Jennifer Huddleston's position going into the segment was one sentence the host read out: "There is a significant difference between companies choosing to pause and a government mandate that could pause all that AI development."
The week's debate has been about whether the frontier labs should slow down. Her argument is that the word "should" is doing two different jobs — a company exercising its own judgment, and a rule applied to everyone at once — and that only the second one has to be justified to the public.
"If an individual company wants to not advance with their product, they have the right as an innovator, as an entrepreneur to say, this is where we're stopping."
Huddleston is a Senior Research Fellow at the Cato Institute who works on technology policy, and her case rests on a mechanism rather than on optimism: a statute is fixed at the moment it passes, and the thing it governs is not.
The full segment is covered here so you can skip it.
Here are the 6 arguments that matter.
👤 Guest: Jennifer Huddleston, Senior Research Fellow at the Cato Institute, where she works on technology and innovation policy
🎙️ Host: Jonathan Ferro, a Bloomberg Surveillance anchor
📰 Published: 14 September 2026 on Bloomberg Surveillance
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | ⏱️ length not available
Key Takeaways
A lab pausing itself and a government pausing the industry are different events, and only one of them stops the benefits too
She named cybersecurity defense, medical research and Defense Department work as what a blanket pause would also halt
Her answer to the coordination problem is industry norms rather than statute, because a rule is fixed the day it passes
The failure mode she names is a pacing problem: a beneficial application blocked by an outdated regulation
The harms people name are mostly already illegal under laws that exist
Cybersecurity and fraud statutes already reach a bad actor using a new tool
Electricity and water are grid and water problems, and she wants them argued as such rather than as AI problems
The regulatory risk she flags is the states, not Washington
AI executive orders go back to the Obama administration and have swung with each one since
By 2028 the harder question may be what still counts as AI at all
Nobody calls email autocomplete artificial intelligence any more
1. Pausing Vs A Mandate
The host's opening question was the blunt version: if the labs want to slow down, why don't they just slow down?
Huddleston said that is genuinely one of the options. "If an individual company wants to not advance with their product, they have the right as an innovator, as an entrepreneur to say, this is where we're stopping."
What follows from one firm stopping is that the others do not. She described a very competitive market in which the rest continue with what she called the beneficial applications of more advanced AI
Her objection to a mandate is that it stops both sides of the ledger. "While a lot of the conversation around this has been about the potentially risky impacts of AI, we also have to consider what this means for many of the significant benefits of AI" — she named the ability to respond to the same cybersecurity concerns that worry people, and advances in fields such as medical research
She added the Defense Department to the list, saying the department finds this research paramount to what it is doing right now
2. Norms Move, Laws Don't
Pressed on whether a single company holding back is useless unless all of them agree, Huddleston moved the answer away from law.
What she expects to emerge on specific issues is industry norms. On narrow concerns such as cybersecurity, she said, the response could take the form of self-regulation rather than top-down government rules
The advantage she claims for that is speed of revision. Norms can adapt as new risks appear and as new benign or beneficial uses appear
The disadvantage of the alternative is stated as a property, not a complaint. "When we're looking at government regulation, government regulation is a very static element." It takes a long time to change as the technology changes, and AI is moving quickly
The failure mode has a name. "And so you could end up with a sort of pacing problem where even if technology is good and beneficial, it's unable to go forward if a government regulation is particularly outdated."
3. Congress, Then The States
Asked whether any US action would come from Congress or the executive branch, Huddleston gave the conventional answer and then flagged where she thinks the risk actually sits.
On an issue this size, she expects legislation eventually. "When we're talking about this type of major issue, you're likely to need some sort of congressional framework at some point."
The thing she says is already a problem is subnational. "What's also been concerning is the number of states that have started to act around AI issues."
Her evidence that executive action is an unstable foundation is historical. Executive orders on AI have changed significantly between administrations, and "There have also been AI executive orders going all the way back to the Obama administrations." The point is that this predates the current argument by several presidencies
The host's framing was a patchwork, and asked how disruptive a state-by-state patchwork would be both to the industry and to the development of the norms Huddleston prefers
4. Regulate Apps, Not AI
The host laid out the range of worries — electricity costs, water use, and now human extinction — and asked whether AI should be treated like a nuclear product or like a technology with growing pains.
Her first move was to narrow the object of regulation. The concerns people raise are about specific applications, so it is unlikely that regulating the technology as a whole is the right response, and an attempt to do so should give people pause
For the misuse cases, she says the statute book is already written. "When it comes to a lot of those concerns, there are existing laws in play already around things like cybersecurity, around things like fraud" — and, more broadly, any case where a bad actor uses a technology to do a bad thing
On electricity and water, she wants the argument moved rather than settled. Those pressures, on her account, are an opportunity to have a broader policy discussion about the grid and about water use generally, not only as they relate to AI
5. Railroads, Not Warheads
The host put the choice as a historical one: is this the arrival of the railroads, or is it the Cold War nuclear treaties?
Huddleston did not pick either, and explained why the question keeps recurring. "One of the things that makes artificial intelligence so unique is how general purpose of a technology it is." That is why several very different conversations run at once
Her example of the breadth is access to information. She called AI an incredible tool for reaching information, while noting there are many different kinds of information one can reach
The question she says belongs at the center of the regulatory debate is about authority. What does it mean for government to be the body that dictates how the technology evolves, as against something emerging from industry norms
6. What Counts As AI In 2028
Asked as a thought exercise what regulation looks like by the next presidential race, Huddleston gave two answers and the second one was the more interesting.
The first is about which venue wins. She expects the state level to be where the action is, and the question to be whether a congressional framework arrives and, if it does, whether it keeps what she called the light touch that allowed the US to excel in areas such as the internet, or turns out to be much more regulatory
The test she applies to a heavier regime is that it makes benign applications harder to ship, not that it makes dangerous ones harder
The second answer is a definitional one. "Very few of us think about the autocomplete on our emails as a form of artificial intelligence." Nor, she said, do people think of the customer-service chatbot they pass through before reaching a person
The implication for anyone drafting a rule is that the category keeps shrinking behind them, because a technology stops being called AI once people are used to it
Bonus Insights
The host's own framing was the sharpest statement of the problem in the segment. He pointed out that the risks being discussed are not one risk: electricity costs and water use are quality-of-life questions, cybersecurity exposure is a different kind of question, and comparing the technology to a nuclear weapon is a third
Huddleston never disputed that the safety concerns are real. Her answer throughout was about where the response should sit — a firm's own decision, an industry norm, an existing statute, or a new one — rather than about whether a response is needed
She put the beneficial uses in the same sentence as the risks each time, which is how her argument avoids resting on a forecast: the cost of a blanket pause is paid immediately in cybersecurity defense and medical research, whatever happens to the risks
Huddleston's bottom line is that the case for letting a company pause itself is easy, the case for making everyone pause is not, and that most of the harms being cited to justify the second one are already covered by laws that exist.
Products, Companies & Tools Mentioned
Cato Institute (Huddleston's employer; the segment was built around a line from her published writing on the difference between a voluntary pause and a mandate)
US Department of Defense (One of the users she named whose work would be affected by a blanket pause; she said the department finds AI research paramount to what it is doing now)
If this was worth your time, send it to someone who has to have a view on this.
Get the latest market chatter as it happens: