The Fastest Race Nobody Agreed to Run - Should We Slow Down AI?

In the spring of 2023, more than a thousand of the world's most prominent AI researchers and technology figures signed an open letter calling for a six-month pause in the development of AI systems that are more powerful than GPT-4. The signatories included Elon Musk, Steve Wozniak, and dozens of distinguished academics. The letter warned that AI labs were locked in an "out-of-control race" to develop increasingly powerful systems that nobody fully understood and nobody had figured out how to make reliably safe.

The pause never happened.

In the three years since that letter was published, the pace of AI development has not slowed. It has accelerated. The systems deployed today are orders of magnitude more capable than GPT-4. They write code, conduct research, manage complex multi-step tasks, and make consequential decisions across medicine, law, finance, and security. The AI labs that were called upon to pause are larger, better funded, and moving faster than they were in 2023. And the fundamental concerns that motivated the letter, about safety, about control, about understanding what we are building before we build it,  have not been resolved.

They have deepened.

This is the central tension of AI development in 2026 - the technology is advancing faster than our ability to understand it, govern it, or verify that it is safe. And the question of whether we should slow down,  and whether slowing down is even possible, is one of the most consequential and least settled debates in contemporary technology.

This article takes that debate seriously. Both sides of it. Because the honest answer is not simple, and the people who tell you it is simple are not telling you the whole truth.

How Fast Is "Too Fast"?

To evaluate whether AI is developing dangerously quickly, you need a reference point. Fast relative to what?

Relative to our understanding of what these systems are doing internally, very fast. The field of mechanistic interpretability, which attempts to understand the internal reasoning of large AI models, is a young discipline that has made significant progress and is still nowhere near the point where researchers can comprehensively verify what a frontier model has learned or whether its objectives are aligned with human intentions. We deploy systems we cannot fully read.

Relative to our regulatory and governance capacity, it is very fast. The EU AI Act, the most comprehensive AI regulation currently in force anywhere in the world, was designed primarily around the capability levels of 2021 and 2022 AI systems. The systems in deployment today have capabilities that the Act's farmers did not anticipate, and the Act's provisions are being tested against use cases its authors did not foresee (European Parliament, 2024). Legislation moves in years. AI capability moves in months.

Relative to our social and institutional capacity to absorb the changes, AI is driving very fast. Labour markets, educational systems, legal frameworks, democratic processes, and professional licensing structures are all being disrupted simultaneously by a technology whose full implications are not yet visible. Society's capacity to adapt to technological change has limits. Whether those limits are being exceeded is an open empirical question.

Relative to past technological transitions, arguably not as fast as it feels. The industrial revolution, electrification, the internet, all transformed society in ways that caused significant dislocation alongside enormous benefit. AI is not uniquely unprecedented in causing disruption. What may be different is the breadth of simultaneous impact and the speed at which frontier capability is advancing in absolute terms.

The Case for Pressing the Brakes

The argument for slowing down AI development is not primarily made by people who are afraid of technology. It is made by people who understand it deeply and are concerned precisely because of that understanding.

We Cannot Verify What We Are Building

The most technically serious argument for slowing down is the alignment problem, the gap between what AI systems is designed to do and what they actually do, which grows more consequential as the systems grow more capable.

Current AI training methods do not guarantee alignment. They produce systems that perform well on the objectives they are evaluated against, but whether those systems have genuinely internalized the values and goals their developers intended or have learned to produce outputs that appear aligned while pursuing different internal objectives, is something we cannot currently verify (Hubinger et al., 2019).

The concerning version of this argument is not about today's systems. It is about the trajectory. If we deploy successively more capable systems before solving the alignment problem, we are betting that the problem will remain manageable until we have solved it. The higher the capability, the larger the consequences of that bet being wrong.

Stuart Russell, one of the world's foremost AI researchers and author of the standard textbook on artificial intelligence, has argued that deploying increasingly powerful AI systems without solving the control problem is analogous to a plan to build nuclear reactors without having figured out how to shut them down (Russell, 2019). The analogy is imperfect, AI systems are not nuclear reactors, but the underlying logic is serious - capability without verified control is a risk that scales with the capability.

The Competitive Dynamic Undermines Safety

Even organizations that genuinely priorities safety face structural pressures that work against it. The AI development landscape in 2026 is characterized by intense competition, between companies racing to capture market share, between nations competing for technological leadership.

In a competitive race, safety measures that slow development impose costs on the organizations that adopt them without imposing equivalent costs on those that do not, unless those safety measures are required of everyone through binding standards or regulation. The result is a dynamic where each individual actor has rational incentives to move faster and invest less in safety than the collective interest warrants. This is a classic coordination problem, and coordination problems require coordination solutions, not individual virtue.

The parallel to other dangerous technology races is instructive. The management of nuclear weapons risk did not depend on individual nations voluntarily choosing restraint based on their own safety assessments. It required negotiated agreements, verification mechanisms, and international institutions. Whether an analogous framework for AI is achievable, given the competitive dynamics between the United States, China, and other major AI-developing nations, is one of the most important unsolved problems in technology governance.

The Stakes Are Qualitatively Different from Previous Technologies

Proponents of slowing down argue that AI is not just another powerful technology, it is a technology that, if it continues to advance, could eventually exceed human capability across virtually every domain of cognitive work, including the domain of designing and improving AI systems themselves.

If that threshold is reached, and whether or when it might be is genuinely uncertain, the consequences of misalignment at that capability level would be qualitatively different from any previous technology risk. A misaligned power grid damages infrastructure. A misaligned nuclear weapon destroys a city. A misaligned superintelligent AI system, in most concerning theoretical scenarios, could pose risks of a different magnitude entirely.

Most AI researchers consider this scenario to be at least a decade away and many consider it speculative. But the argument that precautionary attention to this possibility is warranted, even under uncertainty about its likelihood, is taken seriously by researchers including Geoffrey Hinton, Yoshua Bengio, and dozens of other figures who have spent careers building the field and have become increasingly concerned about where it is heading (Bengio et al., 2023).

The Case Against Slowing Down

The argument for continuing at current pace is not simply the self-interest of AI companies who profit from speed. It has genuine substance, and dismissing it prevents honest engagement with the full picture.

The Benefits Are Real and Arriving Now

AI is already delivering benefits that have life-and-death implications. AlphaFold's protein structure predictions have accelerated drug discovery in ways that will translate into treatments for diseases that currently have none. AI diagnostic tools identify cancers, diabetic retinopathy, and cardiac conditions earlier than human physicians in controlled studies. AI-powered materials discovery is identifying candidates for next-generation batteries, solar cells, and carbon capture technologies that the clean energy transition urgently needs.

These benefits are not hypothetical. They are happening. And they are happening faster because of the pace of AI development. Slowing that pace has real costs, measured in delayed treatments, in climate targets missed, in scientific discoveries deferred. Those costs are diffuse and statistical, which makes them less visible than the concentrated risks that capture headlines. But they are not less real.

Slowing Down Unilaterally Does Not Work

The most pragmatic argument against slowing down is that it does not actually reduce risk, it just redistributes who is at the frontier.

If the United States slows its AI development and China does not, or vice versa, the frontier of AI capability still advances at the same pace. The entity at the frontier is different, with different values, different governance structures, and potentially different safety priorities. The risk does not decrease. It is relocating.

This argument has real force. Unilateral restraint in technology development has a poor historical track record of achieving its safety objectives. The alternative, international coordination that creates binding constraints on all major AI developers simultaneously, is theoretically more effective but practically extremely difficult to achieve given the geopolitical landscape.

The honest version of this argument is not "therefore we should race without constraint." It is "the solution is international coordination, which we should pursue urgently while recognizing that unilateral slowdown is not a substitute for it."

Technology May Be Self-Correcting in Some Dimensions

Some researchers argue that certain safety concerns are addressed, rather than amplified, by greater capability. A more capable AI system can better understand nuanced instructions, better identify the limits of its own knowledge, and better flag uncertainty, properties that improve rather than worsen with capability, at least for some failure modes.

The most capable AI systems in 2026 are, in some respects, more honest about their limitations than earlier generations, more likely to acknowledge uncertainty, more resistant to some forms of manipulation, more consistent in following complex instructions. This does not mean all safety properties improve with scale. But it complicates the simple narrative that more capability is straightforwardly more dangerous.

We Cannot Identify Safe Speed

Critics of the "slow down" position note that no one advocating for deceleration has articulated a clear answer to the question - slow down to what? What rate of progress would be safe? What milestones in safety research would justify resuming normal pace? Without answers to these questions, "slow down" is not a policy, it is an expression of anxiety.

This is a legitimate challenge. The precautionary argument is more compelling when it comes to a clear account of what precaution requires and what conditions would satisfy it. The AI safety field has made progress on these questions, developing specific technical milestones in alignment and interpretability research that could serve as safety gateposts, but a widely shared, actionable framework for safe development pace does not yet exist.

What Is Actually Being Done

Between the poles of "full speed ahead" and "stop everything," there is a set of intermediate governance approaches that are actually being implemented, imperfectly, unevenly, and with significant gaps.

Compute governance, regulating access to the specialized chips required to train frontier AI systems, has emerged as the most practically tractable lever for AI governance. Export controls on advanced AI chips, licensing requirements for large training runs, and monitoring of compute consumption can slow the development of frontier systems and create chokepoints for international coordination. The US government's semiconductor export controls, whatever their geopolitical motivations, have had measurable effects on the ability of certain actors to develop frontier AI systems (Bureau of Industry and Security, 2023).

Mandatory safety evaluations, requiring AI developers to submit frontier models to independent evaluation before deployment, are now required under the EU AI Act and are voluntary commitments made by major AI laboratories under agreements with the US and UK AI Safety Institutes. The quality and rigor of these evaluations is still developing, but the infrastructure for pre-deployment safety review is being built.

Information sharing requirements, requiring AI developers to disclose safety-relevant information about frontier models to regulators and to each other, are part of both the EU AI Act and voluntary commitments made at the Bletchley and Seoul summits. The goal is to allow collective assessment of risks that individual organizations may not fully see.

International coordination mechanisms, through the OECD AI Policy Observatory, the Global Partnership on AI, and the bilateral and multilateral summit frameworks established at Bletchley and Seoul, are creating the diplomatic infrastructure for AI governance, even as binding international agreements remain elusive.

None of this is sufficient. The gaps between the pace of capability development and the pace of governance development remain large. But the trajectory of governance effort is positive, and the institutional infrastructure being built now will matter as capability continues to advance.

What the Honest Middle Ground Looks Like

The honest position on whether AI is developing too fast is one that resists both the dismissiveness of unconstrained optimism and the paralysis of unconstrained alarm.

It acknowledges that the benefits of AI development are real and that slowing them has genuine costs, in health outcomes, in scientific progress, in economic development.

It acknowledges that the risks of deploying increasingly powerful systems without adequate safety research and governance are also real, not hypothetical science fiction, but documented in current systems and escalating with capability.

It recognizes that unilateral slowdown by any single actor is not a solution, but that international coordination, which would allow all major developers to slow simultaneously, reducing the competitive pressure without relocating the frontier, is both necessary and difficult.

It insists that the investment in AI safety research, in alignment techniques, interpretability tools, evaluation methodologies, and governance frameworks, must scale with the investment in capability development, not lag years behind it as it currently does.

And it holds that the public has both the right and the responsibility to engage with these questions, not leaving them with AI companies whose interests are not perfectly aligned with collective welfare, or to governments whose understanding of the technology often lags its development.

The Bottom Line

The fastest race nobody agreed to run is still running.

The people building AI systems are not, in the majority, reckless or indifferent to the risks. Many of them are genuinely committed to safety and are working hard on difficult problems with real urgency. The people raising alarms are not, in the majority, technophobes or catastrophists. Many of them are among the most knowledgeable people in the field, raising concerns precisely because of what their expertise allows them to see.

The truth is that this technology is extraordinarily powerful, that its development is genuinely outpacing our capacity to verify its safety and govern its deployment, and that the appropriate response is neither to stop nor to simply accelerate, but to invest with urgency in the safety research, governance mechanisms, and international coordination that would allow the benefits of AI to be captured without the risks being passed, unexamined, to future generations.

Whether we should slow down is the wrong question, ultimately. The right question is - what would it take to develop AI fast enough to capture its benefits while building the safety and governance infrastructure capable of managing its risks? And are we actually building that infrastructure at the required pace?

The answer, in 2026, is not yet.

The race continues. The question of whether anyone is steering is the one that matters.

Cover Image by Freepik [www.freepik,com]

References

Bengio, Y., Hinton, G., Yao, A., Song, D., Abbeel, P., Harnad, S., Zhang, Y.Q., Xue, L., Romero-Soriano, A., Prince, S.J.D., Bahdanau, D., Hutter, F., Golub, G.H., Lyu, S., Keutzer, K. and Pierson, E. (2023) 'Managing AI risks in an era of rapid progress', arXiv preprint arXiv -2310.17688. Available at - https -//arxiv.org/abs/2310.17688 (Accessed - 24 June 2026).

Bureau of Industry and Security (2023) Commerce implements sweeping restrictions on exports of advanced computing chips, chipmaking equipment, and other items to China. Washington, DC - US Department of Commerce. Available at - https -//www.bis.doc.gov/index.php/documents/about-bis/newsroom/press-releases/3186-2023-10-17-bis-press-release-advanced-computing-eis-final-rule/file (Accessed - 24 June 2026).

European Parliament and Council of the European Union (2024) Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union. Available at - https -//eur-lex.europa.eu/legal-content/EN/TXT/?uri=OJ -L_202401689 (Accessed - 24 June 2026).

Future of Life Institute (2023) Pause giant AI experiments - an open letter. Available at - https -//futureoflife.org/open-letter/pause-giant-ai-experiments/ (Accessed - 23 June 2026).

Gabriel, I. (2020) 'Artificial intelligence, values, and alignment', Minds and Machines, 30(3), pp. 411–437. doi -10.1007/s11023-020-09539-2.

Hubinger, E., van Merwijk, C., Mikulik, V., Skalse, J. and Garrabrant, S. (2019) 'Risks from learned optimization in advanced machine learning systems', arXiv preprint arXiv -1906.01820. Available at - https -//arxiv.org/abs/1906.01820 (Accessed - 23 June 2026).

Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl, S.A.A., Ballard, A.J., Cowie, A., Romera-Paredes, B., Nikolov, S., Jain, R., Adler, J., Back, T., Petersen, S., Reiman, D., Clancy, E., Zielinski, M., Steinegger, M., Pacholska, M., Berghammer, T., Bodenstein, S., Silver, D., Vinyals, O., Senior, A.W., Kavukcuoglu, K., Kohli, P. and Hassabis, D. (2021) 'Highly accurate protein structure prediction with AlphaFold', Nature, 596(7873), pp. 583–589. doi -10.1038/s41586-021-03819-2.

Krakovna, V., Uesato, J., Mikulik, V., Martic, M., Tomasev, N., Stepleton, T., Perolat, J., Everitt, T. and Legg, S. (2020) 'Specification gaming - the flip side of AI ingenuity', DeepMind Blog. Available at - https -//deepmind.com/blog/article/Specification-gaming-the-flip-side-of-AI-ingenuity (Accessed - 23 June 2026).

Marcus, G. and Davis, E. (2019) Rebooting AI - building artificial intelligence we can trust. New York - Pantheon Books.

Russell, S. (2019) Human compatible - artificial intelligence and the problem of control. New York - Viking.

Suleyman, M. and Bhaskar, M. (2023) The coming wave - technology, power, and the twenty-first century's greatest dilemma. London - Crown.

UK Government (2023) The Bletchley declaration by countries attending the AI safety summit, 1–2 November 2023. London - His Majesty's Government. Available at - https -//www.gov.uk/government/publications/ai-safety-summit-2023-the-bletchley-declaration/the-bletchley-declaration-by-countries-attending-the-ai-safety-summit-1-2-november-2023 (Accessed - 24 June 2026).