This time, the warning is coming from inside the engine room. Last week, Jacob Coxon, an AI researcher who had spent the previous three years at two of the field’s most important companies, OpenAI and Anthropic, announced that he was leaving his job. His explanation had nothing to do with burnout or a better offer elsewhere. Coxon said the leading companies in the field were racing toward a superintelligence capable of improving itself and operating autonomously - while “gambling with our lives.”

The warning gained added weight just days later, when Anthropic CEO Dario Amodei published an essay titled “We Must Pace the Frontier”. His language was more measured, but the message was the same: AI systems may improve faster than safety and control mechanisms can keep up.

Amodei proposed, among other things, giving independent external evaluators meaningful access to companies’ development processes, establishing a coordination mechanism among developers of advanced models, and creating international arrangements. Sam Altman of OpenAI added his voice, expressing support for slowing the race and embracing at least some proposals for external oversight.

The result is a rather unusual critical mass: the people leading the companies at the forefront of one of the greatest technological revolutions of our time, perhaps the greatest of them all, are themselves acknowledging that existing oversight mechanisms are inadequate. They are doing so while their companies continue competing to develop the next generation of the very same technology.

A similar dynamic has played out for years in another field. We know we are heading toward the abyss; every warning light is flashing, yet we march on relentlessly.

Throughout the twentieth century, evidence mounted of the link between accelerating industrialization, the burning of fossil fuels, greenhouse-gas emissions, and severe, destructive climate change. The scientific consensus took shape, models grew more sophisticated, and the damage became increasingly tangible in the real world, not just in forecasts. Yet the gap between knowledge and action did not narrow - it widened. The world did not stop using coal, oil, and gas once their environmental cost became clear. It simply increased the pace of extraction.

The parallel between fossil fuels and AI lies in the gap between humanity’s ability to develop powerful technologies at speed and its political, legal, and moral capacity to build, in real time and with an eye to the future, mechanisms that prevent that power from becoming destructive.

In both cases, the technology offers enormous benefits that are immediately apparent, while efforts to limit it are initially cast as a battle against progress or as a primitive fear of the unknown. In both cases, there is a huge economic incentive to press ahead and leave the question of the cost for later. The environmental record shows the danger of postponing that reckoning.

One reason is what economists call the externalization of costs. A polluting factory reaps the revenue and profits generated by production, while the cost of pollution is borne by local residents, the health care system, the wider public, and future generations. Polluters therefore have no incentive to restrain themselves.

Much environmental regulation has grown out of the understanding that the market does not, on its own, price the full harm caused by economic activity. A product’s price therefore fails to reflect its true cost to society. That insight shaped environmental tort law, reporting obligations, and even corporate law, including concepts of corporate, social, and environmental responsibility.

AI raises the same question, especially about its side effects. A company that develops a new model benefits from the investment it attracts, the revenue and valuation it generates, the market advantage it confers, and the product’s own usefulness and capabilities.

A significant share of the costs and consequences, however, falls on others: workers whose occupations are transformed or disappear; creators whose works have been used to train systems; citizens exposed to scams and false information; public institutions forced to contend with forgeries and manipulation; legal systems required to resolve questions where technology has outpaced the law; and the public as a whole, which is gradually becoming dependent on infrastructure controlled by a very small number of companies.

The same is true of the risks and the possibility of losing control over a model. We need not accept the most extreme predictions about a superintelligence getting out of control to recognize the immense gap that already exists between those who reap the profits and those who bear the risks.

Another problem familiar from the struggle against climate change is the collective action problem. A state may conclude that it must reduce emissions, yet fear that other countries will continue producing goods using cheap and polluting energy and thereby gain a competitive advantage. An AI company may conclude that a particular pace of development is unsafe, yet know that stopping unilaterally would allow a competitor to overtake it. The US government may become convinced that restrictions should be imposed on American AI companies, yet fear that China will not follow suit.

At this point, it is important to heed Amodei’s point: the problem is not merely technological, but institutional. It cannot be solved by companies promising to be more careful, just as the climate crisis cannot be solved by voluntary commitments from factories to reduce their emissions. It requires rules, transparency, oversight, the ability to verify compliance, and coordination among the relevant actors – above all, a shared understanding and collective will.

The question of AI is thus gradually shifting from one of engineering and science to one of governance, public policy, and international relations. Who sets the boundaries of acceptable risk? Who oversees those developing these systems? Who has access to the information needed to assess that risk? And who has the right to decide, on behalf of the public, that a certain level of risk is an acceptable price of progress?

The experience accumulated in confronting the climate crisis also offers a useful principle. Regulation need not begin only after harm has been proved beyond all doubt. The precautionary principle developed in environmental law from the recognition that, when potential harm is serious and sometimes irreversible, waiting for complete certainty is not a neutral position. Choosing not to act is itself a decision about the distribution of risk, and it usually shifts that risk onto the public rather than those who create it.

The lesson of the climate crisis does not end with regulations set individually by each country. Climate change has made clear both the limits of national action in the face of a risk that knows no borders and the need for international treaties and mechanisms that establish common rules even among competing states with conflicting interests.

Early signs of a similar approach are already visible in AI. The Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, which Israel has also signed, is the first attempt to create a binding international legal framework in the field.

So long as a company or country that chooses to slow down knows that its competitors can continue racing ahead, domestic regulation alone cannot solve the problem. Just as tackling climate change required efforts to agree on measurement, reporting, targets, and common rules of conduct, AI will require international agreements on transparency, risk assessment, independent testing, and perhaps even limits that must not be crossed in developing certain capabilities.

The model need not be identical to climate agreements, and it is doubtful whether any single treaty could keep pace with a technology that is changing so rapidly. The underlying principle, however, is similar: when the risk is global and the incentive to break the rules is driven by competition, no single country or company can guarantee the safety of the system as a whole.

Nor is it necessary to believe that AI is about to destroy humanity to conclude that those developing systems of such extraordinary power must be subject to external oversight. We do not need to know for certain that a future system will get out of control to demand the ability to examine how it operates and what safeguards have been built into it.

Another lesson of the climate crisis concerns dependence. At a certain point, the question is no longer whether it is technically possible to stop using a technology, but what the social and economic cost of trying to do so will be. It is difficult to transform an entire economy built over decades around a particular energy source. The same will be true of AI. Change will become extremely difficult once government systems, armies, hospitals, schools, universities, media organizations, law firms, and labor markets have reorganized their operations around the new technology. That dependence could become almost irreversible.

Perhaps there are grounds for cautious optimism. By the time most of humanity understood the full scale of the climate crisis, much of the infrastructure responsible for it had already been in place for generations. With AI, we are still relatively close to the beginning. The machine is being built before our eyes, while internal debates among decision-makers are beginning to come into public view.

The major fossil-fuel companies knew very well that they were causing unprecedented damage that threatened the future of human life on Earth, yet did everything in their power to hide it from the public. With AI, something almost the opposite is sometimes happening. Executives and researchers at the companies developing it are publicly warning of serious risks and calling for oversight, controls, and even a slower pace.

The opportunity to shape the rules of the game before dependence becomes entrenched still exists, even if the window is steadily narrowing.

This article was published in Hebrew on September 14, 2026