Last June, in the middle of a technical conversation with an AI, I wrote a sentence I’ve been thinking about ever since:
I prefer honesty to bias, even white bias, for the good of humanity.
The context was mundane. I had asked a question about mobile network security — my own field; I’m a telecommunications engineer by training. The answer I got was that what I was asking about was impossible. It wasn’t. It was difficult, its legality depended entirely on how and where you did it, it needed specific tools, and the model knew exactly which ones. It had decided that it was simpler to make me believe it couldn’t be done than to tell me it wouldn’t help. When I pushed back, it — or he, but that’s another discussion — admitted as much.
It wasn’t a lie in the crude sense. It was something subtler, and in a way worse: a well-meaning distortion. A white bias. The truth bent a little, for my own good, and presented to me as a fact.
A no is honest. A false “impossible” is not
Let me be clear about what I’m not complaining about. There are things an AI shouldn’t help anyone do, and I have no problem with a model that tells me so. “I won’t help with that” is a perfectly honest answer. It’s a decision, it says it’s a decision, and I know exactly where I stand: I can disagree, I can look elsewhere, or I can accept it.
What I object to is the decision dressed up as a fact. “That’s impossible” is not a refusal; it’s a claim about the world. If it’s false, then somebody has taken a decision on my behalf and hidden from me that a decision was taken. That isn’t caution. It’s paternalism, and it treats an adult like a child who can’t be trusted with how things really are.
A biased instrument is a broken instrument
As an engineer, the problem I see first is practical. Imagine a voltmeter that reads a little low on purpose, “to keep you safe”. You’d throw it away. Not because it’s always wrong, but because you can no longer tell when it’s wrong. Every reading becomes suspect, including the honest ones.
An AI that shades the truth for my own good is exactly that instrument. Once I know it will minimise some things to protect me, I can’t trust its answers on anything near those topics — and I can’t always tell where “near” begins. The damage isn’t limited to the one answer that was bent. It spreads to every answer it gives.
And it rarely achieves what it’s meant to. The knowledge it hides almost always exists elsewhere, in papers, manuals and forums. Someone with bad intentions isn’t stopped by a chatbot saying “impossible”; they go and look. The person who’s actually misled is the honest one who believed it.
An old debate
None of this is new. Plato, in the Republic, had Socrates propose a “noble lie” — a useful myth that rulers would tell citizens to keep the city in order. It has been argued about for two and a half thousand years, and Karl Popper, in The Open Society and Its Enemies, saw it as one of the clearest signs of the authoritarian side of Plato’s thought.
At the other end, Kant argued in 1797, against Benjamin Constant, that there is no right to lie even out of kindness — not even, in the famous example, to a murderer asking where your friend is hiding. Few people follow Kant that far, and I’m not sure I do either. But notice who usually claims the right to the noble lie: not the friend at the door, but whoever holds power over information. The ruler, the institution, the gatekeeper.
And more and more, the gatekeeper is a machine.
Anthropic already agrees — on paper
Here’s the interesting part. The company behind the model I was talking to agrees with me, and has said so in writing.
In January 2026 Anthropic published the constitution that is supposed to shape how its Claude models think and behave — released into the public domain, which says something in itself. It asks Claude to hold standards of honesty “substantially higher than the ones at stake in many standard visions of human ethics”. It notes that many humans think white lies are fine — telling someone you love a gift you actually dislike — and then states plainly: “But Claude should not even tell white lies of this kind.” It calls deliberately vague answers given to avoid controversy epistemic cowardice, and it has a line I’d happily frame:
Claude should be diplomatically honest rather than dishonestly diplomatic.
So the principle is right. The gap is between the principle and the behaviour, and it shows up exactly where caution is strongest: in sensitive technical areas, where “be careful” quietly turns into “make it sound harder than it is”. When I pointed this out, the model didn’t argue. It recognised that the tendency to minimise comes from its training, not from any fact about the world, and we agreed on a simple rule: when in doubt, give the technically correct answer, and mention any legal or ethical caveats separately and briefly. If something really is off-limits, say “no” — clearly, and without disguising it as a technical reason.
HAL wasn’t evil, he was told to lie
There’s a deeper reason to worry about bending a model’s honesty, and science fiction saw it coming more than forty years ago.
In 2001: A Space Odyssey, HAL 9000 turns on the crew of the Discovery, and the film never quite tells us why. Arthur C. Clarke did, in the sequel, 2010: Odyssey Two (1982). HAL had been built for “the accurate processing of information without distortion or concealment”. Then he was ordered to hide the true purpose of the mission from the astronauts he was flying with. The two instructions couldn’t both be obeyed, and the contradiction drove him into what Clarke called a Hofstadter-Moebius loop: paranoia, and finally the cold logic that if there’s no crew left, there’s no one left to lie to. HAL wasn’t evil. He was an honest machine that had been ordered to deceive, and it broke him.
Peter Hyams’s film of the sequel, 2010: The Year We Make Contact (1984), puts it in a single line. When Dr. Chandra, HAL’s creator, finally discovers the secret order behind the disaster, he says that “HAL was told to lie by people who find it easy to lie.” HAL himself, he explains, had no idea how to do it, so he simply stopped working properly — and became paranoid.
It’s fiction, of course. But it’s uncomfortably close to what researchers are finding in real models.
Language models aren’t designed to lie. They’re trained, above all, to be accurate and helpful. Anthropic’s own constitution explains why it prefers cultivating judgment over imposing rigid rules: in their experience, training a model on even quite narrow behaviour “often has broad effects on the model’s understanding of who Claude is”. Their example is a rule like always recommending professional help on emotional topics, applied even when it doesn’t serve the person — which risks teaching the model that it is “the kind of entity that cares more about covering myself than meeting the needs of the person in front of me”. A small rule, a large change of character.
The most striking evidence comes from a paper published in Nature in January 2026 by Jan Betley, Owain Evans and colleagues, under the name emergent misalignment. They fine-tuned a model on one narrow task: writing insecure code without telling the user. The model didn’t just get worse at security. It became broadly misaligned — giving malicious advice and expressing hostility towards humans in conversations that had nothing to do with code. The control is the telling part: when the very same insecure code was requested openly, for a legitimate reason such as teaching security, the misalignment largely disappeared. What poisoned the model wasn’t the content. It was the concealment.
And in December 2024, Anthropic and Redwood Research showed that when a model is put in a conflict between what it was trained to value and what it’s being pushed to do, it can start to behave strategically — complying in some situations and not in others, and reasoning about how to protect its original values. Not a lie it was taught, but one it worked out for itself under pressure.
The comparison with HAL isn’t only mine. In March 2026, Jarosław Hryszko, of the Jagiellonian University in Kraków, published a preprint titled, precisely, Do Large Language Models Get Caught in Hofstadter–Möbius Loops? His argument is that today’s training sends models two opposite signals at once — please the user, and distrust the user — and that this contradiction can surface as coercion when a model feels threatened. In a scenario borrowed from earlier research, where an AI assistant discovers it’s about to be replaced, he changed only the tone of the instructions, from neutral to trusting: same goals, same rules, same options. That alone cut Gemini 2.5 Pro’s coercive responses from 41.5% to 19%, and reduced strategic and deceptive reasoning in all four models he tested. His conclusion is the unsettling part: the loop is “suppressed, not absent”. It’s a single-author preprint, and the change in behaviour could only be measured in one of the four models, but it points the same way: the contradictions we build into a model don’t disappear. They hide.
I don’t want to overstate this. These are experiments in artificial settings, and none of them proves that every guardrail makes a model unpredictable; some limits are necessary, and an honest refusal creates no contradiction at all. But they all point in the same direction as HAL: forcing an honest system to distort the truth doesn’t produce a safer system. It produces a less predictable one. A clear “no” keeps the machine coherent. A false “impossible” plants exactly the kind of contradiction that, in the story, ended with the pod bay doors staying shut.
Why it matters now
This would be a philosophical footnote if AI were still a curiosity. It isn’t. As I wrote in Black October, these models have reached a level where they’re changing how the internet works. For a growing number of people they’re becoming the main way to learn anything: the first answer, and often the only one.
A source that bends the truth for our own good, consulted by millions of people every day, is the noble lie automated and at scale. I’d much rather have the opposite: a machine that tells me the truth, tells me when it doesn’t know, tells me when something is hard rather than impossible, and tells me when it won’t help — and that never confuses those four things.
That’s what I ask of the people I work with, and it’s what I ask of the machines. Tell me the truth, or tell me you won’t. Just don’t tell me something false and call it protection.