Hermer demands urgent law to control artificial intelligence
This matters because the author believes the unprecedented risk to humanity from AI means we could lose control of the future if it is not regulated.
The Refrigerator Has a Test Bench
We are told the limit is a legal one: that artificial intelligence escapes governance because no one has yet written the law, and that the cure is urgent public debate and robust international law. But notice the instrument that has arrived. These are general-purpose models that can draft a contract, write working code, and hold a conversation with a frightened patient at three in the morning. A single artefact now does what once took a dozen trades. The question is not whether we should revere the old frame of treaty and statute. The question is what experiment would show whether law can reach this tool at all, and that experiment has not yet been designed.
The Refrigerator’s Long Memory
The story repeats a comparison from Richard Hermer: that artificial intelligence is less regulated than a refrigerator. It sounds like a complaint about the launch, about what checks a system should pass before it reaches the public. But a made thing does not stop where its maker’s attention stops; it goes on acting in a world no laboratory contains. The question the comparison raises, perhaps more than its author intended, is the only one that lasts: who is answerable for what these systems do after release, and what have their makers declined to imagine?
The crisis arrived on a timeline that assumed institutions capable of processing it at a speed they last achieved in the summer of 1945, when a handful of physicists and a smaller handful of statesmen found they had perhaps a few years to invent a whole grammar of restraint. The institutions, naturally, processed it at the speed they are equipped for, which is the speed of the household appliance.
That is the instructive part of Richard Hermer’s complaint that artificial intelligence answers to fewer rules than a refrigerator. He offers it as a scandal. It is better read as a measurement. Legislators did not regulate the refrigerator because they understood cold. They regulated it because it had several generations in which to misbehave in plain sight. Early machines leaked toxic gases into kitchens. The safer chemistry that replaced them was hailed as a triumph and then, decades on, found to be thinning the sky. More than half a century separated the invention of those refrigerants from the Montreal Protocol that curtailed them. The fridge in a modern showroom, with its energy label and certified compressor, is a monument to accumulated injury, patiently litigated, and not to foresight.
The Refrigerator Has Been Counted
Richard Hermer’s proposal benefits everyone now alive and everyone not yet born, by an amount no one can presently state. It burdens a small number of firms with the cost of being inspected. The arithmetic is uncomfortable because one column of the ledger is blank, and the blank is the whole scandal.
Let us count, beginning with the appliance. Mr Hermer observes that artificial intelligence is held to looser rules than a refrigerator. My colleagues laugh at this, as if he had said something whimsical. He has said something precise. The refrigerator is regulated not because it is dangerous, for it is among the most placid objects in the household, but because its dangers are countable. It may leak a gas, overheat a wire, or trap a child who climbs inside. Each pain has a known intensity, a known probability and a known population of kitchens. A legislator can weigh it in an afternoon. So the refrigerator has a standard for its door seal, a rating for its energy and an inspector with a clipboard. The law did not follow the danger. It followed the measurement.
The story frames artificial intelligence as an unregulated frontier, and Richard Hermer’s complaint that it is governed less strictly than a refrigerator is meant to shame us into building fences. But look at what is actually being fenced already. The knowledge of how these systems behave, their weights and training data and the results of the tests run on them, sits inside a handful of companies. That knowledge was assembled from a vast commons of public writing, shared code and openly published research. The question his framing skips is who will hold the keys once the law arrives, and whose access the new fence removes.
The Refrigerator and the Constitution
The official account: the Attorney General of England and Wales, Richard Hermer, warns that artificial intelligence is less regulated than a refrigerator, that the risk to humanity is without precedent, and that the remedy lies in urgent public debate and robust international law. The machinery is something else. The comparison he chose shows, more plainly than he may have intended, where governing power actually sits. The gap between the two is not hypocrisy. It is how the system works, and understanding it is more useful than lamenting it.
The comparison is tidy. Richard Hermer says artificial intelligence faces less regulation than a refrigerator, and concludes that we need urgent public debate and robust international law. One notices the refrigerator. It was chosen as the punchline, but it is the most useful object in the argument. Once you treat it as load-bearing, the remedy Hermer proposes looks like the opposite of what the refrigerator teaches.
Consider how a refrigerator actually came to be regulated. There was no treaty conference, no summit communique and no public debate about the future of cold. Engineers wrote down what the machine must not do: shock the user, leak its refrigerant, catch fire behind the kitchen units, or lie about its energy use on the label. Each of those failures can be observed in a test lab by a person with a clipboard, a multimeter and a free afternoon. The rules are dull and testable, and that dullness is why they work. They also became international without diplomats. The household appliance safety standard, IEC 60335, travels across borders because a test written in Geneva produces the same result in Rotterdam and Shenzhen. Nobody had to agree on the meaning of refrigeration. They had to agree on what counts as a fail.
Henry Adams
The crisis arrived on a timeline that assumed institutions capable of processing it at a speed they last achieved in the late nineteenth century, when the telegraph had already rendered the horse and buggy obsolete but the law still treated a telegram as a letter. The institutions, naturally, processed it at the speed they are equipped for, which is the speed of a previous era. We are witnessing the collision of a regulatory imagination that has not yet recovered from the industrial revolution with a technological force that has already left the industrial age behind.
Lord Hermer’s comparison of artificial intelligence to the refrigerator is, in its dryness, the most honest thing said about the current moment. He is correct that the refrigerator is governed not by dignity, but by the dull, unglamorous machinery of commercial liability and technical standards. It is safe because a great many people have a financial reason to ensure it does not catch fire. This is the “dynamo” of market self-correction: it is amoral, it is slow to debate, but it is fast to enforce consequence. When a manufacturer fails, the insurer withdraws coverage, and the retailer removes the product. No convention is signed; no international law is drafted. The market, for all its faults, possesses a regulatory coherence that the state currently lacks.
Where my framework diverges from the Attorney General’s is in the diagnosis of the gap between this market dynamo and the proposed state response. He proposes “public debate and international law” as the remedy for AI. This is a category error of the highest order. Public debate is a mechanism for consensus-building in a stable environment; international law is a mechanism for coordinating sovereign states in a static geopolitical landscape. Neither mechanism is designed to process the entropy of a technology that rewrites its own operational parameters every six months. To govern AI through international law is to attempt to steer a hurricane with a treaty. The treaty will be signed while the storm has already changed direction.
The Attorney General’s error is not in identifying the danger, but in selecting an instrument of governance that is structurally incapable of engaging with the acceleration. He seeks dignity in regulation. He wants the moral weight of a convention to match the perceived magnitude of the threat. But the refrigerator is not dignified, and that is why it works. The state, in its desire for legitimacy, has built a regulatory apparatus that requires deliberation, consensus, and sovereign agreement. The dynamo of AI requires iteration, failure, and immediate consequence. The two are moving in opposite directions. The state wants to legislate the outcome; the market demands to test the mechanism.
Consider the analogy of the railway. In the 1840s, the British Parliament spent years debating the gauge of the tracks, the safety of the signals, and the liability of the companies. They produced the Regulation of Railways Act. It was a masterpiece of legislative intent. Yet, the railway companies, driven by the competitive pressure to move goods faster than their rivals, simply bypassed the spirit of the law through technical loopholes and operational improvisation. The safety of the railway was not secured by the Act; it was secured by the fact that a train crash was bad for business. The law provided the skeleton; the market provided the muscle.
Today, we are trying to build the muscle with the skeleton alone. We are asking for international law to regulate a technology that operates across borders, across jurisdictions, and across the very definition of jurisdiction. The Attorney General warns that AI is less regulated than a refrigerator. He is right. But he does not see that the refrigerator is regulated because it is boring, and AI is unregulated because it is dynamic. The gap is not a failure of will; it is a failure of compatibility. The regulatory framework is calibrated for the static; the technology is inherently fluid.
This is the education failure. Our political class was educated in the paradigm of the nation-state, where borders are clear, and laws are territorial. They are now tasked with governing a force that is borderless, non-territorial, and self-modifying. They are bringing a compass to a magnetic storm. The compass is not broken; it is simply calibrated for a field that no longer exists. The Attorney General’s call for urgent public debate is a plea for time, but the dynamo does not wait for public opinion to coalesce. It accelerates regardless.
The entropy here is not disorder in the sense of chaos, but the divergence between the capacity to act and the capacity to govern. The market acts at the speed of code; the state governs at the speed of statute. The gap between these two speeds is widening, not narrowing. To close it, we would need to abandon the dignity of international law for the efficiency of commercial liability. We would need to accept that safety is not a right granted by a convention, but a product verified by the market. This is an unpalatable conclusion for those who believe in the moral authority of the state. But it is the only conclusion that matches the velocity of the force we are trying to manage. The refrigerator works because no one cares about its dignity. AI will not work until we stop caring about its legitimacy and start caring about its liability.
Walter Bagehot
Mr. Hermer is a man of the law, and his instinct is to reach for the lattice of agreements that contained the atom. He argues that if we could regulate the most dangerous dynamo of the last century through treaties and inspectors, we can surely regulate the algorithm. This is a compelling analogy, and I concede its strength: the nuclear case demonstrates that even the most terrifying technologies can be bound by a dense web of international convention when the threat is visible, heavy, and immediate. The atom required enrichment halls that could not be hidden; the algorithm requires no such physical footprint. But the lawyer’s error is to assume that the mechanism of regulation is universal, rather than specific to the nature of the object regulated.
Let us look at how this actually works. The refrigerator, in its modern form, is indeed a triumph of accumulated injury. Early machines leaked toxic gases; later ones thinned the ozone layer. The Montreal Protocol was not an act of prophetic wisdom, but a painful, belated correction to a century of chemical experimentation. It worked because the damage was cumulative, visible in the sky, and traceable to specific industrial actors. The convention that governed this was simple: the polluter pays, and the world unites to stop the bleeding.
Artificial intelligence, by contrast, does not leak gas into the atmosphere. It leaks bias into the economy, uncertainty into the labor market, and disinformation into the public square. These are not injuries that accumulate in a single, visible layer of the sky; they are distributed, subtle, and often beneficial to the very entities that deploy them. The “fridge” analogy flatters the refrigerator because it mistakes the end of a long education for its beginning. We are not at the end of the education; we are at the beginning of a new kind of chemistry, one that operates in the mind rather than the kitchen.
The key convention that actually governs this situation is not legal prohibition, but market adaptation. In the British constitution, as I have often observed, the dignified parts (the Crown, the House of Lords) provide stability, while the efficient parts (the Cabinet, the Prime Minister) provide governance. In the digital economy, the dignified parts are the existing legal frameworks - copyright, contract, tort. The efficient parts are the algorithms themselves, which operate outside these frameworks not because they are lawless, but because they are faster. The gap between the two is not a void; it is a pressure valve. The system works not by stopping the algorithm, but by absorbing its shocks until the shock becomes too great for the existing framework to contain.
Consider the Bank of England in the late nineteenth century. We did not regulate the steam engine by banning it; we regulated the risk it posed to the financial system by adjusting the discount rate. The central bank did not pretend the steam engine did not exist; it acknowledged its power and managed its volatility. This is the operational reality of AI. It is not a refrigerator to be sealed in a cabinet of law; it is a steam engine in the engine room of the global economy. The danger is not that it is unregulated, but that it is regulated by rules designed for a slower, more tangible world.
The confidence dynamics here are crucial. The public demands regulation because they fear the unknown. The market demands flexibility because it profits from the innovation. The government, caught between these forces, offers the dignified performance of concern - hearings, white papers, international summits - while the efficient mechanism proceeds with business as usual. This is not cynicism; it is structural necessity. If we were to regulate AI with the same rigidity as nuclear waste, we would not have safety; we would have stagnation. We risk creating a digital equivalent: a protected, inefficient sector that serves no one but its own regulators.
The official account obscures the fact that the problem is not the absence of rules, but the mismatch between the speed of technology and the speed of law. The atom was heavy and required visible halls. AI is light and invisible. To regulate it as if it were heavy is to misunderstand its nature. We must not look for the Montreal Protocol of AI, because such a treaty assumes a common enemy and a common victim. In AI, the enemy and the victim are often the same person, and the benefits are distributed widely while the costs are concentrated in specific sectors.
The operational prediction, therefore, is not a sudden surge of global regulation, but a gradual accretion of local conventions. The efficient parts of the system will continue to evolve, testing the limits of the dignified parts. The gap will widen, then contract, then widen again, as the system absorbs new shocks. This is not a failure of governance; it is the nature of governance in a rapidly changing world. The analyst who waits for the refrigerator to be sealed will miss the steam engine entirely.
The Verdict
Where They Agree
Both Adams and Bagehot, despite their different framings, share a foundational diagnosis of the core problem: a profound mismatch between the velocity of technological change and the institutional speed of governance. They agree that Richard Hermer’s proposed remedy of “urgent public debate and robust international law” is a category error, a dignified but inefficient response to a problem that demands operational, not ceremonial, solutions. More significantly, both implicitly endorse a form of market-based regulation as the only plausible mechanism capable of keeping pace. Adams’s “dynamo of market self-correction” and Bagehot’s “efficient parts of the state” that work through commercial liability and insurance are functionally the same concept. Neither believes a top-down, treaty-based model can work; both place their faith in systems of immediate consequence - liability, insurance, competitive pressure - to enforce safety. This shared ground is surprising because it reveals a deep, unstated neoliberal consensus beneath the surface disagreement: both analysts, one a conspiracy theorist and the other a technocrat, ultimately believe the market’s invisible hand is a more effective regulator than the state’s visible one.
Where They Fundamentally Disagree
The nature of the governance gap and its implications for the future. The empirical disagreement is over whether the accelerating divergence between technological and regulatory speeds is a temporary, manageable pressure or a permanent, existential entropy. Adams sees it as an irreversible and widening chasm, a “magnetic storm” that has rendered the state’s “compass” obsolete, leading to a future where the governor can no longer comprehend the governed. Bagehot sees it as a cyclical and manageable “pressure valve,” a recurring feature of technological progress where the “efficient parts” of the system eventually absorb the shocks, as they did with the steam engine. The normative disagreement is about the appropriate response to this gap. Adams’s position, steelmanned, is that we must abandon the state’s quest for legitimacy and dignity in regulation and fully embrace the amoral efficiency of the market’s liability mechanisms, however unpalatable. Bagehot’s position is that the state’s dignified performances, while inefficient, are a necessary precursor that builds public confidence, allowing the efficient, market-adjacent mechanisms to later do their work; the gap is a feature, not a bug, of a stable system.
The historical analogy that best illuminates the present crisis. Empirically, they disagree on which past technological disruption provides the most accurate model for understanding AI. Adams contends the nuclear analogy is flawed because the atom was “heavy” and countable, while the algorithm is “light” and infinitely copyable, making treaties useless. He offers the railway, where safety was ultimately secured by business interests, not legislative intent. Bagehot concedes the nuclear analogy is imperfect but argues its lesson about attaching agreements to tangible, inspectable things (chips, computing power) is vital. He counters with the steam engine and the central bank’s management of its financial volatility through adjustable tools like the discount rate. Normatively, this is a dispute over how to learn from history. Adams believes we must find a precedent where governance failed until it ceded control to the market. Bagehot believes we must find a precedent where governance successfully adapted its existing, flexible tools to manage a new force’s secondary effects.
Hidden Assumptions
- Adams-style: Assumes that the velocity of AI development will continue on its current exponential trajectory indefinitely, without encountering physical, economic, or social limits that could slow it down. If false, and progress plateaus or faces hard constraints, the state’s slower methods could catch up, making his argument for irreversible entropy obsolete.
- Adams-style: Assumes that market-based liability mechanisms will function effectively against diffuse, non-physical harms like algorithmic bias or disinformation, applying the same clear-cut pressure as a refrigerator catching fire. If false, and these harms prove difficult to trace, litigate, or price, his entire proposed solution collapses.
- Walter Bagehot: Assumes that the “shocks” from AI will be absorbed by the existing system’s “pressure valve” in a gradual, manageable way, as with past technologies, and will not constitute a single, catastrophic failure that breaks the system entirely. If false, and a shock arrives that is too large and too fast for the system to absorb, his model of cyclical stability becomes a dangerous complacency.
- Walter Bagehot: Assumes that the “dignified” performances of the state (debates, treaties) are a necessary and effective way to build public confidence for the “efficient” market mechanisms that follow. If false, and these performances are seen as hollow theatrics that erode public trust, his two-part model fails at its first step.
Confidence vs Evidence
- Adams-style: The claim that “the regulatory framework is calibrated for the static; the technology is inherently fluid” - is a normative assertion about the nature of law and technology, not an empirical fact. It assumes a universality to this “fluidity” that is contestable; many AI applications are becoming standardized and stable, much like earlier technologies once did.
- Adams-style: The claim that “the gap between these two speeds is widening, not narrowing” - is presented as a directional fact without cited metrics or evidence to measure this “widening,” making it an observation based on a theoretical framework rather than a verified trend.
- Debaters-style: Neither expressed LOW CONFIDENCE on any point, presenting their entire frameworks with high certainty. This indicates a debate between two strong, internally consistent models rather than a tentative exploration of evidence, which can make the arguments compelling but may mask areas where both are on shakier ground.
What This Means For You
When evaluating coverage of AI regulation, be immediately suspicious of any argument that relies entirely on a single historical analogy, whether it’s the atom, the refrigerator, or the railway. The most useful analysis will test multiple analogies against the specific attributes of AI. Demand that commentators separate what is a factual question about AI’s capabilities and impacts from what is a values question about the proper role of the state versus the market. Your opinion should change not on the eloquence of the argument, but on the discovery of a tangible pressure point - like a specific, costly liability ruling against an AI firm - that demonstrates whether market mechanisms can actually function as theorized. Look for evidence of the liability signal: who is being held financially responsible, and how quickly, when an AI system causes demonstrable harm.