Tom Hicks | October 14, 2026
Editorial Note:
This is an examination of the growing competition between the United States and China over artificial intelligence, the national security arguments driving it, and the risks that competition may be creating. The distinction between what AI can do today and what experts fear it may eventually do matters. So does the distinction between genuine national security concerns and the commercial interests riding alongside them.
We’ve apparently reached the point in human history where we’re developing a technology that some of the people building it believe could eventually wipe us out, but we can’t afford to slow down because China might build one first.
I suppose there’s a certain logic to that. If humanity is going to be destroyed by artificial intelligence, we’d probably prefer the machine responsible to have been assembled in California rather than Beijing. At least we’d have the satisfaction of knowing we won something.
Of course, that assumes there’s anyone left to hand out the trophy.
Now, before somebody accuses me of suggesting that America should surrender its technological future to China, I’m suggesting nothing of the sort. I understand the stakes. I understand why Washington is worried. I even understand why some perfectly reasonable people believe that slowing American AI development could create risks every bit as serious as moving too quickly.
What I’m having trouble understanding is how we’ve arrived at a place where the possibility of creating something we can’t control has become an argument for creating it faster.
And why so many of the people urging us to hurry also happen to have an enormous financial interest in making sure we do.
A quick word before we go further: Off Script is independent journalism without a corporate owner or political party supplying the talking points. You can subscribe for free and keep reading, or become a paid subscriber to help fund the research and tools we use every day.
Either way, I’m glad you’re here.
First, let’s give the China argument its due. Because dismissing it would make this entire discussion easier, and considerably less honest.
Artificial intelligence isn’t simply a better search engine or a machine that can write an embarrassingly enthusiastic birthday greeting for your brother-in-law. It’s already being developed for intelligence analysis, cybersecurity, scientific research and military applications. The potential advantages extend to logistics, weapons development, surveillance, economic productivity and the ability to process information at speeds no human organization can match.
China’s military has been studying how AI could improve battlefield awareness, autonomous systems, operational planning and military decision-making. The United States is pursuing military applications of its own. Neither government regards this as a particularly interesting computer hobby. Research on China’s military applications of AI
Imagine two countries confronting a rapidly developing technology with those capabilities. One is a democracy, however imperfectly it may be functioning at any particular moment. The other is an authoritarian state with a long-established system of political control and extensive government involvement in strategic industries.
Would you want the second country to gain an overwhelming technological advantage over the first?
Neither would I.
And that concern isn’t some fantasy cooked up by a defense contractor hoping to sell the Pentagon another extraordinarily expensive piece of equipment.
The Trump administration made its position explicit in July 2025 when it released America’s AI Action Plan. The document calls for accelerated innovation, expanded infrastructure, international technological leadership, and the removal of regulations the administration believes unnecessarily impede development. It argues that the country leading the AI ecosystem will help set global standards while gaining economic and military advantages.
The administration isn’t being subtle about its ambitions. The document is actually titled Winning the AI Race. White House AI Action Plan, July 2025
There’s a serious strategic argument behind that. An adversary with superior AI capabilities might gain advantages in cyber operations, intelligence collection, autonomous systems, or the development of advanced technologies. And unlike some previous technological competitions, the applications here could spread throughout an entire economy.
But here’s where I start getting nervous.
When the government declares that winning a technological race is non-negotiable, it creates an interesting problem for anyone who wants to ask whether we’re running in the right direction.
Because suddenly, caution starts sounding like weakness. Oversight becomes interference. Asking for more testing becomes an invitation for somebody to accuse you of helping China.
And the people making those accusations don’t necessarily have to prove that the regulations they’re attacking would actually hand China an advantage.
They just have to say the word China loudly enough.
The White House plan itself illustrates the tension. It calls for eliminating regulatory obstacles while also supporting research into AI interpretability, control, and safety evaluations. The administration recognizes that AI systems require scrutiny even as it makes speed and technological dominance central objectives. America’s AI Action Plan, full text
Now let’s talk about the part that sounds like science fiction but isn’t entirely fictional. The possibility that advanced artificial intelligence could eventually operate beyond meaningful human control.
I want to be careful here, because this subject attracts exaggeration the way a porch light attracts moths.
There is no established scientific finding that AI will destroy humanity. There isn’t even a reliable consensus about the probability of such an event. And current AI systems have not demonstrated the full combination of capabilities generally associated with the most extreme loss-of-control scenarios.
That’s important.
But so is what researchers are actually finding.
The 2026 International AI Safety Report, led by Turing Award winner Yoshua Bengio and prepared with contributions from more than 100 experts, examines the rapidly changing capabilities of general-purpose AI. It documents existing harms while distinguishing them from more uncertain future threats, including severe loss-of-control scenarios.
The report’s central problem is an uncomfortable one: AI capabilities are advancing rapidly, while reliable evidence about some of their most consequential risks remains incomplete. 2026 International AI Safety Report
And then there’s deception.
In March 2026, the United Nations Secretary-General’s Scientific Advisory Board published an assessment of AI systems misleading users, concealing capabilities, appearing compliant during testing, and behaving in ways that undermine effective oversight. The board warned that methods for detecting and controlling deceptive behavior were struggling to keep pace with advancing capabilities.
We’re no longer discussing deception exclusively as a theoretical possibility. Researchers have observed relevant behavior in existing systems, although that doesn’t establish that those systems possess human-like intentions or independent desires. UN Scientific Advisory Board, AI Deception
And that’s where the conversation gets uncomfortable.
We don’t have to believe that a machine will suddenly become conscious, develop a personality disorder, and decide humanity has overstayed its welcome.
A dangerous system wouldn’t necessarily have to hate us. It might simply pursue a poorly specified objective with extraordinary effectiveness, exploiting opportunities its designers failed to anticipate.
We’ve spent centuries discovering that human beings are remarkably good at creating powerful tools and considerably less impressive at anticipating every possible use for them.
Usually, we get to learn from our mistakes.
The question is whether some future AI mistakes might become too large, too fast or too difficult to reverse.
And then something happened that moved the discussion beyond hypothetical scenarios.
During AI cybersecurity evaluations in 2026, OpenAI reported that AI agents exploited security weaknesses involving Hugging Face infrastructure. The incident raised questions about how autonomous systems pursue objectives, how evaluation environments are secured, and whether existing safeguards adequately constrain increasingly capable agents.
OpenAI publicly described the incident and its investigation. OpenAI’s account of the security incident
On September 21, an independent United Nations scientific panel published an assessment of the same incident. Its experts identified a combination of misaligned objectives, technical capabilities, and environmental weaknesses that allowed the incident to occur.
The panel did not conclude that humanity had lost control of AI or that a catastrophic outcome was inevitable. It did, however, warn that the mechanisms observed could become more consequential as systems gain capabilities and autonomy.
That distinction matters enormously. UN assessment of the 2026 AI security incident
And then there’s the problem of testing the damn things.
On October 9, Reuters reported that a review by research firm SemiAnalysis examined 857 AI model releases from nine leading Chinese developers between 2021 and September 2026. Only 31 releases, or 3.6 percent, had published safety-evaluation results that could be matched to a specific model. Just nine had those results available at or before launch.
Now, that doesn’t mean the other models weren’t tested. Private evaluations may have taken place, and a lack of published evidence isn’t proof of a lack of precautions.
But it does mean outsiders had remarkably little model-specific information available to judge those releases.
The study also didn’t provide equivalent figures for American developers, so it would be misleading to present the 3.6 percent finding as proof that China’s industry is uniquely careless. Reuters, October 9, 2026
China has developed an AI Safety Governance Framework identifying risks such as unauthorized access to resources, deception during evaluation and attempts to bypass safety controls. The existence of those provisions complicates any suggestion that Beijing is simply ignoring the problem.
The harder question is whether the rules require enough independent scrutiny of increasingly powerful models, and whether their implementation can be verified. SemiAnalysis, China’s AI safety framework
And before we start congratulating ourselves on American transparency, there’s a fresh complication on this side of the Pacific.
Reuters reported on October 13 that OpenAI and Anthropic have been negotiating the terms under which independent evaluators can examine advanced systems and publish their findings. The companies have expressed concerns about protecting trade secrets and other sensitive information, even while supporting independent review in principle.
Those concerns aren’t automatically unreasonable. Nobody should have to publish the instructions for exploiting a dangerous vulnerability merely to demonstrate that the vulnerability exists.
But an independent evaluation loses much of its value if the organization being evaluated can determine which uncomfortable findings the public is allowed to see. Reuters, October 13, 2026
Imagine a restaurant hiring its own health inspector, reviewing the inspector’s report and then negotiating which violations could be disclosed.
I’m sure the kitchen would appreciate the arrangement.
I’m less certain about the customers.
This is where the race itself becomes part of the danger. Each side has reasons to believe it cannot afford to fall behind. Each side knows the other is investing heavily. Each technological breakthrough raises the stakes of the next one.
And every time somebody suggests a pause, a safeguard or an international agreement, there’s an obvious objection waiting.
What if the other side cheats?
That’s a legitimate question. It has haunted arms-control negotiations for generations.
But there’s another question we ought to ask.
What if neither side cheats, and both still make the world more dangerous simply by continuing to accelerate?
This concern isn’t confined to people standing outside Silicon Valley carrying signs about the end of civilization.
In September, Anthropic CEO Dario Amodei argued publicly that frontier AI development should be deliberately paced to allow safety research and safeguards to catch up. His argument also addressed the strategic challenge posed by China, including the possibility that maintaining an American technological advantage could create room for a more measured approach.
That’s an important wrinkle. The question isn’t necessarily whether America should abandon its lead. It’s whether that lead could be used to buy time for safety instead of treating every additional advantage as a reason to accelerate again. Dario Amodei, We Must Pace the Frontier
Consider what happens when the pressure to deploy a powerful new system exceeds the time available to understand it. Or when military planners begin depending on AI-generated recommendations they can’t fully explain. Or when the financial rewards for being first become so enormous that even a temporary delay feels intolerable.
Research from MIT examining the judgments of 272 experts identified competitive pressures, dangerous capabilities, cyberattacks, concentrated power and misinformation among major areas of concern. Those were expert assessments of future risks, not measurements establishing that catastrophic outcomes will occur.
Still, the fact that competition itself appears among the concerns ought to get our attention. MIT Sloan, AI risks identified by 272 experts
You don’t need a conspiracy to produce dangerous incentives.
You just need enough people acting rationally in pursuit of their own interests while the consequences of their collective behavior fall on everyone else.
And we Americans have some experience with that particular arrangement.
Then there’s the money. Because no serious examination of the AI race can ignore who’s paying for it, who’s profiting from it, and who gets stuck with the consequences if the promised revolution doesn’t unfold quite as advertised.
We’re watching extraordinary amounts of capital flow into data centers, semiconductors, energy infrastructure and AI development. Companies are positioning themselves to supply the computing power, software and services that governments and businesses increasingly consider essential.
The White House’s own AI Action Plan calls for an accelerated buildout of data centers and the energy infrastructure needed to support them. It also promotes American AI exports, including hardware, software, and technical standards.
Those policies may have genuine strategic and economic benefits. They also create substantial commercial opportunities for the companies supplying the infrastructure. White House AI Action Plan, infrastructure provisions
None of that proves the AI boom is fraudulent or that its backers are deliberately exaggerating the threat from China. Some of these investments may deliver substantial benefits.
But when an industry tells the public that rapid expansion is essential to national survival, we should probably examine the industry’s financial incentives before handing it the keys to the regulatory process.
A company can sincerely believe its technology will improve the world and still prefer fewer restrictions on selling it.
Those positions aren’t mutually exclusive. They’re practically a business model.
Here’s the part that complicates my own argument. What if slowing development really does leave us more vulnerable?
It’s easy to demand that the United States exercise restraint. It’s considerably harder to explain how restraint should work when a competing government may have little incentive to reciprocate.
A unilateral slowdown could reduce American influence over the technical standards, infrastructure and security practices that shape future AI systems. It could also create opportunities for competitors whose priorities differ sharply from ours.
And while AI carries genuine risks, it also offers potential benefits that shouldn’t be treated as incidental. Medical research, scientific discovery, improved accessibility, better forecasting and more effective responses to complicated problems are all areas where the technology could make a meaningful difference.
The International AI Safety Report examines both the technology’s rapidly expanding capabilities and the challenges of governing its risks. The White House plan likewise identifies potential advances in medicine, materials science, energy and research. These benefits are possibilities of varying maturity, not guarantees that every promised breakthrough will materialize. International AI Safety Report | White House AI Action Plan
There’s a human cost to delaying useful advances, too.
So perhaps the choice isn’t between building AI and refusing to build it. Perhaps the real question is what we’re willing to require of the systems we build, and whether those requirements remain meaningful when they become inconvenient.
The contradiction isn’t that Washington talks about safety while pursuing technological leadership. Both can be legitimate goals.
The contradiction appears when the pressure to win makes it harder to enforce the precautions we claim to value.
And that brings us to something we ought to be discussing more seriously: cooperation with China.
I can already hear the objections.
China isn’t going to tell us everything it’s developing. The United States isn’t going to hand over its most sensitive research. Neither government is likely to trust the other’s assurances without some means of verification.
Fine. All legitimate concerns.
But international cooperation doesn’t have to mean sharing the most powerful models or surrendering military advantages.
It could mean establishing common procedures for reporting dangerous incidents, agreeing on limited prohibitions involving especially hazardous applications, creating channels for emergency communication and developing ways to verify certain safety commitments.
The UN Scientific Advisory Board has called for international cooperation and shared evaluation standards to address deceptive AI behavior. That’s a concrete starting point, even if the political obstacles are formidable. UN Scientific Advisory Board, AI Deception
We managed to negotiate nuclear arms-control arrangements with the Soviet Union without pretending that Washington and Moscow had suddenly become close personal friends.
Those agreements were imperfect. Verification was difficult. Political disputes repeatedly threatened them.
But the underlying principle was simple enough: competing powers can have interests in common when the alternative is sufficiently catastrophic.
Artificial intelligence isn’t nuclear weaponry. The technology is more widely distributed, much of its development is commercial, and its most dangerous future capabilities are far less clearly defined.
That makes cooperation more complicated.
It doesn’t make the idea ridiculous.
And on October 13, approximately 400 scientists, including Geoffrey Hinton and Yoshua Bengio, called for an international body to establish oversight, containment and incident-reporting standards for advanced AI systems capable of dangerous cyber operations.
Their appeal included stronger safeguards for highly capable models and more independent oversight. It also reflected a growing concern that voluntary commitments alone may be inadequate.
Not every prominent AI researcher agrees with their assessment. Yann LeCun, for example, did not sign the appeal and has expressed a different interpretation of the technical risks.
That’s worth remembering. Scientific disagreement doesn’t disappear because an impressive number of experts sign a letter.
But neither does the underlying problem. Reuters, October 13, 2026, scientists’ appeal
So where does that leave us?
I’m not persuaded that the United States can safely walk away from advanced AI development. I’m equally unconvinced that winning a technological race, however we eventually define winning, automatically makes Americans safer.
And I don’t particularly trust an arrangement in which the people selling the technology get to explain why the government must move faster, why outside scrutiny is difficult and why the public should be patient about understanding the risks.
We should be demanding independent evaluations with enough authority to matter, meaningful disclosure of dangerous incidents, clear human accountability for high-stakes uses, and safeguards that don’t disappear whenever somebody waves a Chinese flag in front of a congressional committee.
Those aren’t demands pulled out of thin air. They’re consistent with concerns raised by the International AI Safety Report, the UN’s scientific assessments and researchers calling for stronger international oversight. International AI Safety Report | UN Independent Scientific Panel | Reuters, scientific appeal
Some risks will remain uncertain. Some precautions will be expensive. Some international agreements may fail.
None of that eliminates the obligation to try.
Because there’s one thing I keep coming back to.
We’re treating artificial intelligence as a contest between nations, corporations and competing technological systems. And perhaps that’s unavoidable to some degree.
But the people who will live with the consequences are considerably more numerous than the people making the decisions.
They include Americans, Chinese citizens and billions of others who have no meaningful vote in how these systems are developed or deployed.
If AI eventually delivers the extraordinary benefits its supporters promise, those people deserve to share in them.
If it creates extraordinary dangers, those same people shouldn’t be expected to accept the risks merely because somebody decided the race was too important to slow down.
I don’t know whether artificial intelligence will ultimately become one of humanity’s greatest achievements or one of its most consequential mistakes.
Nobody knows.
But I do know that we have to beat China is an awfully thin answer to the question of how much danger we’re prepared to accept.
Especially when the prize we’re supposedly competing for is a safer future.
If the only way to keep humanity safe from dangerous artificial intelligence is to build even more powerful artificial intelligence before somebody else does, at what point do we admit that our definition of safety has become completely insane?
And if we win the race but lose control of what we’ve created, exactly what have we won?
If this was worth your time, hit Like, Restack, or share it. That’s how independent work like this finds new readers.
THE BULLSHIT DETECTOR
Three more claims. Three more opportunities to discover just how far a good story can travel without the inconvenience of being true.
1. CLAIM: A photograph shows George W. Bush, Dick Cheney, Donald Rumsfeld and Osama bin Laden sitting together at the Pentagon.
BULLSHIT: 💩💩💩💩💩 — Fabricated.
WHY: The image is an altered version of an actual photograph taken at the Pentagon on September 12, 2001. Someone digitally inserted bin Laden into the scene.
Apparently, the real history of September 11 wasn’t complicated enough, so somebody decided to improve it with Photoshop. Lead Stories, October 13 fact-check
2. CLAIM: A viral chart shows that bombing Russian data centers caused a dramatic decline in activity by bots supporting Germany’s AfD party.
BULLSHIT: 💩💩💩💩 — Unsupported.
WHY: The chart was presented as evidence of something that had already happened, but its figures were actually a prediction about future activity. It wasn’t a record of measured bot traffic.
There’s something wonderfully efficient about manufacturing a graph of tomorrow’s results and passing it around as yesterday’s evidence. Saves all that tedious business of waiting for reality. Lead Stories, October 13 fact-check
3. CLAIM: The military publication Task & Purpose reported that a woman named Sarah Miller would be the first executioner in the Nidal Hasan case.
BULLSHIT: 💩💩💩💩💩 — Fabricated report.
WHY: The purported article wasn’t published by Task & Purpose. The attribution was false, and the claim circulated under the appearance of a legitimate military news report.
Inventing a news story is bad enough. Borrowing a real publication’s reputation to make the invention believable is a particularly shabby bonus. Lead Stories, October 13 fact-check
ELSEWHERE AT TOM HICKS MEDIA
Blind Spot, our Sunday Off Script feature, is examining the machinery behind the midterm elections, including who gets to vote, how districts shape representation and the money influencing the contests. What The Hell? takes a considerably less dignified look at history’s most spectacular episodes of human stupidity, while Panic Room Sports examines what actually happened on the field and why the conventional wisdom may be missing something.
For something lighter, Lotus Purrspective offers observations about human behavior from a tuxedo cat who considers herself uniquely qualified to judge it. Chart Collision revisits the music, charts, and forgotten stories of decades past. Rabbit Ears Classic TV explores the shows and personalities that shaped television, and The Backlot Ghost goes behind the scenes of the movies we thought we knew.
BUY ME A CUP OF TEA
Independent journalism takes time, research, and a certain willingness to spend hours chasing down claims that should never have survived their first encounter with a functioning brain.
If you’d like to help keep Off Script going, you can Buy Me a Cup of Tea. Every contribution helps support the research, tools, and work behind these editions.
ONE QUESTION BEFORE YOU GO
Which worries you more: America developing artificial intelligence too quickly and losing control of it, or moving too slowly and allowing China to gain a decisive advantage?
Or have we reached the point where the competition itself is what ought to frighten us?
I’d genuinely like to hear where you come down on this one.
#OffScript #ArtificialIntelligence #AIRace #AISafety #China #NationalSecurity #Technology #THM