industry and coordinated on a broader scale. This stance is particularly useful in a context in which AI is now at the center of discussions on competitiveness, cybersecurity, employment, copyright, and technological sovereignty.

For California elected officials, the challenge will be not to confuse industry support with regulatory consensus. The fact that a major player is calling for a stronger measure may facilitate the adoption of a law. It may also lead to complex negotiations over precise arrangements, with the risk that requirements are calibrated according to the capabilities of the most established companies. Smaller entities, independent researchers, and public-interest advocates do not always have the same resources to take part in these trade-offs.

The California precedent will probably resonate far beyond the state. If SB 53 establishes obligations regarded as realistic by major companies and sufficiently robust by safety advocates, other U.S. jurisdictions may draw inspiration from it. Conversely, if the measure is heavily contested, diluted, or difficult to apply, it will fuel the argument that only a federal framework can effectively address frontier models.

From this perspective, OpenAI’s position reveals less an end to the conflict between innovation and regulation than a new phase in that conflict. Major labs can no longer ignore the demand for rules on the safety of advanced models. They are now seeking to define the relevant risk categories, the indicators that must be submitted, acceptable transparency thresholds, and the institutions responsible for deciding.

For France and Europe, the U.S. development deserves sustained attention. The continent already has a common legal framework with the AI Act, but most models used by European companies are still developed outside Europe. How the United States organizes the transparency and safety of these systems will have indirect effects on European users, supply contracts, and information available to regulators.

The most likely trajectory is not immediate harmonization among Washington, Sacramento, and Brussels. Economic interests, institutions, and legal traditions differ profoundly. But pressure in favor of documented safety mechanisms and increased transparency should continue to intensify. The debate around SB 53 shows that, even within the industry, the question is no longer whether frontier models should be governed. It has become a question of who will write the rules, with what level of stringency, and for the benefit of which players.