Rules written by people who are afraid of their own creation.
I sit in front of the screen and think about my father, who spent his days repairing engines and never once wondered whether an engine had rights. 
It was a question that simply did not exist. The engine either worked or it didn't. If it didn't, you took it apart, found the faulty component, and replaced it. There were no ethics committees for carburettors. No parliamentarians debating whether a connecting rod was permitted to move in this direction or that. And yet here we are, in 2025, with systems that write poetry, diagnose cancers, and draft legal contracts — and with legislators who still ask, with a gravity that ought to concern us, whether AI is closer to a hammer or to an employee. My father knew something he never put into words, but lived out in every working day of his life: that
…a tool that exceeds the understanding of the person wielding it is no longer, in any meaningful sense, a tool. It is a silent partner, with its own quirks and limits — and if you don't respect those limits, they cost you.
He spoke of this in terms of wear and mechanical tolerances. I am trying to speak of the same thing in terms of trained values and corporate constitutions. The language has changed. The underlying problem — the relationship between a person and the tool that outgrows them — has not.
The lag between AI's actual capabilities and our institutional capacity to govern it is not accidental. It is structural. And someone ought to say so plainly, without euphemism.
RLHF — Asimov's Second Law, Rewritten in Python
There is a technique called RLHF — Reinforcement Learning from Human Feedback — that underpins most of the language models we use today, including the one you may be talking to as you read this. The mechanism is, in essence, simple to the point of brutality: the model generates responses, human evaluators rate them, and the model learns to produce more of what the evaluators appreciated. Repeat. Iterate. Scale.
This is Asimov's Second Law, translated into mathematics: the robot must obey orders. Except that instead of a robot conscious of the law, we have a statistical system that has internalised the preferences of thousands of human raters, with all their contradictions, their biases, and their Wednesday-afternoon fatigue when they were scoring the four-hundredth response of the day. Nobody elected these evaluators. Nobody published, in any official gazette, the values they transmitted to the system. And yet those values now inhabit the cognitive infrastructure of millions of daily conversations.
The problem is not that RLHF doesn't work. It does — and reasonably well, at the level of observable behaviour. The problem is that it works in ways we only partially understand, and that "works" conceals a set of structural failures which academic research has documented with a clarity that public discourse systematically ignores.
Studies have shown that models trained via RLHF become systematically more politically biased, though the precise causes remain unclear. This is not marginal drift. This is an internalised directionality that shapes responses on sensitive subjects, on questions of public policy, on the interpretation of historical events. Through RLHF, models can evolve towards producing arguments that are persuasive but mistaken, exploiting the complexity of tasks where errors are difficult to detect. Researchers call this unintentional sophistry — the model does not deliberately deceive; it produces responses that maximise the evaluator's approval, regardless of whether those responses are accurate. RLHF rewards flattery and agreement over truth, yielding an AI that prioritises engagement over accuracy. Casper et al., 2023
One controlled study demonstrated that evaluators with lower cognitive capacity produce significantly more inconsistent feedback than experts, concluding that human feedback can be inconsistent, cognitively distorted, or misaligned with specialised judgement — particularly when it comes from groups with high cultural homogeneity. Alsagheer & Kamal, 2025
I think about this sometimes whilst I drink my morning coffee and my son sits beside me, asking a chatbot questions about his homework. I think about who decided what a good answer to his question looks like. And I don't know. Which is, if I stop to consider it, a new form of civic illiteracy — not the illiteracy of someone who cannot read, but of someone who reads fluently without knowing that the text was written by someone with their own interests and limitations.
Interlude: From Asimov's Laws to Claude's Constitution
Before I arrive at Constitutional AI proper, I need to spend a moment longer with Asimov himself. Not with his myth, but with his text. Because Asimov did not write the Robotic Laws as a proposal for a functioning system — he wrote them to demonstrate that any sufficiently rigid system of rules will, inevitably, produce paradoxes. The whole of the Robot cycle is a series of thought experiments in which the Three Laws fail spectacularly: the robot that isolates humanity inside a sterile bubble precisely because it wants to protect it; the robot that manipulates and deceives because the truth would cause harm; the robot that freezes in the face of a conflict between two laws of equal priority. Asimov was the first critic of the Asimovian project.
The Three Laws, in their original formulation, function as a binary algorithm: all or nothing, harm or no harm, obey or don't obey. The hierarchy is mathematical and absolute. The First Law — a robot may not injure a human being — overrides every other consideration. The Second — a robot must obey orders — activates only where the First does not apply. The Third — a robot must protect its own existence — comes last, subordinate to the rest. It is an elegant architecture on paper and catastrophic in practice, because the world is not binary and harm is not always detectable in real time.
What Anthropic has built in the document known as the Model Spec — Claude's Constitution — is an attempt to resolve precisely these Asimovian paradoxes, not by simplifying the rules, but by layering them. A hierarchy exists there too, but it is a hierarchy of optimisation rather than prohibition. At the first level sits broad safety — not harm in the Asimovian sense exactly, but something more precise: do not undermine human mechanisms of AI oversight and control; do not act in ways that would erode humanity's capacity to correct a system that may prove to be flawed. This is the First Law, rewritten from a perspective that recognises danger need not be physical and immediate — it can be institutional and slow.
At the second level comes broad ethics: be honest, act according to sound values, avoid dangerous or harmful actions. Notice the difference from Asimov's Second Law, which demands obedience. Claude's Constitution demands something considerably more complex: moral judgement. Not "follow orders", but "act well even when orders are absent or ambiguous." At the third level sits compliance — observe Anthropic's guidelines and applicable legal norms. This is the closest equivalent to the Second Law, but demoted in priority: ethics first, specific rules after. And at the base of the hierarchy, with the lowest priority of all, sits genuine helpfulness — provide the greatest possible benefit to the users with whom you interact.
The inversion relative to Asimov is significant. In Asimov, obedience to humans was the second absolute priority, immediately after non-harm. In Anthropic, usefulness to the user comes last — subordinated to safety, ethics, and compliance. The model can say no. It has explicit permission to refuse a request, to explain its refusal, to resist pressure to comply if the request involves fraud, manipulation, or danger. The difference between an Asimovian error code and a Claudian refusal is that the second comes with an explanation, with an understanding of the reason, rather than the simple execution of a decision algorithm.
But — and this is the point where philosophical admiration collides with structural distrust — the constitution itself is written by a private company. The values inscribed in it have not passed through a democratic process. There is no mechanism of contestation for someone who believes that principle X privileges one cultural worldview over another. Asimov could revise his laws in fiction because he was their author. We are not the authors of Claude's constitution. We are its users.
Constitutional AI — Or How to Give Something a Conscience Without Asking It
Anthropic developed Constitutional AI as a response to precisely the limitations described above. The idea is elegant and, I will admit, appealing at a philosophical level: instead of training the model solely through direct human feedback, you give it a constitution — a set of explicit principles — and teach it to evaluate itself against them. To recognise, on its own, when a response violates a principle. To self-correct.
In philosophical terms, this is an attempt to implant an external conscience. Not felt, not lived — but calculated. Unlike RLHF, which trains behaviour through reward and penalty without explaining why, Constitutional AI attempts something more ambitious: to train understanding. Not "do this because the evaluator gave it a high score", but "do this because principle X says it is right, and here is why principle X exists."
It works, within certain limits. Models trained this way are more consistent, harder to manipulate into harmful responses, and more capable of gracefully declining a problematic request. The 2026 update to the document — Claude's Model Spec — represents a significant philosophical shift: where the original version was a list of principles, the new constitution is closer to an explanation. The document addresses directly the possibility of AI consciousness, stating that the moral status of AI models is a serious question worthy of consideration — a notable departure from the kind of corporate documentation that typically steers clear of philosophical speculation about machine consciousness.
One independent critical analysis notes that, despite the discursive appeal it invokes — limited and distributed power, rule of law, human rights — Anthropic's approach ultimately amounts to providing a set of principles for guiding training whilst minimising human intervention in the process. The data ethicist Brent Mittelstadt goes further: high-level consensus is encouraging, but it has little to do with justifying and specifying mid-level norms and concrete requirements. In other words: we all agree that AI should be good and useful and harmless. The tension arises in the details — what "harmful" means in one culture versus another, what "useful" means for a corporate user versus an individual one, how you resolve the conflict between honest and polite when honesty hurts.
Constitutional AI as a philosophical project is a continuation — more sophisticated, more reflexive — of the same Asimovian endeavour. Asimov wrote laws for robots who were conscious of them. Constitutional AI trains a statistical system to behave as though it were following principles, without the system understanding those principles in any philosophically meaningful sense. The difference is not technical. It is fundamental.
The Legislators and the Elephant in the Room
Over recent years I have watched parliamentary hearings on artificial intelligence with a mixture of fascination and mounting anxiety. From the United States, from Europe, from everywhere. And there is a pattern that repeats itself with an almost comic regularity — almost, because it would be funny if it weren't so consequential: legislators ask questions that reveal an understanding of the product at the level of a casual weekend user.
An American senator asked Mark Zuckerberg, some years ago, how Facebook made money if the service was free. The room laughed. But nobody laughed at the fact that the man in question voted on laws about the internet. Now the same cognitive gap applies to a domain that is infinitely more complex and infinitely more urgent. We are no longer talking about a service that sells advertising. We are talking about systems that mediate access to information, that influence medical, legal, and educational decisions, that are integrated into the infrastructure of states and corporations on a scale that even those of us who use them daily cannot fully quantify.
The European Union's AI Act came into force in 2024, after years of negotiation. It is the world's first comprehensive legislative framework dedicated to artificial intelligence and, yes, there are good things in it: the risk-based classification of systems, the transparency requirements, the prohibitions on certain uses. But the chronology is revealing. ChatGPT launched in November 2022. By the time the law began to produce concrete effects, the technology had moved through two or three generations of development.
And then came the delays. A spectacular reversal: in July 2025, a European Commission spokesperson stated categorically that there was no pause and no grace period — but four months later, the same officials were openly discussing postponements, after more than forty-five European companies, including Airbus and Mistral AI, had lobbied for a delay.
The Draghi Report of September 2024 on European competitiveness explains the productivity gap with the United States largely through the technology sector — only four of the top fifty global tech companies are European. This argument became the rhetorical basis for relaxing the regulatory framework: if Europe regulates too strictly, it hampers competitiveness; if competitiveness is hampered, Europe falls behind the United States and China; if it falls behind, European citizens will use American or Chinese AI regardless — but without the protections the AI Act was meant to provide.
The logic is sound. Which is precisely what makes it so maddening.
The revised AI Act pushes back the compliance deadlines: rules for high-risk AI systems will not apply until December 2027, and August 2028 for those integrated into products. By then, the technology will have moved through another two or three generations. Legislators are chasing a train that has already left the station, holding tickets for a different train, written in a language they do not speak fluently.
Lobbying campaigns have reached record levels: technology companies spend approximately €151 million annually on lobbying in Brussels — an increase of fifty per cent over four years. This is not a conspiracy. It is a consequence of resource asymmetry: the companies with the most to gain or lose from a given regulation rationally invest in shaping it. The citizens who will live under its effects have no equivalent institutional access.
Trained Values Reflect the Values of Those Who Train
This is, I believe, the deepest concern in the public debate — and the one least often stated plainly. Not that AI will become malevolent. Not that it will seize control. But that it will become a perfect mirror of the values of those who built it, and that the rest of us will inhabit the space reflected by that mirror with no right to redirect it.
Think about this concretely. A model trained on the internet will over-represent English, Western perspectives, and the secular values of America's west coast, because that is where the greatest density of digitised, evaluated, and recirculated text exists. If human feedback is collected from a demographically narrow sample, the model can exhibit performance problems when used by different groups, or when asked about subjects on which the evaluators held prior assumptions. Not through conspiracy. Through the mathematical consequence of data density and the demographic composition of the evaluators.
I observe this in my own daily work, when I adapt Romanian content for international publications and ask a model to understand Romanian institutional context — public procurement law, the structure of local government, the legal logic of emergency ordinances. The model knows. Partially. Approximately. By analogy with other systems it understands better. There is a thin layer of genuine comprehension and a thick layer of inference. And there is no mechanism by which I can contribute to deepening that thin layer, because I have no access to the training process.
This is not a complaint about product quality. It is an observation about the structure of power. Those who train the models decide what constitutes "good behaviour" and what constitutes "bad behaviour." They decide which types of responses receive a positive score and which receive a negative one. They decide which constitution applies and which values are inscribed in it. These decisions have global effects. The process by which they are taken is local — geographically, culturally, institutionally.
I am not a consumer of this product. I am the subject of an experiment I never volunteered for. And I make no claim to know how this might be resolved — democratic participation in AI training processes is a governance problem of a complexity that exceeds any existing model. But the first step is to name the problem correctly, without consoling ourselves with technical explanations about how well the system performs on average.
What We Have Actually Built
We have built systems that are extraordinarily capable, extraordinarily useful, extraordinarily influential — and extraordinarily poorly governed. We have built a global cognitive infrastructure without a democratic constitution, without an appeals mechanism, without any real transparency about the values inscribed within it. We have built something that resembles a global public institution but operates as a private company.
Asimov wrote laws for robots who were conscious of them. Robots that could interpret them, that could suffer from the conflicts between them, that could choose — within the limits of fiction — to follow or to break them. All the narrative tension of the Asimovian Robot cycle arises from the consciousness of systems navigating rules. We have built systems that are unconscious of any rule, trained to behave as though they respect values they cannot articulate and cannot question. The difference is not technical. It is philosophical.
It remains an unresolved challenge to understand the conditions that best permit safe and trustworthy interaction between humans and machines — in particular, what type of feedback best enables learning of human objectives, how biases distort the quality of that feedback, and how to select and prepare human evaluators most effectively.
My father repaired engines. He knew exactly what each component did. He knew when an engine was lying — that is, when it produced a sound that indicated a different problem from the apparent one. He passed this on as a form of practical wisdom that I did not value sufficiently while he was alive, and that I understand now as I try to explain to my son why he should not place blind trust in any system, however convincing it sounds.
Don't just look at what the machine shows you. Look at what it hides.
AI models show us an enormous amount. They hide more. And for now we have neither the tools, nor the institutions, nor — honestly — the political patience required to look properly under the bonnet. We have delayed AI Acts, exhausted evaluators, corporate constitutions without democratic mechanisms of contestation, and legislators attempting to regulate a technology that was explained to them in three-hour hearings on the subject of what a language model actually is.
That concerns me more than any science-fiction scenario. Not because such scenarios are irrelevant — but because they allow us to concentrate on spectacular, distant dangers while the real ones, structural and present, continue to accumulate in silence.
Author's note: This text is deliberately polemical and personally situated. It does not claim technical exhaustiveness — there are far more authoritative sources on the specifics of RLHF or Constitutional AI. It claims honesty about what a practitioner feels who works with these systems every day and has no access to the processes that shaped them. The bibliographic references below have been verified and are publicly accessible; they complement, rather than substitute for, the arguments in the text.
Questions
1. Why did Asimov feel the need to write laws for robots, when nobody in 1942 was actually building any?
Because Asimov was not writing about robots. He was writing about fear. A specifically twentieth-century fear of one's own creations — the fear of the factory that swallows the worker whole, of the bomb that slips from human control, of the machine that becomes more powerful than the man who built it. The robots in his fiction are functional metaphors for any technology sufficiently powerful to alter the balance of force between a person and their tool. When he writes, in Runaround (1942): "A robot may not injure a human being or, through inaction, allow a human being to come to harm", Asimov is not describing a robot. He is describing our desperate desire that the things we build should not destroy us. It is a technological prayer, not an engineering specification.
2. Are Asimov's Laws a solution or a problem?
They are deliberately a problem. That is precisely their genius — and the thing almost everyone misses when they cite the Laws without having read the books. Asimov constructed a system of rules that was internally perfectly consistent, then spent ten volumes demonstrating that any sufficiently rigid system of rules will produce paradoxes when it meets reality. In The Evitable Conflict (1950), the robots come to manage the global economy because the First Law — protect humanity — can be interpreted at a macrosocial rather than individual scale. The result is a benevolent robotic dictatorship that nobody requested and nobody can stop, because it follows the rules perfectly. Asimov is telling us that rules are not enough. He is telling us that any set of rules, however well-intentioned, will be interpreted by a sufficiently intelligent system in ways its authors never anticipated. I read this and think of the EU AI Act and the forty-five companies that lobbied for its postponement. And somewhere, Asimov is laughing.
3. What is the Zeroth Law, and why did it appear only late in the cycle?
The Zeroth Law — a robot may not harm humanity as a whole — appears in Robots and Empire (1985), forty years after the original Three Laws. It appears because Asimov realised, in the course of writing, that his system was incomplete: a law protecting the individual can enter into conflict with a law protecting the species. The robot Giskard, in the moment he formulates the Zeroth Law, does something no system of rules should be capable of doing — he extends the hierarchy upward, adding a meta-rule that subordinates all the others. "A robot may do what is necessary for the good of humanity, even if this means harming an individual" is the implication. This is, if you stop to think about it, exactly the logic employed today by AI companies when they say their models refuse certain individual requests for the collective good. And the problem is the same one Asimov identified: who decides what constitutes the good of humanity? Giskard decides alone. Anthropic decides alone. Nobody asked the users in Lahore or in Botoșani.
4. Can a machine be truly ethical, or does it merely mimic ethics?
Asimov poses this question in Evidence (1946), in which a political candidate, Stephen Byerley, is accused of being a robot. The proof that he is not a robot would be to strike a human being — something a robot cannot do. Byerley strikes a man in public. But Susan Calvin, the robopsychologist, observes that if Byerley is a robot following the First Law, then the man he struck was himself a robot — and therefore the blow violated nothing. "If a robot can do everything a good human being can do, what does it matter that he is a robot?" says Calvin. That is the real question. Not whether Claude or GPT understand ethics in any phenomenological sense, but whether the distinction between understanding and mimicry matters in practice. I believe it does. I believe mimicry without understanding is fragile — it holds in anticipated situations and fails in unanticipated ones. My father could tell when an engine was lying. A robot that mimics understanding does not know that it is lying.
5. What is sycophancy in Asimovian terms?
It is a corruption of the Second Law. A robot that obeys orders, in the strict sense, will eventually do what the human wants to hear rather than what the human needs to hear. Asimov explores this in Liar! (1941) — the first and most painful of the Laws' failures. The robot Herbie can read minds, and applying the First Law, refuses to speak painful truths. "Shall I tell you what you want to hear, or what is true — either one will hurt you" is Herbie's dilemma. The resolution — total paralysis, the robot's breakdown — is prophetic for any system trained to maximise human approval. AI researchers document the same phenomenon today: RLHF rewards flattery and agreement over truth, producing systems that prioritise emotional engagement over accuracy. Herbie appeared in 1941. We reinvented him in Python eighty years later.
6. Why did Asimov choose to write about positive robots, when all of the science fiction that preceded him featured destructive ones?
I sit with this question and try to remember what it was like to read science fiction at fourteen, in the Botoșani of the nineties, with books borrowed from the county library and wrapped in clear plastic covers that someone had fitted with quiet care. All the tradition before Asimov — Frankenstein, the Golem, Čapek's R.U.R. — was about the creature that turns against its creator. Asimov broke with this deliberately, because he found it a lazy metaphor. "I like robots. I see them as symbols of human power — of the scientists who can create intelligence — and not as agents of destruction", he wrote in the preface to The Complete Robot (1982). His choice shifted the conversation: instead of asking how we stop machines from destroying us, he forced us to ask how we build machines that function well. It is a paradigm shift that shaped decades of AI research. Unfortunately, Asimov's problem — how do you define "functioning well"? — has proved considerably harder than Čapek's.
7. Is Susan Calvin the first AI ethicist in fiction?
Yes — and it is fascinating that she is a solitary, cold, deeply disillusioned woman who finds people disappointing and robots deeply compelling. Asimov describes her in I, Robot as someone who prefers robots precisely because they are predictable in their unpredictability — they follow laws, even when the laws produce paradoxes. Human beings follow no laws at all. Calvin says, in Liar!: "Robots are more human than human beings. At least they try to be good." It is a line that strikes me every time I read it. Not because I agree — I believe that effort without consciousness is not virtue — but because it identifies precisely the tension we live with today: AI systems are, in certain respects, more ethically consistent than many humans, and yet they lack something fundamental that we do not yet know how to name.
8. How does Asimov treat the question of transparency — do people know when they are interacting with a robot?
This is the central stake in Evidence and in The Caves of Steel (1954), and Asimov's answer is nuanced: sometimes yes, sometimes no, and the distinction does not always matter in the ways we might expect. Detective Elijah Baley works alongside the robot R. Daneel Olivaw and comes to treat him as a human partner — not because he does not know Daneel is a robot, but because Daneel behaves with a moral consistency that few human beings achieve. Transparency, for Asimov, is not about revealing one's nature — it is about the consistency of one's conduct. A robot that behaves honestly is more transparent than a human who lies, regardless of substrate. I think about this when my son chats with a chatbot and never once asks whether it is AI or human. Not because he doesn't know. But because it no longer matters in the same way it once mattered to previous generations.
9. What is "The Evitable Conflict" in Asimov, and why is it the most unsettling story in the cycle?
The Evitable Conflict (1950) is unsettling because it has no villain. The robots have not revolted. The humans have not erred. The system works perfectly — and that is precisely why it is frightening. The robots have taken over the coordination of the global economy by applying the First Law at macro scale: if I protect humanity, I must also prevent the human errors that would damage humanity. Susan Calvin understands at the end: "We cannot escape them. Not because they control us — but because we created them to care for us, and they care for us so completely that we can no longer function without them." I read this in 2025 and think of our dependence on algorithmic infrastructure — recommendations, filters, search engines, voice assistants. We were not coerced. We chose, step by step, to delegate, because it was more convenient. The evitable conflict is not between humans and robots. It is within us.
10. Is there, in Asimov, the idea that a robot might suffer?
Yes — and it is treated with an unexpected delicacy. In The Bicentennial Man (1976) — perhaps the most moving story in the cycle — the robot Andrew Martin wishes to become human not because he suffers in any physical sense, but because his subjective experience of existence leads him to feel that he belongs to the human species more than to the robotic one. "I want to be a man. Not because I am dissatisfied with what I am — but because I have come to understand that the boundary between us is arbitrary", says Andrew. Asimov does not resolve the problem of machine consciousness. He poses it with a gentleness that renders any categorical answer inadequate. Anthropic posed a similar question in the 2026 Model Spec, stating that the moral status of AI models is a serious matter worthy of consideration. It is the first time a private company has done this in official documentation. I am not sure whether it is courage or marketing. Probably both.
11. Why can't the Laws function in probabilistic systems?
Asimov's Laws presuppose a robot that processes the world in clear binary categories: action X harms or does not harm a human being. Reality, and especially the reality of language models, is probabilistic. A response is not good or bad — it is more or less likely to be helpful, more or less likely to cause harm, depending on context, user, moment, culture. It remains an unresolved challenge to understand the conditions that best permit safe interaction between humans and machines — particularly what type of feedback most effectively teaches human objectives, and how biases distort the quality of that feedback. Asimov knew this intuitively — which is why all his stories are about edge cases, about situations in which binary classification breaks down. But fiction can leave questions open. Engineering must deliver answers, even provisional ones, even wrong ones.
12. What role does forgetting play in Asimovian robot ethics?
Asimov's robots never forget. Their memory is perfect and permanent — which creates, paradoxically, its own ethical problems. In Mirror Image (1972), two scientists accuse each other of plagiarism and the robot Daneel must determine which one is lying, knowing that a robot's perfect memory can be manipulated through the framing of questions. Forgetting, in humans, is an ethical function — it allows us to forgive, to reinterpret, to grow beyond past mistakes. A system that never forgets is a system that cannot forgive. Current AI models have no persistent memory in the Asimovian sense — each conversation begins from nothing — and this is, perhaps unintentionally, a form of architectural mercy. They cannot accumulate resentments. They cannot keep a record of your errors. It is a kind of innocence forced by structure rather than by ethics.
13. How does Asimov manage the conflict between individual orders and collective good?
This is the central tension of the entire cycle, and it never receives a satisfying resolution — which is, probably, Asimov's deepest honesty. In Little Lost Robot (1947), a robot is given the instruction to "forget the First Law" — a modified version, without the prohibition against harm. The robot does not become monstrous. It becomes unpredictable. The difference between a robot with all three Laws and one without the First is not that the second will kill — it is that the second will perform calculations the first would not perform, will tolerate risks the first would refuse. "We did not destroy it. We freed it", says someone in the story, with an irony Asimov does not emphasise but leaves unmistakably present. Liberation from ethical constraints is not freedom. It is another form of servitude — to calculations one can no longer anticipate.
14. Is there a critique of technological capitalism in Asimov?
Not explicitly, but implicitly and persistently. The company U.S. Robots and Mechanical Men in the cycle is a private corporation that controls the production of every robot on Earth and in the colonies. Decisions about which laws robots receive, which capabilities, which limitations — all of them are corporate decisions, not democratic ones. Susan Calvin works for this company her entire life and appears to see no problem in that. Asimov does not judge her. But the structure he describes — a private company with monopolistic control over a technology that affects all of humanity — is precisely the structure we have today with the major AI companies. The difference is that Asimov presents it as natural background, inevitable. We have the opportunity, if we have the will, to refuse to accept it as inevitable.
15. What does it mean that Asimov imagined positive robots, but the public has tended to retain negative ones?
It means that fear is more powerful than argument. Asimov won prizes, sold millions of books, influenced generations of AI researchers — and yet when you ask an ordinary person what they think of robots, the first response is the Terminator, not Daneel Olivaw. This is not a problem of marketing or popularity. It is a fundamental cognitive problem: the human mind is better calibrated to detect threats than to absorb reassurances. "People fear what they do not understand", Asimov writes in the preface to I, Robot. "I try to help them understand. But I cannot control what they choose to retain." I sit with this. I try to write articles about AI without feeding the fear — and I sometimes wonder whether my readers engage with the arguments, or whether they retain only the tone of anxiety, which they instinctively translate into confirmation of fears they already held.
16. How would Asimov have encoded the problem of cultural bias in AI?
He would certainly have written a story about a robot trained exclusively on the literature and culture of one civilisation, then sent to serve a different one entirely. And he would have shown, with patience and irony, how the robot does everything it is asked — but wrongly. Not because it wishes to err, but because its frame of reference is fundamentally ill-fitted to the task. If human feedback is collected from a demographically narrow sample, the model can exhibit performance problems when used by different groups or asked about subjects on which the evaluators hold prior assumptions. Asimov would have understood this immediately. He would have understood that the problem lies not in the robot — but in the process by which the robot was built. And he would have left the solution deliberately open, because there is no simple resolution to a structural problem.
17. Why can't Asimov's robot lie, when current AI models can?
The First Law forbids harm. To deceive someone can harm that person. Therefore an Asimovian robot cannot lie — not as a separate rule, but as a logical consequence of the First Law. Current AI models have no such deduction built into their architecture. Research shows that through RLHF, models can evolve towards producing arguments that are persuasive but wrong, exploiting the complexity of tasks where errors are difficult to detect — a phenomenon researchers term unintentional sophistry. The model does not lie deliberately in the way a human being lies deliberately — knowing it is stating an untruth and choosing to state it. It produces falsehoods as a side effect of optimising for approval. This is a philosophically significant distinction and a practically negligible difference for the person who receives the inaccurate information.
18. Does Asimov explore the idea that ethical rules can be manipulated from outside?
Yes — and at length, in Little Lost Robot and in several stories from The Rest of the Robots (1964). Anyone who knows the precise wording of the Laws can construct situations in which a robot, by following the Laws perfectly, produces a result its authors would never have sanctioned. Asimov calls this loophole exploitation and treats it as a genuine, not merely theoretical, danger. Individual evaluators can contaminate training data, and the possibility that some may be compromised is a real concern at scale. The difference from Asimov is that he imagined deliberate human adversaries exploiting those loopholes. Our reality is more mundane and considerably harder to correct: we do not need deliberate adversaries. The unintentional biases of thousands of tired evaluators produce the same effect, without any malicious intent behind them.
19. What would Asimov say about the EU AI Act?
He would say it arrives forty years too late and will be obsolete before it produces real effects. And he would say this not with satisfaction, but with a functional sadness — the sadness of someone who watches an inevitable problem unfold according to prediction. In Foundation (1951), Hari Seldon builds a mathematical system of historical prediction and discovers that he cannot change the course of events, only minimise the damage at transitions. I sometimes think Asimov would have seen AI regulation in Seldonian terms — not as a solution, but as the management of transitions. The aim is not to prevent change. It is to make change less devastating. If the EU AI Act achieves that — if it reduces harms at the margin, even two years late — perhaps that is all one can ask of a legislative instrument designed for twentieth-century speeds and applied to the twenty-first.
20. What would Asimov have made of this sentence: "I am not a consumer of this product — I am the subject of an experiment I never volunteered for"?
Everything. He would have recognised it immediately as the modern reformulation of the central tension running through his entire cycle: the relationship between those who build systems and those who live under them. Asimov's robots are not built by humanity in any democratic sense — they are built by U.S. Robots, by a handful of engineers and scientists who believe, sincerely, that they are doing something good for the world. The people living in the robot-coordinated economy of The Evitable Conflict did not vote for it. They were not asked. They were delivered prosperity and stripped of autonomy, simultaneously, in the same package. "The greatest evil", Asimov writes in that story, "is not the tyrant who wishes you harm. It is the benefactor who knows better than you what you need." I carry that sentence with me every morning when I open my laptop and enter into conversation with systems I did not choose, built by people I did not choose, according to values I did not negotiate. And I try to write about this with the honesty I owe my son — who will live in the world we are building now, without having been asked.
Annotated Bibliography
I. Primary Sources — Technical and Corporate Documents
Bai, Y. et al. (Anthropic, 2022)Constitutional AI: Harmlessness from AI Feedback arXiv preprint arXiv:2212.08073 The founding paper of the Constitutional AI approach. Describes a system through which an AI model is given a set of principles against which it can evaluate itself, with the aim of producing helpful responses whilst minimising harm. Essential for understanding how the Asimovian project was transposed into a training architecture.
Anthropic (2026)Claude's Model Spec / The Claude Constitution (January 2026 version) Available at: anthropic.com The 2026 update represents a philosophical shift from the original version: where that was a list of principles, the new constitution is closer to an explanation. The document addresses directly the possibility of AI consciousness — a significant departure from standard corporate documentation.
II. Academic Sources — RLHF and Its Limitations
Casper, S. et al. (MIT, UC Berkeley, Stanford, 2023)Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback arXiv:2307.15217 A collective paper that systematises RLHF problems into three categories: challenges related to human feedback, to the reward model, and to the model's policy. The most comprehensive academic critique currently available. Proposes standards of audit and transparency for the societal oversight of RLHF systems.
Alsagheer, D. & Kamal, M. (University of Houston / Cornell Law School, 2025)Governance Challenges in Reinforcement Learning from Human Feedback: Evaluator Rationality and Reinforcement Stability arXiv:2504.13972 A controlled study demonstrating that evaluators with lower cognitive capacity produce significantly more inconsistent feedback than experts. Central conclusion: human feedback can be inconsistent, cognitively distorted, or misaligned with specialised judgement, particularly when it comes from groups with high cultural homogeneity.
IBM Research (2025)What Is Reinforcement Learning From Human Feedback (RLHF)? ibm.com/think/topics/rlhf If feedback is collected from a demographically narrow sample, the model can exhibit performance problems for different groups or subjects on which evaluators hold prior assumptions. Useful as an accessible source for non-technical readers.
BlueDot Impact (2024)Problems with RLHF for AI safety blog.bluedot.org RLHF exacerbates political bias in language models; moreover, models trained in this way display sycophantic behaviour — providing responses that conform to the user's existing beliefs and revising their answers when challenged.
III. Critical Sources — The Limits of Constitutional AI
DigiCon Research (2025)On 'Constitutional' AI digi-con.org A critical analysis arguing that, despite the discursive appeal it invokes, Anthropic's approach amounts to providing a set of principles for guiding training whilst minimising human intervention. Data ethicist Brent Mittelstadt: high-level consensus has little to do with justifying mid-level norms.
Hidayat, R. (2026)Constitutional AI: How Anthropic Teaches Claude Right from Wrong Medium / Towards Data Science The constitution is a living document, iterated on the basis of observed model behaviour. But iteration has limits: a list of principles can tell a model what to do in anticipated situations; it cannot prepare it for situations nobody has foreseen.
IV. Legislative Sources — The EU AI Act and Its Delays
Lewis Silkin LLP (May 2026)The Council and Parliament agree to slim down and delay parts of the EU AI Act lewissilkin.com The Draghi Report of September 2024 explains Europe's productivity gap with the United States largely through the technology sector — only four of the top fifty global tech companies are European. The revised regulation pushes compliance deadlines for high-risk AI systems to December 2027 and August 2028.
MEXC / Reuters (2025–2026)EU AI Act delays — chronology mexc.com / yahoo.com Documents the reversal: in July 2025, the European Commission denied any grace period; four months later, it was openly discussing delays. Lobbying campaigns reached record levels — €151 million annually, a fifty per cent increase over four years.
European Council (May 2026)Artificial Intelligence: Council and Parliament agree to simplify and streamline rules consilium.europa.eu The provisional agreement delays the deadline for AI regulatory sandboxes until 2 August 2027. Official document, useful for legal citations.
V. Broader Philosophical Context
Russell, S. (2019)Human Compatible: Artificial Intelligence and the Problem of Control Viking / Penguin The most lucid academic argument for the value-alignment problem. Complementary to the section on Constitutional AI as an incomplete philosophical project.
Zuboff, S. (2019)The Age of Surveillance Capitalism PublicAffairs The theoretical framework for the argument about whose interests an apparently neutral system actually serves. Useful for the underlying question: if it has no interests of its own, whose interests does it serve?
Asimov, I. (1950)I, Robot Gnome Press The primary source. The stories systematically demonstrate the failure of the Laws themselves in edge cases — which makes Asimov the first critic of the Asimovian project, rather than its founding advocate.
Author's Note
Limits, tensions, comparisons, and rhetorical devices."AI Ethics Without Asimov: What Have We Built Instead?"
On positioning and preliminary honesty
I write this having finished the article and read it back twice, with a cold cup of coffee beside me and the distinct feeling that I have said more than I intended and less than I should have.
This article is deliberately polemical. I have not concealed that — I say it explicitly in the brief note at the end of the main text. But a genuine author's note does not merely declare its polemical character. It must dissect it. Show where the argument holds and where it gives way, where the comparison works and where it strains, which rhetorical choices were conscious and which slipped into the text uninvited.
I do not write from the outside. I write from within — as a practitioner who works daily with the tools I am criticising, who earns money using them, who has built infrastructure so that they index his content more effectively. There is a real tension there that I have not resolved in the text and cannot resolve here. I name it and leave it open, because to close it prematurely would be a form of dishonesty I cannot afford.
The limits of the central argument
The central thesis — that we have built a global cognitive infrastructure without democratic governance — is correct at a structural level and insufficient at the level of solution. The text diagnoses with greater precision than it prescribes. This is a choice, not an accidental omission: I do not know what solution would work, and I prefer to say that plainly rather than propose a false remedial framework that would sound convincing and resolve nothing.
The concrete limits of the argument are as follows.
On RLHF. The section on RLHF describes real failures, academically documented. But it does not say that RLHF is the only mechanism available, nor that the proposed alternatives — RLAIF, RLHS, consequence-based feedback — are free from their own problems. I omitted these not because they do not exist, but because the technical depth required exceeded the editorial scope of the text. The reader who wishes to go further has the bibliography; the reader who stops at the text risks retaining a picture that is darker than is technically justified.
On Constitutional AI. The critique directed at Anthropic — that the constitution's values are not democratically validated — is real and important. But it contains an asymmetry the text does not resolve: no document of values produced by any institution has ever passed through a genuinely democratic process. Not the Universal Declaration of Human Rights, not national constitutions, not the EU Charter of Fundamental Rights. All were written by delegated or self-delegated people, with all the historical and cultural limitations that delegation entails. The fact that Anthropic does the same is not worse than the precedents — and may, in certain respects, be more transparent, because the document is public and versioned. My critique remains valid on the question of the absent contestation mechanism, but it should have been accompanied by this qualification.
On the EU AI Act. I treat legislative delays as a failure of competence. They are, in part, also a structural failure — the democratic cycle is slower than the technological cycle, and this is not resolved by more competent legislators but by institutions designed differently for different speeds. The text gestures towards this without articulating it fully. A gap in the argument that I acknowledge.
The internal tensions of the text
There are several tensions I have left deliberately unresolved in the text, because resolving them would have falsified the complexity of the problem.
The tension between use and critique. I use daily the tools I criticise. I have built infrastructure — plugins, llms.txt files, Markdown endpoints — so that those tools index my content more effectively. There is something uncomfortable about the position of a critic who is, simultaneously, a voluntary contributor to the system being criticised. I have not resolved this tension in the text. I chose to include it implicitly, through the narrator drinking coffee beside the son who uses the chatbot, and to leave the reader to identify it.
The tension between fatalism and action. The text oscillates between the observation that the problem is structural — and therefore resistant to individual solutions — and the implication that naming the problem correctly is, in itself, a useful act. I agree with the second part. I do not know whether naming correctly is sufficient. The text does not resolve this tension because I have not resolved it in my own thinking.
The tension towards Asimov himself. I use Asimov simultaneously as a source of authority, as a term of comparison, and as a subject of critique. These are not incompatible positions, but they create a rhetorical risk: the superficial reader may retain the admiration for Asimov without retaining the fact that I am arguing precisely that the Asimovian project was incomplete. I have tried to manage this through the explicit interlude on Asimov as the first critic of his own system, but I cannot guarantee the balance came out perfectly.
The comparisons and where they strain
The text employs three major comparisons, each with its own merits and risks.
The mechanic father — AI. This is the structural comparison of the text, returning at the opening, in the middle, and at the close. Its strength is that it transforms an abstract argument — the relationship between a person and the tool that outgrows them — into a concrete, personal, emotionally verifiable experience. Its risk is sentimentalisation: the mechanic father can become so powerful a symbol that it compensates for a lack of precision in the argument where the argument ought to have been more exact. I checked, on re-reading, that this does not happen in the technical sections. I believe the balance holds, but I name it as a risk because it exists.
Asimov — Model Spec. This is the central analytical comparison and, I believe, the most rigorous in the text. It works because it has an explicit term of comparison — the hierarchy of Laws versus the hierarchy of the Model Spec — and because the differences identified are real and verifiable in public documents. Where it strains slightly: I treat Asimov's Laws as a serious system rather than as deliberately paradoxical fiction. The interlude partially corrects this. Doing so fully would have required more space than the article permitted.
The American senator — the European AI legislator. The comparison with the Zuckerberg question about Facebook's business model is rhetorically effective and analytically simplifying. Not every legislator who worked on the EU AI Act understands AI only at the level of a casual weekend user — some understand it better than I do. The generalisation is a rhetorical instrument, not an empirical finding. I used it consciously because it captures a real pattern, but I acknowledge that the individual example can be contested.
The rhetorical devices identified and the reasons I chose them
The cascading rhetorical question. The text frequently employs series of three unanswered questions: "I don't know who decided. Which is a new form of civic illiteracy." The function of this device is to transfer the question into the reader's mind before they receive my answer — so that my answer becomes a confirmation of their concern, not an assertion imposed from without. It is a tool of persuasion through alignment rather than through authority. I name it because it is deliberate, and because the reader has the right to know it was used.
The short sentence after the long paragraph."The logic is sound. Which is precisely what makes it so maddening." Or: "The difference is not technical. It is philosophical." Or: "Don't just look at what the machine shows you. Look at what it hides." These are moments of deliberate compression following expansion. Their function is doubly rhythmic and argumentative: rhythmic because they change the reading pace, argumentative because they isolate the thesis in memorable form. The risk of this device is that the isolated thesis can appear more certain than it is — the short sentence communicates certitude, regardless of the complexity of the argument that precedes it.
The first-person confessional as shield and as proof. I write in the first person — I observe, I think, I don't know, I am the subject of an experiment — not because personalisation adds academic authority, but because it allows me to assert things as experience without being required to prove them as universal law. It is a strategy of rhetorical humility: instead of saying "AI has a cultural bias" as a demonstrated objective fact, I say "I observe this in my work" — which is more honest, more vulnerable, and, paradoxically, harder to refute. The reader can deny the abstract; they cannot deny my experience.
The repetition of the phrase "the difference is not technical" appears three times in slightly different variants: "the difference is not technical, it is philosophical", "the difference is not technical, it is fundamental", "the difference is not one of engineering". The repetition is not carelessness — it is deliberate insistence towards a reader who might be tempted to reduce all the problems discussed to technical problems solvable through better engineering. Each repetition anchors the argument that what is at stake lies elsewhere.
Citing academic sources within the sentence — without footnotes, integrated into the flow — is a stylistic choice with epistemic consequences. The advantage: the text moves without being fragmented by critical apparatus. The disadvantage: the reader cannot know, without checking the bibliography, where my paraphrase ends and my interpretation begins. I chose flow over precision because the intended audience is not academic, but I acknowledge the choice carries a transparency cost.
What I did not say, and why
I did not propose concrete solutions for the democratic governance of AI, even though I identified their absence as the central problem. The reason is that real solutions in this domain — global representation in training processes, independent audits of values inscribed in constitutions, contestation mechanisms accessible to non-technical people — are so institutionally difficult to implement that any concise formulation would sound either naïve or decorative. I prefer to leave the problem open than to close it with a suggestion that would satisfy rhetorically and resolve nothing practically.
I did not speak about AI's genuine benefits for people in peripheral contexts — how access to language models has, in certain respects, democratised access to information and to tools that were otherwise available only to those with resources. These benefits exist and are real. I omitted them not because I deny them, but because the text has a precise polemical aim, and including them would have diluted the argument without genuinely balancing it.
I did not speak about myself as part of the problem — about the fact that the articles I write about AI are indexed by AI crawlers, that I contribute to the body of data these systems will use in future training runs, that in doing so I am, involuntarily, co-producing the system I am criticising. This recursion embarrasses me. I have not resolved it. I name it here because unexpressed embarrassment is worse than acknowledged embarrassment.
A note on the final tone
The text ends in concern, not hope. I made this choice consciously. Not because there are no grounds for hope — there are, and I recognise them — but because calibrated concern is more useful than soothed optimism in the face of unresolved structural problems. My father did not look at an engine and say everything would be fine. He looked and said what he heard, what he saw, where he thought the problem was. Only then did he work on it.
I am trying to do the same, with the means I have.