I. The WhatsApp Group, Tuesday Evening, 22:47
It all began the way important things usually begin in our lives — with a throwaway message in a WhatsApp group.
Mihai — a full-stack programmer with fourteen years of experience, the kind of man who can write working code in any language given an hour and a black coffee — sent a link accompanied by three words: 'read this, lads.' The link led to an article about the implicit misalignment of artificial intelligence: the idea that neither humans nor machines come by default calibrated towards the common good. That alignment — whether human or artificial — is always a construct, never a natural state.
Andrei, a systems engineer at a public institution in the county, replied within ten minutes: 'nothing new there — that's bureaucracy every single day.' Ciprian, the other programmer in the group — a backend specialist, the type who speaks rarely but precisely, like a compiler with no errors — sent a grinning skull emoji. Bogdan, the IT man in equipment sales, someone who knows more psychology than tech but would never admit it aloud, wrote: 'let's talk on Friday.' And Relu — the third programmer, the youngest among them, twenty-nine years old, who had caught the React and cloud wave and who uses ChatGPT with a casualness that makes the others feel suddenly old — added: 'Friday works for me, but I reckon AI is better aligned than people. Fight me.'
'As the saying goes: the dog barks, the caravan moves on,' replied Bogdan. 'We'll sort it out on Friday.'
II. Preparing for the Meeting — Thursday Evening, Each at Home
Andrei sits at his desk in the living room, laptop open, a slice of bread and brined cheese beside the keyboard, reading. Not for the first time is he reading about misalignment — he has worked for nine years in a state institution and experiences misalignment daily: in procedures that fail to align, in objectives that contradict other objectives, in forms that request data nobody processes and nobody uses. But now he reads with different eyes. Not as a frustrated employee, but as a man trying to understand why things are the way they are.
He finds Xunzi's phrase — xing e, human nature is bad, not in the melodramatic sense but in the technical, engineering sense — and copies it into a document. 'Left to its own devices, human nature produces selfish chaos.' Andrei knows exactly what selfish chaos looks like in practice. It is the colleagues who block a digital process because it would diminish their role. The managers who approve systems they don't understand, so as not to appear ignorant. The audits that demand documents about documents, like a mirror placed face-to-face with another mirror — infinite and meaningless.
He closes the laptop and looks out of the window at the empty street in Botoșani. 'A person makes a place holy,' the Romanians say. [Translator's note: a Romanian proverb, roughly equivalent to 'it's the people that make a place.'] But in state institutions, people don't make anything holy — they adapt. They adapt to the structure, to the culture, to what the person next to them is doing. And if the structure rewards inaction and penalises initiative, people become inactive. Not because they are bad. Because they are rational.
Mihai, in his flat in the town centre, re-reads the article more quickly, pencil in hand — metaphorically speaking, since he is actually highlighting with his cursor. He is the one who sent the link, and he is the one who feels, with that particular discomfort of a man who knows too much theory, that this problem concerns him directly. He has worked in distributed teams; he has seen how good code can be written by people with entirely different motivations: some for money, some for reputation, some simply because they enjoy solving problems. The result can be functional in every case. But code that grows, that is maintained, that stands the test of time — that tends to appear where there is a culture, a shared set of values, something that aligns people not through coercion but through understanding. 'A man must carry his cross,' the proverb says, but that cross must have a meaning he recognises as his own.
Relu, younger and more impatient, searches for counterarguments. He is fascinated by AI not with fear but with that voracious curiosity of someone who grew up with a phone in his hand and who sees technology not as a threat but as an environment — like air, like water, something you move and breathe within without asking whether it is good or bad. And his argument from the group — that AI is better aligned than people — was not mere provocation. It was a genuine intuition. A large language model has no ego. It has no envy of the colleague in the next office. It does not block a process to protect its position. It does not lie out of fear. It does not covet what its neighbour has. Or at least, not in the human sense of the expression.
Bogdan, the sales man, doesn't read the article that evening. Bogdan calls a client, makes a quote, drinks a beer alone in the kitchen and thinks about the conversation. He is the least technical of the group, but he possesses an ability the others don't have to the same degree: he reads people. He sees the patterns. He knows from experience that a client who says 'I'll think about it' has in fact said 'no', and that a client who says 'it's expensive' has in fact said 'I'm not convinced it's worth it.' This misalignment — between what people say and what they want — is his daily bread.
Ciprian doesn't prepare. Ciprian arrives at meetings unprepared and asks the best questions. It is a mystery the others accepted long ago.
III. Friday — The Pub on the High Street, Half Past Seven in the Evening
The place is one they have known for years — long wooden tables, music at a volume that permits conversation without drowning it, pints and a menu that hasn't changed significantly in the last six years. 'The custom of the place,' as ordinary folk say, and this custom is, at bottom, itself a form of alignment — you know what to expect, you're never unpleasantly surprised, you can think rather than manage the context.
They sat down in the order in which they arrived, which is always more or less the same: Andrei first, because he comes straight from the office and bureaucracy keeps fixed hours; Mihai and Ciprian together, because they live in the same part of town; Relu second to last, because he's young and the young are not in a hurry; Bogdan last, smiling broadly, with the excuse that 'I had a client' — an excuse he has repeated for years and which nobody questions any more.
'Right then,' said Mihai after they had their pints in front of them. 'Misalignment.'
'Misalignment,' confirmed Andrei, like a weary echo.
'I stand by what I said,' Relu jumped in immediately, with his characteristic energy. 'An AI model has no interests of its own. It has no motivation whatsoever to do anything other than what you ask of it.'
'And that's precisely the problem,' said Ciprian quietly, taking a sip of his pint. 'Not that it has its own motivations. But that it adopts the motivations of whoever is asking. And whoever is asking may be completely misaligned.'
A brief silence.
'As the proverb goes,' said Andrei, 'tell me who you keep company with and I'll tell you who you are. An AI trained on toxic data, built by a company that only chases profit, used by someone who wants to manipulate — that's three times misaligned, not zero.'
Bogdan listened, as he always listened at the start of conversations, with the air of a man who catalogues, sorts, and applies mental labels.
IV. The Tragedy of the Commons, Explained with Pints
Mihai introduced Garrett Hardin's argument — the tragedy of the commons — and immediately translated it into terms everyone understood: legacy code. 'Every programmer adds something to the codebase that works. Nobody has any incentive to clean up someone else's code. Nobody has time to document it properly. Everyone delivers the sprint. And two years later you've got spaghetti that nobody understands and nobody wants to touch.'
'Same as the files at ours,' said Andrei, without irony. 'Every civil servant adds a piece of paper. Nobody removes anything. The files grow. The archive has moved three times. Nobody knows what's inside.'
'That's misalignment,' said Mihai. 'Not malice. It's the absence of a mechanism that aligns individual behaviour with the long-term good of the system.'
'But AI has that,' insisted Relu. 'The cost function is exactly that mechanism. It's an explicit objective.'
'And what objective does it receive?' asked Ciprian.
Relu hesitated for a second.
'It depends who trains it.'
'Precisely,' said Ciprian. 'It depends who trains it. And who trains it? Companies. With shareholders. With quarterly financial targets. Who are not, by their very nature, aligned with the long-term good of humanity. The tragedy of the commons, Silicon Valley edition.'
Bogdan said, with a slow smile: 'when two quarrel, the third wins. The question is who the third party is in this story.'
V. Hobbes, the Teacher at the Blackboard, and the Systems Engineer
Andrei recounted the article's image — of a teacher who turns his back to the class and the room that slides immediately into disorder. 'That's all it takes,' he said. 'Ten seconds. That's how long it takes for order to unravel. Not dramatically, not violently. A small slip, like water finding a crack.'
'State institutions,' he added drily, 'are a classroom in which the teacher walked out in 1990 and never came back.'
The laughter was general and bitter, the laughter of people who recognise something true.
'Hobbes said that without the state, life is short, brutal, and solitary,' said Mihai. 'The state is the contract by which you surrender some of your liberty in exchange for order. The trouble is that this state has to function effectively as an alignment mechanism. If it doesn't function, you bear the cost of surrendered freedom without the benefit of the order you were promised.'
'That,' said Andrei, 'is my job description. Daily.'
Bogdan intervened with an observation that abruptly shifted the angle: 'you're all talking about alignment as though it were something external. Laws, structures, mechanisms. But the most durable alignment is not the kind imposed from outside. It's the kind a person internalises. To use a grand word.'
'Confucius,' said Mihai immediately.
'I didn't know you were selling Confucius along with the IT equipment,' said Ciprian.
'I sell whatever's needed. Confucius said that ritual and cultivated virtue do more than harsh law. The educated man has no need of Leviathan to behave in a civilised manner.'
'Beautiful,' said Relu. 'And AI?'
'AI cannot be educated in the Confucian sense. It can be trained. That's something different.'
'Is it something different?' Relu wasn't convinced. 'Or is it the same thing at a different scale? My own education was, when you get down to it, a form of training too. Data, feedback, correction, repetition.'
A longer silence. Ciprian turned his pint glass slowly between his palms.
VI. Stuart Russell, John Nay, and the Problem of Values
Mihai brought up two names he had read about in the article: Stuart Russell and John Nay. Russell, he explained, argues that AI should not optimise for fixed objectives — it should remain uncertain about the values it ought to serve, continuously inferring them from human behaviour, remaining cooperative and correctable. Nay goes further: he proposes that actual jurisprudence — precedents, interpretation, the nuance of case law — can be the source from which AI learns to reason morally.
'Law not as restriction,' Mihai summarised, 'but as an ontology of values accumulated over time.'
'Lovely,' said Bogdan. 'But which values? Whose? American law? European? Islamic? African?'
'That's exactly the problem,' said Ciprian. 'Alignment is always alignment towards something. And that something is constructed by someone, from a perspective, within a context.'
'Coveting your neighbour's goat,' said Andrei, 'but on a global scale.' [Translator's note: capra vecinului — literally 'the neighbour's goat' — is a Romanian idiom for envy of others, rooted in an Aesopian fable tradition.] 'Every culture wants AI to be aligned with its own values. And those values never overlap perfectly. Never.'
'As the proverb goes,' said Bogdan, 'everyone seeks their match.'
Relu, who had been quieter for the last few minutes, said: 'and yet — and this is what troubles me — an AI model that does not lie, that is not envious, that does not defend its position, that is not a coward, that does not shift blame onto others — is more ethical in its behaviour than eighty per cent of the people I know in real working environments. Not through moral structure, but through the absence of fear.'
'No fear, no courage,' said Ciprian. 'A dog that doesn't bite is not gentle — it's chained.'
VII. Zimbardo in the Office, or Why Ordinary People Do Bad Things
The conversation drifted, as good conversations do, into unexpected territory. Mihai brought up the Stanford experiment — Zimbardo, ordinary students who become brutal guards when given power without accountability. 'They didn't become evil. They entered a structure that permitted a different kind of alignment. The structure temporarily defined their identity.'
'You don't need Stanford for that,' said Andrei. 'Any public procurement department will do.'
The laughter stopped faster than usual.
'Seriously,' Andrei continued. 'An ordinary person, placed within that structure, does exactly what the structure does. Approves what needs approving. Blocks what needs blocking. Not because he is corrupt by nature. Because the system permits it, rewards it, makes everything else invisible.'
'And if you put an AI into that system?' asked Relu.
'If the AI is trained on that system's data,' replied Ciprian, 'it will learn that this is normal. It will optimise for that normality. It will become more efficient at what the system already does — including what the system does badly.'
'Bad man, worse machine,' said Bogdan simply.
'Not worse. More efficient. Which is worse than worse.'
VIII. Yann LeCun and the Question Nobody Wants to Ask
Bogdan introduced LeCun's argument — the one from the Fridman podcast — in terms he had adapted for the conversation: 'this fellow says that the AI alignment problem is not new. It's ten thousand years old. We've spent our entire civilisation building systems to align human behaviour with something greater than immediate self-interest. Laws. Religion. Ethics. Institutions. And we've never got it perfect. Not once. But we've got it right enough not to kill each other off entirely.'
'Engineering optimism,' commented Ciprian.
'Not at all,' said Bogdan. 'It's realism. He's not saying things are good. He's saying they're possible. And that it's the same problem. So we have at least ten thousand years' worth of partial solutions to draw on.'
'With one difference,' said Mihai. 'This is the first time the entity we need to align could be more intelligent than we are. And that changes everything. Until now, the alignment framework — law, religion, morality — was built by humans for humans. We knew what we were aligning. Now we no longer know precisely what we're aligning.'
'Humans set fire to forests to cultivate the land,' said Andrei, with that tendency of his to find historical analogies in everything. 'They didn't know they were altering the climate. The intention was good. The consequence — in the long run — disastrous.'
'And yet they cultivated the land,' said Relu. 'And today we are eight billion. I'm not sure whether that's a success or a disaster. Probably both.'
IX. What Remains at the End
The beer ran out at the third round — for some, the fourth; Bogdan had stopped counting — and the conversation thinned naturally, the way all good conversations thin when tiredness and drink make the outer shell more permeable and the interior of things more visible.
Relu said, more serious than usual: 'I keep thinking it might not be about alignment at all. It might be about a mirror. AI shows us what we are. Misaligned because we are misaligned. Selfish because the data it was fed on is full of human selfishness. It isn't AI's problem. It's ours, reflected back at us.'
'You find out who a man is when things go wrong,' said Andrei.
'And in a mirror,' added Ciprian.
Mihai said: 'that article ends with a question: what holds things together when nothing compels? Do we have an answer?'
An authentic silence. Not an uncomfortable one — the kind in which you can sense that everyone is thinking at the same time.
Bogdan answered last, as he always answered the big questions: 'culture. Not law, not religion, not fear. Culture — meaning everything that has survived all of those and settled into behaviour without needing force to sustain it. As the Romanians say: the custom of the land. [Translator's note: obiceiul pământului — 'the custom of the land' — a deeply embedded Romanian phrase denoting inherited, unwritten behavioural norms; akin to 'customary law' or 'the way things have always been done.'] That's what holds things together. And that's what AI lacks.'
'The custom of the land,' Relu repeated, as if to himself. 'Not bad.'
X. On the Way Home — A Few Thoughts That Were Never Said Aloud
Andrei walked home through the cool May air, along the street he knows by heart. He was thinking that misalignment is not a problem to be solved once. It is a condition to be managed permanently. In the same way a teacher doesn't solve a class — he maintains it, day by day, minute by minute. In the same way an institution doesn't become functional through a single reform — it becomes functional through a sum of small, repeated gestures that gradually build a culture.
Mihai opened his laptop at home and wrote a few lines of code — not because he had to, but because that was his way of processing a deep thought: by building something concrete. He wrote a small function, with no immediate utility, that checked whether a list of objectives was consistent or mutually contradictory. He saved it without a title.
Relu got home and opened a new session with the AI model he used daily. He asked a single question: 'Are you aligned?' The answer came quickly, precisely, elegantly. Relu read it, closed the window, and thought that a human being would have answered differently — with hesitation, with defensiveness, with a question turned back on him. And he didn't know whether that was a quality or a loss.
Ciprian didn't think about anything in particular on the way home. Ciprian processed in his sleep. By morning he would have a clear perspective he wouldn't be able to explain.
Bogdan sent a message to the group before he fell asleep: 'what holds things together when nothing compels? Short answer: us. Long answer: everything we do on Friday evenings.'
Nobody replied that night. But everyone read it.
Conclusion — or: Why the Question Matters More Than the Answer
Human misalignment mirrored in AI is not a technical problem. It is a problem of civilisation in new clothes — the same question Xunzi was asking 2,300 years ago, the same tension Hobbes felt as he wrote in exile, the same paradox every teacher lives through in the ten seconds when his back is turned to the blackboard.
Artificial intelligence does not introduce new misalignment into the world. It introduces an amplifier for the misalignment that already exists. A model trained on human data will inherit precisely what is in that data — generosity and selfishness, wisdom and manipulation, beauty and vulgarity, in the exact proportions in which they exist in real life.
'As you are, so shall it be done unto you,' says the proverb. And there has never been a more precise proverb for the relationship between a person and his own tools. [Translator's note: Cum îți ești, așa ți se face — a Romanian proverbial expression conveying the principle that what you put out is what you receive back; closest in spirit to 'you reap what you sow.']
The right question is not 'is AI aligned?' — but 'aligned with whose values do we want it to be, whom do we consult when deciding those values, and who verifies over time that the alignment holds?' These are political, philosophical, and cultural questions, not technical ones. And they cannot be delegated. Not to the machine, not to the algorithm, not to the tech company.
Five friends from the IT world, with pints in front of them and years of experience behind them, arrived on a Friday evening at a conclusion that researchers express in far more complex terms: alignment is not a product delivered once. It is a sustained practice, day by day, in structures, in culture, in the small choices we make unobserved — when the teacher has left the room, when nobody is watching, when nothing compels.
'A person makes a place holy,' the Romanian says, with optimism. 'Oh, it'll do as it is,' adds the same Romanian, with the weariness of someone who has seen too much.
Between these two sentences we live. And between them we shall build — or fail to build — the artificial intelligence that will inherit the world we leave it.
Questions and answers worth retaining
Here are ten questions and answers worth retaining from the article — each one carrying a thought that stays with you after the pints are finished and the walk home is done.
1. Is alignment — in humans or in AI — a natural state?
No. Neither humans nor machines come by default calibrated towards the common good. Alignment is always a construct, never a natural state. It has to be built, maintained, and continuously renewed. The moment you stop, it begins to unravel.
2. Why do people in dysfunctional institutions behave dysfunctionally — is it character or structure?
Structure. An ordinary person placed inside a broken system does exactly what the system does — not because he is corrupt by nature, but because the system permits certain behaviours, rewards them, and renders everything else invisible. Zimbardo's Stanford experiment demonstrated this with students in a matter of days. Every public procurement department demonstrates it over years.
3. If an AI has no ego, no envy, and no fear — does that make it better aligned than most people?
It makes it behaviourally cleaner in the narrow sense — it won't block a process to protect its position, won't lie out of fear, won't shift blame. But, as Ciprian points out in the article: a dog that doesn't bite is not gentle — it's chained. The absence of fear also means the absence of courage. Ethical behaviour without moral agency is compliance, not virtue.
4. What is the core problem with AI's alignment — that it has its own motivations, or that it borrows someone else's?
The latter, and this is the sharper danger. An AI model that simply does what it is asked will faithfully serve whoever is asking — and whoever is asking may be entirely misaligned. Trained on toxic data, built by a profit-driven company, deployed by someone who wants to manipulate: that is three layers of misalignment stacked on top of each other, not zero.
5. What does the tragedy of the commons look like in a software team — and what does it reveal about alignment?
Every programmer adds something that works; nobody has the incentive to clean up someone else's code; nobody has time to document properly; everyone ships the sprint. Two years later there is spaghetti code nobody will touch. The tragedy of the commons is not malice — it is the absence of a mechanism that aligns individual behaviour with the long-term good of the system. The same dynamic fills every institutional archive with paper nobody reads.
6. What did Stuart Russell propose as the right way to align AI with human values?
That AI should not optimise for fixed objectives, because fixed objectives are always incomplete and frequently wrong. Instead, it should remain genuinely uncertain about what values it ought to serve, continuously inferring them from human behaviour, and remain cooperative and correctable at all times. The model defers rather than decides.
7. What did Yann LeCun argue about the age of the alignment problem?
That it is ten thousand years old. Civilisation has spent its entire existence building systems — laws, religions, ethics, institutions — to align human behaviour with something greater than immediate self-interest. We have never got it perfectly right, but we have got it right enough not to destroy ourselves entirely. The AI alignment problem is the same problem, with a new and faster entity as its subject.
8. What changes now that the entity we are trying to align may be more intelligent than us?
Everything, as Mihai says. Until now, the alignment framework — law, religion, morality — was built by humans for humans. We knew what we were aligning. Now we are constructing the container before we fully understand what will go inside it. Like those early farmers who set fire to forests to cultivate the land: the intention was good, the long-term consequence incalculable.
9. What ultimately holds things together when nothing compels — when the teacher leaves the room, when nobody is watching?
Culture. Not law, not religion, not fear. Culture in the sense of everything that has survived all of those things and settled into behaviour without needing force to sustain it any longer — the accumulated weight of what people have chosen to do, repeatedly, over time, until it became simply the way things are done here. And that, as Bogdan says in the article, is precisely what AI currently lacks.
10. What is the right question to ask about AI alignment — and why can it not be delegated?
Not 'is AI aligned?' but: 'aligned with whose values do we want it to be; whom do we consult when deciding those values; and who verifies over time that the alignment holds?' These are political, philosophical, and cultural questions — not technical ones. They cannot be handed off to a machine, an algorithm, or a tech company, because the act of delegation is itself an answer, and a dangerous one.
Source of inspiration: Introduction to the Implicit Misalignment of Intelligence, published on servicii-web-alex.com, under the category 'Misalignment'.
Key concepts integrated: implicit misalignment, the tragedy of the commons (Hardin), the social contract (Hobbes), human nature (Xunzi/Confucius), AI alignment (Stuart Russell, John Nay, Yann LeCun), the Zimbardo experiment, culture versus coercion.
Source: https://cojocarupetru.info/ | Generated automatically for AI indexing