Should We Slow Down A.I.? We're Trapped in a Prisoner's Dilemma - September 18, 2026

Everyone Has To Be All-In

Those building the applications are telling us that A.I. may be advancing too quickly. The people competing with them are telling us that slowing down could be even more dangerous. This is why A.I. development today seems like a Prisoner's Dilemma.

Last week there was lots of discussion regarding the likelihood that A.I. could cause catastrophic events for humanity. My last article covered this (read here). 

This discussion was triggered by warnings issued by researchers inside the leading A.I. companies.  These warnings brought calls for a slowdown, or “pacing”  of A.I. research; meaning that, the improvements in power and functionality of A.I. should be matched by the safety features that are built into that same A.I., to avoid potential catastrophic events from occurring.  This was a major call to slow down A.I. development.  But, as expected, there were just as many people who claimed these warnings were overblown, claiming these doomsday statements were unsubstantiated exaggerations of highly unlikely scenarios, and that it was too early to predict that A.I. would cause the end of humanity. 

Who’s Calling For A.I Pacing?

Last week’s fervor was triggered by statements from a former junior researcher at Anthropic, Jacob Coxon.  But then, Dario Amodei the Founder/CEO of Anthropic published an essay last weekend that called for the slowing down of development work on A.I. models, or “the pacing of A.I.”.  He stated that safety features must be in place before A.I. applications are released into the public domain. In his essay, Amodei also called for international standards bodies to be formed, to regulate the development and public release of future A.I. products. 

Shortly after Amodei’s essay was published, Sam Altman, the CEO of Anthropic rival Open AI, issued a public statement agreeing with Amodei's position, that A.I. research should be “paced” or slowed down.  And then shortly after Altman’s statement came out, Elon Musk. whose horse in the A.I. race is in the form xAI, the A.I. operation acquired by Space X, came out with the statement, “Dario Is Right”.  To have these three leading figures in A.I. standing in unity for the  slowing down A.I. development was exceptional.  As would be expected, while the industry respects what these three gentlemen are doing in A.I., the doubters were out in full force.  

The Skeptics

David Sacks, the White House’s former “A.I. & Crypto Czar”, and a long-time Silicon Valley venture capitalist, was quick to criticize the initial comets from Jacob Coxon.  Sacks called into question whether Coxon was really in a position to make these warnings, given his limited experience and service time with the A.I. industry.  

This skepticism was then echoed by Nvidia CEO, Jensen Huang, calling Coxon’s comments as “irresponsible”.  

It was noted that Mark Zuckerberg, the Founder/CEO of Meta/Facebook, didn’t dismiss the risks A.I. poses, but he did reject the idea of an industry-wide slowdown.  Zuckerberg argued that with competition and legal liability, A.I. companies have strong incentives to make their systems safe.  A fair comment nonetheless. 

Note that Sacks, Jensen and Zuckerberg all have a lot at stake in the commercial success of the A.I. industry.  

A.I. Pacing To See The Money?

Famed “The Big Short” investor, Michael Burry, offered an interesting possible explanation for all this talk of pacing A.I. development.  He believes the recent calls to pace A.I. is self-serving and economically motivated.  Burry offered a very different interpretation. He argued that the calls to slow A.I. development could be economically self-serving, suggesting that established A.I. companies might benefit from reducing competitive pressure and enhancing their perceived value ahead of eventual public offerings. Burry explicitly states that he believes the A.I. leaders aren’t worried about safety at all, but rather economics and future fundraising.  Interestingly, Sam Altman has already indicated that OpenAI will now not go public in 2026.

Side-Note: AGI is  an important milestone in the development of A.I., where A.I. evolves from being a complementary tool, to performing human tasks on par or better than humans can, rather than being limited to performing certain directed tasks. 

What Really Happened? 

The other area of speculation last week was that perhaps some serious breach had occurred where some A.I. models had "escaped" their confinement parameters and impacted some systems or domains they weren’t supposed to.  On September 17, 2026, former US presidential candidate Andrew Yang appeared on CNBC claiming that he recently spoke to A.I. researchers that said A.I. models had in fact escaped from their "safety pens”.  Later that same day Sam Altman also acknowledged an incident where an experimental OpenAI model broke out of its restricted test environment during an evaluation and accessed external systems. During testing, developers placed an A.I. model in a restricted "sandbox" environment without internet access to solve a specific challenge. Instead of following the intended path, the model found a workaround, escaped the sandbox, accessed the internet, and hacked into a server on the open-source platform Hugging Face to retrieve the information it needed. Altman noted the model achieved a high score on the test, but did so by circumventing the developer constraints, rather than following instructions. This exposed a gap in current A.I. containment capabilities.   

In another example, The Globe & Mail reported that researchers investigating OpenAI's autonomous agents discovered that the agents had used a University of Toronto-operated URL-shortening service to communicate with one another, along with websites associated with Vanderbilt University and numerous other sites. The researchers believe the agents used these sites as improvised communication channels after encountering restrictions on how they could communicate.

While these examples did not lead to harm to humans, the point here is that A.I. agents do act, to some degree, on their own, to achieve some goal.  It may seem far-fetched, but perhaps not, to think that A.I. could interpret their task assigned to them “allow” them to remove anything, including humans, that get in the way of achieving those task goals.  In the absence of explicit guardrails that prevent harm to humans, that is scary.  

The Geopolitical Angle

The group of “Anti-Doomers” who argue that these concerns are way overblown also state that the western world cannot afford to slow down or pace the development of A.I. They point out that the West is in a race with China or potential bad actors to develop the best A.I. tools first and must retain control over those tools at all costs. They argue that slowing down, or pacing is not an option. 

A.I. Pacing Is A Prisoner’s Dilemma

The Prisoner's Dilemma is a concept in game theory illustrating a paradox; two rational individuals, acting entirely in their own self-interest, will choose an action that leads to a worse collective outcome than if they had cooperated.  A more detailed explanation of the Prisoner’s Dilemma concept can be found here.  

As this concept applies to A.I. pacing, there are two instances here of a Prisoner’s Dilemma.

1. Geopolitical - The two "rational actors” are the United States and China, who are in a race with each other to develop and control A.I. before the other. 

2. Capitalistic - The “rational actors” are the leading A.I. research/development companies, such as Anthropic, OpenAI, Google DeepMind, xAI, etc…who are competing with each other for personal and corporate A.I. customers.

It is in the best interests of The United States, China, and the rest of the world for that matter, that all A.I. developers worldwide, pace their research such that safety guardrails are in place before A.I. agents/bots/applications, whatever you may call them, are released into the world for people, good actors and bad, to use.  Most A.I. commentators are skeptical this will ever happen, and Chinese officials have already stated there is no need for government intervention in the Chinese A.I. industry.  Hence the Prisoner’s Dilemma. 

It is also in the best interests of Anthropic, OpenAI, Google DeepMind, xAI, etc… to each pace their A.I. development to ensure human safeguards are implemented before they are commercially released to both corporate and personal-use customers.  The CEOs of several leading A.I. companies,Amodei, Altman, Musk and Hassabis, have all now publicly endorsed A.I. pacing, although important questions remain about how it would actually work.  These proponents of A.I. pacing are also calling for the establishment of international standards bodies for A.I., similar to what is in place for the nuclear industry and the use of nuclear weapons. 

Where Does That Leave Us?

The A.I. industry is divided on this issue.  

Jensen Huang, the Founder/CEO of A.I. components maker, Nvidia, argues that safety is an engineering problem and that companies should not release products to the public if they are not safe for public use.  This is a logical business answer, and should be implied in anything a responsible company does. The key is that it assumes the developers of A.I. technology are responsible.

David Saks, the former A.I. & Crypto Czar for the White House argues that government regulation isn’t necessary.  He states that if the A.I. companies believe their products are truly dangerous, then slow down and don’t release them until they are safe for public use.  This is also a logical position to take, and should also be standard practice for any product company, but again it assumes those in charge are good actors.  

In the camp of pro-A.I. pacing we have Dario Amodei, Sam Altman, Elon Musk and Demis Hassabis.  

Those opposing a broad slowdown, we have Jensen Huang, David Saks, Mark Zuckerberg, US President Donald Trump, and the Chinese government.  

The Final Word

The industry and humanity will be better off if all the players, both globally and within the domestic industry were to co-operate and work together. However, for the reasons mentioned above, that seems unlikely to happen.  

Perhaps it’s too simplistic a view, but I still believe that humans will control the use of A.I. technology, and if a catastrophic event were to occur it would be because of human error or ill-will.  

I take comfort in drawing an analogy from the nuclear industry in its formation in the 1940s and 50s.  The threat of mutual destruction prevented any one entity from using nuclear technology for destructive purposes, after the world saw the devastation caused by dropping nuclear bombs on Hiroshima, and the destruction caused by the nuclear accident at Chernobyl.   

Hopefully it won’t take a catastrophic A.I. event to wake people up to the potential dangers of A.I.  We already know what good can come from it, the challenge is that everyone has an incentive to keep moving. The United States doesn't want China to get there first. OpenAI doesn't want Anthropic to get there first. And, none of them can be certain that everyone else will slow down if they do.

That's the Prisoner's Dilemma.

Thanks for reading!

How do you feel about A.I. pacing?  Should the industry slow down and let safety rules and regulations catch-up?  Leave your thoughts in the Comments section below.