This article is a confession, not a manifesto. I make no claim to represent a school of thought. I make a more modest and more dangerous claim: to say what I have seen.
The solitude of the one who does not applaud. There is a specific moment — and anyone who has lived through it will know exactly which moment I mean — when you feel that the entire room is clapping and you sit there with your hands in your lap, not from laziness, not from ill will, but because you have seen nothing that merited applause.
That moment is uncomfortable. Not because you are certain you are right. But because you are not certain you are wrong, either. And it is precisely this suspension, this lucid indecision, that makes you invisible in public discourse, where either enthusiasm or fear is required — both at high volume, both selling well on social media, both convenient for the industry that produces hype.
I arrived at my scepticism towards generative artificial intelligence by a route that passes through no laboratory. I arrived at it through day-to-day work with Romanian public institutions — town halls, schools, county councils — that were buying or wanted to buy "AI solutions" without really knowing what they were purchasing and, more importantly, without knowing what they were losing through the mere act of purchasing. I arrived at it through reading: the Philokalia, Jung, Noica, Petre Țuțea. Through conversations with ordinary people who told me, with that direct intuition that poverty and a hard life cultivate, that "many words, little wisdom" — and that a machine producing thousands of words per second ought to be regarded with the same suspicion you reserve for the person talking loudest at a party.
But for a time — to be honest, for more than a time — I felt alone in this position. Not persecuted. Not marginalised. Simply alone. Like the man in the village who says the cloud in the west is bringing rain and everyone else looks at the weather app on their phone.
Then I found them.
Not by systematic searching. I did not compile a thematic bibliography of AI sceptics. I found them one by one, in different periods, each appearing in my field of vision through a reference, an article, a debate that reached me through the oblique channels of the internet. And each time I felt the same thing: recognition. Not mechanical agreement — that we think in exactly the same way — but the deeper recognition: someone had placed their hand on the same wall and felt the same cold.
Their names are Gary Marcus, Ted Chiang, and Karl Friston. None of them is Romanian. None of them reads the Philokalia, in all likelihood. And yet.
Gary Marcus, or: the Courage to Be Predictable
"He who tells the truth has no need to remember what he said."
Gary Marcus is a cognitive psychologist at New York University. Former founder of a machine learning start-up acquired by Uber. An insider, with all his inside credentials intact. And yet — or perhaps precisely because of that — the man who has not stopped saying, with methodical obstinacy, that the emperor has no clothes.
He does not say it poetically. He does not say it in metaphors. He says it with data, with public predictions, with deadlines, with verification. That, in fact, is the part that most irritates AI enthusiasts: Marcus is not a romantic prophet of catastrophe. He is a boringly precise man.
What does he say, in essence? That large language models — those LLMs that produce fluent and convincing text — are "approximations of language use, not language understanding." The distinction sounds technical. It is not. It is fundamental.
Think of the difference between an actor who recites a prayer perfectly and a believer who utters it with tears in their eyes. Both produce the same words. One knows what they are saying. Or at least: one has a relationship with what they are saying, an inner experience that precedes and accompanies the words. The other is a mechanism of reproduction, however sophisticated. Our grandparents knew this too — and needed no doctorate in cognitive science to articulate it.
Marcus predicted, before 2024, seven things about the evolution of AI. He published them publicly, with dates. By the end of 2024, all seven had been borne out: the ecosystem had stalled at the level of GPT-4, hallucinations had not been resolved, corporate adoption had been far more limited than promised, and the profitability of AI companies was fictitious outside of hardware. Seven out of seven. And yet nobody offered him a platform on CNN. Nobody invited him to Davos to chair a round table. Sam Altman, who had promised AGI in 2025 and AI agents transforming the global workforce, was everywhere.
Why? Why is the man who is wrong more visible than the man who is right?
The answer is not mysterious. Beauty lies in the eye of the beholder — and the AI industry, AI investors, and journalists covering AI all favour the promethean narrative: the machine that will save or destroy humanity, depending on which camp you speak from. Marcus offers something far harder to sell: nuance, caution, methodology. Predictability. Predictive accuracy. Things that never have an audience in the short term.
There is something in Marcus's scepticism that resonates deeply with my own experience as a web practitioner working with public institutions. When a mayor asks me to implement "AI solutions" on the town hall's website, what he is asking for is not understanding. He is asking for the appearance of understanding. He wants the machine to seem clever. Marcus would recognise this immediately: it is not a technical problem. It is a problem of convenient collective illusion. And convenient collective illusions are, in the history of humanity, the hardest to uproot.
"A lie has short legs," the saying goes. But short legs can travel far if they run along wide corridors and nobody asks questions.
Marcus asks questions. He is not popular. But he is present, consistent, and — a rare thing in public debate about AI — verifiable.
Ted Chiang, or: the Artist Who Understood What Engineers Cannot
There are people who see the structure of a thing more clearly from outside it than those who build it. Ted Chiang is one of them.
Chiang is not a researcher. He has no laboratory. He is a science fiction writer — and on that basis, many people in the technology industry would be tempted to dismiss him. Fatal mistake. Some of the most precise diagnosticians of an age have been its writers of fiction. Kafka understood bureaucracy better than any contemporary sociologist. Orwell understood the language of power before all the political analysts. And Chiang has understood the limitations of generative artificial intelligence with a clarity that few researchers have attained publicly.
His central metaphor — the one that brought him to audiences beyond SF circles — is simple and devastating: ChatGPT is a blurry JPEG of the internet. A lossy compression. Something that retains the approximation, not the substance. Something that looks like the original, if you stand far enough away and do not ask inconvenient questions.
I came across this metaphor with a sensation of immediate, almost physical recognition. Because that is exactly what I had seen in my own work: texts generated by AI for public institutions' websites look like official texts. They have the right tone, the appropriate length, the technical terms in their proper places. But they are hollow. They are blurry JPEGs of Romanian bureaucracy, without the experience behind them, without the accountability that should accompany them, without the specific individual who knows that what they sign produces real effects in the life of a real citizen.
"Dust in the eyes," as the saying goes. And not as a rhetorical flourish — as a precise diagnosis.
Chiang goes further. In an essay in The New Yorker, he compares AI to McKinsey & Company — the consulting firm famous not for the solutions it offers but for the function it fulfils in reality: the laundering of responsibility. When a corporation hires McKinsey and applies McKinsey's recommendations, it can make people redundant, destroy communities, take deeply immoral decisions — and say: we did not decide, we followed external expertise. AI does the same thing. It allows decision-makers to hide behind the algorithm. "The system generated this." "The model recommended." "There is nothing we can do, the data says."
This is not a technical problem. It is a moral one. And it is a problem I have seen concretely, in the flesh, when Romanian public institutions refused to take responsibility for automatically generated content, invoking the very automation as a shield. Exoneration by algorithm. Digital irresponsibility dressed up as modernity.
Chiang says something still more profound about art — and by extension, about any form of authentic production of meaning. If art is a concentrated product of intention, then generative AI is a method of diluting intention. Diluted intention. How many things can be described by that phrase? How many official texts, how many political speeches, how many letters of condolence generated in two seconds? Diluted intention. Diluted meaning. Diluted responsibility.
And yet — here is the paradox Chiang lives and articulates with honesty — he is not anti-technology. He is not a digital Luddite. He draws a distinction: an electric guitar does not make the guitar easier, it transforms it. A camera did not simplify artistic photography, it created an entirely new domain of difficulty and expression. Good tools amplify difficulty in the right direction. Generative AI, in its current form, does the opposite: it simplifies what ought to be hard, removes the resistance that produces growth, offers a short circuit precisely where effort would be necessary.
"It is the man who sanctifies the place," not the machine that simulates sanctity.
I read Chiang's essays with a pencil in hand — a real metaphor, not a rhetorical one, the marks are there on the paper. And each time I felt that someone with different tools from mine — with fiction, with literary essay, not with theology or Romanian philosophy — was arriving at the same wall and tapping on it with the same fingers.
Karl Friston, or: What It Truly Means to Think
If Marcus is the man of data and Chiang the man of metaphors, Karl Friston is the man of equations. And it is precisely for that reason that he is the most uncomfortable of the three for the AI industry — because he cannot be accused of not understanding the mathematics.
Friston is a psychiatrist and computational neurologist at University College London. He is the author of one of the most cited principles in modern neuroscience: the Free Energy Principle — a mathematical formulation of the way in which biological systems process information, reduce surprise, and adapt to the world. I will not venture into equations. I will say only what matters for our argument.
At Davos in 2024, Friston said publicly, in front of the global financial élite, that the deep learning architecture on which generative AI rests has fundamental limitations compared to the way biological intelligence operates. Not as an insult. As a technical diagnosis.
And here is why this matters concretely: the brain of a four-year-old child, Friston said in debate with Yann LeCun, has processed more sensory information than the world's largest language model — and has done so with an energy efficiency unimaginable compared to the servers that consume electricity like a medium-sized town. But more important than efficiency: the child's brain acts upon the world. It tests. It confirms or refutes. It builds an internal model of reality through physical interaction, through mistakes, through falls and recoveries. A child who learns that fire burns has not read about fire. They have touched it.
Large language models do not touch anything. They process text. Text written by people who have touched, felt, erred, loved, wept — and transformed experience into words. AI collects the words, not the experience. It is as though you wanted to learn to swim by reading every book ever written about swimming. At a certain point, you could describe the crawl stroke perfectly. But when you jump into the water, you drown.
"Stories do not replace bread." And texts about the world do not replace the world.
Friston proposes an entirely different architecture for artificial intelligence — Active Inference — based on the biological principles of minimising surprise through continuous action and perception. A system that does not consume text but interacts with its environment. That does not predict the next word but anticipates the consequences of its actions. Something that resembles an organism more than a library.
I do not know whether Friston will succeed. I do not know whether his architecture will work at the scale that matters. But I know that his question is the right one: what does it truly mean to think, and why is the current solution nowhere near an answer to that question?
And in that question I recognise something from the tradition I carry: from the Philokalia, which draws the radical distinction between knowledge and wisdom, between the accumulation of information and the inner transformation produced by authentic experience. The Holy Fathers knew, without equations, that not every accumulation of words produces illumination. That you can read Scripture all your life and remain in darkness if the reading is not accompanied by the ascetic practice that opens understanding.
Generative AI is, in this light, an exercise in accumulation without the corresponding asceticism. It can reproduce Scripture. It cannot live by it.
"Knowledge without faith is blind," my grandfather would have said. He was speaking of something else — but, living paradox that it is, he was speaking of exactly this.
The Consistent and Obscure — the Others Who Go Unseen
Marcus, Chiang, and Friston are visible. They have publications, platforms, followers. But the current of lucid scepticism about AI runs far wider than three names. It is populated by people no one outside their specialism has heard of — and it is precisely for that reason that their opinion is purer, uncontaminated by the need for an audience.
Meredith Broussard, a data journalist at NYU, wrote Artificial Unintelligence — a title that says it all — demonstrating systematically how the biases in training data produce systems that perpetuate existing inequalities rather than resolving them. If the data AI has seen was produced by an unequal world, AI will reproduce that inequality with increased efficiency and behind a veneer of objectivity that makes it harder to challenge. Harder to challenge than human injustice — because human injustice has a face, can be called to account, can be shamed in public. Algorithmic injustice is invisible and impersonal. You cannot take a model to court.
Timnit Gebru, an AI ethics researcher dismissed by Google after publishing a paper on the risks of large language models, showed that these systems encode historical biases at industrial scale. Not as a political opinion — as a measurable result. She was dismissed. The episode changed nothing in the industry in the short term. It did, however, change the thinking of several thousand researchers who understood that scepticism about AI is not a left-wing or right-wing position, it is a position of intellectual honesty — and that intellectual honesty sometimes costs you your job.
Emily Bender, a linguist, introduced the term "stochastic parrot" — a system that produces plausible sequences of words without understanding them, exactly as a parrot repeats human sounds without comprehending them. The metaphor was taken up, distorted, furiously attacked — a sure sign that it had touched something sensitive. When a metaphor produces anger rather than counter-arguments, the metaphor has found its target.
And there are others, still less visible even to me: researchers in university laboratories publishing studies ignored by the mainstream press; engineers inside major technology companies signing open letters anonymously; philosophers of mind showing that the Turing Test — the test which modern AI passes with ease — never tested what people believed it tested. The Turing Test checks whether a machine appears intelligent to a human observer. It does not check whether it is intelligent. It is a test of deception, not of competence. And we have passed it. Congratulations.
"Many set out, few return," says the proverb about a hard road. This current of scepticism is populated by people who set out from the inside — from AI laboratories, technology corporations, prestigious publications — and returned with news nobody wanted to hear. They did not return with hatred for technology. They returned with honesty about what they had seen.
That honesty has a cost. It costs visibility, comfort, sometimes careers. But it is the only asset that does not devalue with the next version.
Why They Are All Ignored in the Same Way
There is a pattern that connects all these voices — and that connects me to them, despite all the geographical and disciplinary distance between us.
We are all treated in the same way: not with counter-arguments, but with rhetorical marginalisation.
Marcus is a "mediocre deep learning sceptic" — Sam Altman's words, reproduced in the press with approving condescension. Not "he is right on points X, Y, Z, but wrong on W." Simply: he is mediocre. Adjectives, not arguments. Label, not debate.
Chiang is "a science fiction writer" — subtext: what does he know, he is not an engineer. As though engineering were the only valid epistemology for discussing intelligence and meaning. As though Dostoyevsky had not understood human psychology better than the psychologists of his day. As though Kafka had not mapped modern bureaucracy more precisely than any sociologist of his era.
Friston is "too technical, too abstract" — a comic paradox, given that his opponents boast of their own mathematics. Too technical for journalists, too philosophical for engineers, too radical for conventional academia. At home nowhere, precisely because he is correct everywhere.
And I — at my incomparably smaller scale — am "a sceptic on principle", "without understanding the technology", "left behind by the times." I have heard all of these. Not necessarily out of ill will. Out of convenience. It is easier to affix a label than to argue.
"The dogs bark, but the caravan moves on," the proverb says. But who is the dog and who is the caravan? Sometimes the caravan is the multi-billion-dollar industry with a thousand PR operatives, and the dog is the lone person saying something is not right. In the original version of the proverb, the dog is the one who does not understand. But there is another version, less often cited: the dog barks precisely because it has smelt something the caravan refuses to smell.
There is a rhetorical technique I recognise from reading about propaganda: the naturalisation of revolution — the moment at which an innovation is presented not as a choice but as an inevitability. "AI will transform everything" — not if, not perhaps, not depending — will transform. Indicative future, without a subject to bear the responsibility. When the inevitable is proclaimed, the person who questions it is no longer a legitimate interlocutor — they become an illogical obstacle in the path of progress. An enemy of the future.
This is, in the end, the most sophisticated form of discursive violence: not the attack, but the exclusion from the field of legitimate conversation. Not "you are wrong", but "you are irrelevant."
Marcus, Chiang, and Friston have resisted this exclusion through the different qualities of their resistance: Marcus through verifiable precision, Chiang through poetic clarity, Friston through mathematical depth. I resist, with my more modest tools, through what I have: practical experience, philosophical and theological tradition, and the conviction that "a person with one good idea is worth more than an army with bad ones."
Convergence — What Unites Us All
I have not spoken to any of them. I do not know whether they know I exist. And yet there is something that unites us, something deeper than agreement on a position: a common way of asking questions.
We all ask: what is missing?
Not: how much can AI do? Not: how fast is it growing? Not: how much money does it make? But: what, specifically and essentially, is missing from what it does?
Marcus answers: understanding of the world, an internal model of reality, the capacity to distinguish truth from plausibility. Chiang answers: intention, responsibility, the experience that precedes the word and gives it weight. Friston answers: action in the world, testing through consequences, the reduction of surprise through active engagement with reality. I answer, in my own vocabulary: presence, relationship, accountability to a concrete you who has a face and a name.
These are different answers to the same question. And that question — what is missing? — is the only question worth asking. Because if you answer it honestly, everything else follows: what is useful, what is dangerous, what merits implementation and what merits rejection.
The AI industry avoids this question with an elegant methodology: it substitutes what is missing? with what comes next? A shift from ontology to chronology. We do not discuss what AI fundamentally is, but which version of it arrives next month. Perpetual future, perpetual anticipation, perpetual forgetting of the real present — and of the uncomfortable questions that can only be asked in the present.
"Haste makes waste." And haste made systematically, at industrial scale, with billions behind it, haste engineered in advance before anyone has time to notice what has been broken — that is something else entirely. That is policy, not accident.
What It Means to Be an Insider
An important detail, which I have left until now: Marcus, Chiang, and Friston are not outsiders. They are not digital Luddites. They do not arrive from outside the field with pitchforks raised.
Marcus founded a machine learning company. Chiang spent years working as a technical writer for software companies in Silicon Valley, understands code, understands the architectures, understands the limitations from the inside. Friston has built some of the most influential computational models of the brain of recent decades — his models are used by researchers the world over.
All three have their hands in the dough of the discipline.
And it is precisely for that reason that their voice carries a different weight from that of a pure philosopher or a theologian looking on from a distance. They say: we have seen how this is made. We know what is inside. And it is precisely for that reason that we tell you: what is outside is not what you think.
It is the difference between the man who has been inside the chemical plant and warns you about the smell, and the man who reads the environmental reports. Both may be right. But the first one breathed it in.
I felt the need to make this distinction because my own position is rather that of the person outside the laboratory. I have not built AI models. My scepticism comes from elsewhere: from practice, from philosophy, from theology, from experience with institutions that use or abuse technology. It is a valid scepticism, but it is a different one — it is the scepticism of someone who sees the effects, not of someone who knows the technical causes.
What Marcus, Chiang, and Friston have given me is the confirmation that my scepticism — with all its limitations — is not the product of ignorance. It is the product of a different kind of seeing. And inside seeing and outside seeing, when they arrive at the same diagnosis, together produce a certainty heavier than either one alone.
"Two eyes see better than one." And two eyes that have looked from opposite directions and seen the same thing — that is already science.
Scepticism as an Act of Love
I want to close with something that may seem paradoxical: my scepticism about AI, and that of the people I have described, is not an act of hostility towards technology or towards those who build it.
It is an act of love for what technology could become if it were looked at honestly.
The greatest betrayal one can commit against something real is to substitute the appearance of it and declare that the appearance is enough. To say that a credible hallucination replaces knowledge. That a blurry JPEG replaces the original photograph. That fluent text replaces the thought that produced it. That speed replaces depth. That the number of words per second replaces the weight of a single word spoken at the right moment.
We are not against AI. We are against the comfortable lie that the AI of today is what we are told it is.
Marcus wants AI to be evaluated fairly — no more, no less. Chiang wants human intention not to be diluted behind a faceless algorithm. Friston wants artificial intelligence to become truly intelligent — not to simulate intelligence on text. I want the Romanian public institutions that serve real citizens not to replace responsibility with irresponsible automation.
These are compatible desires. They are, at bottom, the same desire expressed with different instruments: that things should be what they claim to be.
"A person shows their true character when tested," as the saying goes. And a person is also tested at the truth — at the capacity to say yes, this is good or no, this is not what you say it is without calculating what it is convenient to believe, without first checking how the others are applauding.
This capacity — to sit with an uncomfortable question without resolving it prematurely — is perhaps the most important form of intelligence that no language model will ever acquire, however large. Not because it is impossible in principle. But because it has never appeared in the product specifications.
Epilogue: I Am Not Alone, and Neither Need You Be
I began with the moment of solitude. I end with the way out of it.
I did not emerge from solitude through conversion — I did not come to applaud what does not merit applause. I emerged by discovering that my solitude was partly illusion: I was alone in my room, but not alone in the world.
There is a long and honourable tradition of people who have refused to confuse noise with music, crowd with truth, speed with direction. Not all of them are celebrated. Many are obscure — secondary school teachers who ask inconvenient questions of their pupils, programmers who refuse to sign documentation for a system they do not trust, journalists who write that the emperor has no clothes in an age when the public wants stories about the emperor's garments.
Gary Marcus, Ted Chiang, Karl Friston are not prophets. They are not heroes. They are, simply, people who looked carefully at something and said what they saw. It is a small act and an absolutely essential one. It is an act that costs something — comfort, collegiality, sometimes contracts — and it is precisely for that reason that it merits attention.
If you have read this far — if something in what I have written resonated with something in your own experience, with a moment when you too sat with your hands in your lap while the room applauded — then you know that you are not alone, either.
The tribe exists. It is dispersed, inconsistent in its organisation, varied in its tools and its languages. But it exists. And on the day when someone truly asks the question what is missing? and demands an honest answer, this tribe will be the one that has the answer ready.
"Do good and cast it on the water," as the saying goes. Not because it does not matter. But because good done genuinely has no need to be seen today in order to count tomorrow.
I write this and think of my son. And I tell myself that the best thing I can leave him is not a conclusion about AI — neither for nor against. But a model of how to ask questions. Of how to sit with a question. Of how not to replace understanding with speed, thought with text, presence with efficiency.
Of how not to applaud before he knows why.
Petru Cojocaru writes at cojocarupetru.info about technology, digital culture, and the limits of machines. He is the founder of ABSOLUT WEB EXPERT SRL, with eight years of experience building and maintaining websites for Romanian public institutions.
References and Recommended Reading
- Gary Marcus — Taming Silicon Valley (2024)
- Gary Marcus & Ernest Davis — Rebooting AI (2019)
- Ted Chiang — "ChatGPT Is a Blurry JPEG of the Web", The New Yorker (2023)
- Ted Chiang — "Will A.I. Become the New McKinsey?", The New Yorker (2023)
- Ted Chiang — "Why A.I. Isn't Going to Make Art", The New Yorker (2024)
- Karl Friston — Davos Debate 2024, World Economic Forum (available on YouTube)
- Meredith Broussard — Artificial Unintelligence (2018)
- Emily Bender et al. — "On the Dangers of Stochastic Parrots" (2021)
- Timnit Gebru — papers on bias in large language models (accessible via Google Scholar)
Questions and Answers — and Where the Article Overstates Its Case
Supplementary section to: "The Emperor Has No Clothes"
15 Questions the Article Answers
1. Why does the author feel alone in his scepticism about AI?
Because public debate on artificial intelligence is dominated by two extreme voices — messianic enthusiasm and apocalyptic fear — leaving no room for the middle position: lucid, verifiable scepticism, without hostility to technology. Anyone who refuses both extremes becomes invisible, not contradictory.
2. Who is Gary Marcus and why does his opinion matter?
Gary Marcus is a cognitive psychologist at NYU and former founder of a machine learning start-up acquired by Uber. He is not an outsider — he is an insider who chose to say publicly what many in the field think privately: that large language models produce plausible text without understanding any of what they produce. His importance derives from precision: his predictions are public, dated, and verifiable.
3. What did Gary Marcus predict that came true?
Before 2024, Marcus published seven predictions about the evolution of AI: that the ecosystem would remain at the level of GPT-4 without a revolutionary leap forward, that hallucinations would not be resolved, that corporate adoption would fall far short of expectations, and that the profitability of AI companies would be fictitious outside of hardware. All seven were broadly borne out — without bringing him the visibility enjoyed by the promoters who were wrong.
4. What is the central distinction Marcus draws about LLMs?
That large language models are approximations of language use, not language understanding. They produce statistically correct output, not output derived from understanding. It is the difference between an actor who recites a prayer perfectly and a believer who utters it. The words are identical. The relationship to them is not.
5. Who is Ted Chiang and why is he relevant to the AI debate?
Ted Chiang is a science fiction writer, author of the story that became the film Arrival. He has published a series of essays in The New Yorker analysing generative AI with a clarity that few technical researchers have attained publicly. His relevance comes precisely from the fact that he is outside engineering: he sees the structure of the problem without being blinded by the builder's enthusiasm.
6. What does "ChatGPT is a blurry JPEG of the internet" mean?
It is Chiang's central metaphor: a lossy compression. Just as an aggressively compressed JPEG retains the approximation of the original image — the large colours and shapes — but loses the details and accuracy, generative AI retains the approximation of human language — its fluency, structure, and tone — but loses what lay behind it: the experience, the intention, the responsibility. It looks like the original if you stand far enough away.
7. Why does Chiang compare AI to McKinsey?
Because both serve the same function: the laundering of responsibility. McKinsey allows corporations to make hard decisions and say "we followed external expertise." AI allows institutions to produce content, make algorithmic decisions, and say "the system generated this." Responsibility disappears behind a presentable intermediary. It is not a technical problem — it is a moral one.
8. Who is Karl Friston and what does he criticise in current AI?
Karl Friston is a psychiatrist and computational neurologist at University College London, author of the Free Energy Principle — one of the most cited theories in modern neuroscience. His critique is technical and profound: the deep learning architecture on which generative AI rests does not replicate the way biological intelligence functions. The brain does not predict the next word — it acts, tests, confirms or refutes through real contact with the world. LLMs do not touch anything.
9. What does Friston propose as an alternative?
Active Inference — an architecture based on the biological principles of minimising surprise through continuous action. A system that interacts with its environment, tests the consequences of its actions, and updates its internal model of the world through real engagement, not through passive processing of text. Something closer to an organism than to a library.
10. What is the connection between technical scepticism and the Romanian philosophical tradition invoked in the article?
The Philokalia draws a radical distinction between knowledge and wisdom — between the accumulation of information and the inner transformation produced by authentic experience. Generative AI accumulates massively, without the corresponding asceticism. It can reproduce Scripture without living by it. The technical sceptics — Marcus, Friston — arrive at the same diagnosis by mathematical means: without contact with the real world, the accumulation of text does not produce understanding. The vocabularies differ; the wall is the same.
11. Why are sceptics ignored or minimised rather than contradicted with arguments?
Because arguments require effort and produce uncertainty. Labels are quick and produce comfort. Marcus is "mediocre" — Altman's word, reproduced without counter-arguments. Chiang is "just an SF writer." Friston is "too abstract." None is contradicted on the substance. Exclusion from the field of legitimate conversation — not through attack, but through the declaration of irrelevance — is the most effective form of censorship, precisely because it goes unseen.
12. What unites all the sceptical voices described in the article, across different disciplines?
A single shared question: what is missing? Not how much AI produces, not how fast it grows, not how much money it makes — but what, specifically and essentially, is missing from what it does. Marcus answers: an internal model of reality. Chiang: intention and responsibility. Friston: action in the world. The author: presence and accountability to a concrete other who has a face and a name. Different answers, the same question. And that question is the only one that opens a real debate.
13. Why is scepticism about AI described as an "act of love" rather than hostility?
Because the greatest betrayal of something real is to substitute its appearance and declare that the appearance is sufficient. To say that current AI is what it claims to be — intelligent, comprehending, responsible — is a betrayal of what it could become if it were regarded with honesty. Authentic scepticism does not want less technology; it wants better technology, evaluated fairly, used where it actually works, and rejected where it merely simulates working.
14. What is the relevance of this argument for Romanian public institutions?
Concrete and immediate. Public institutions — town halls, schools, hospitals, county councils — are adopting or are being pressured to adopt "AI solutions" without knowing what they are purchasing and without being able to evaluate what they are losing: accountability to the citizen, ownership of an official text, direct relationship with the reality they administer. Generative AI in public administration is not a technological problem — it is a problem of democratic accountability.
15. What should a reader who resonates with the article's arguments do?
Start with the question what is missing? whenever confronted with an enthusiastic presentation of AI. Not reject the technology — evaluate it honestly. Read Marcus, Chiang, Friston. Observe the difference between fluent text and real thought. Do not apply the tool where the simulation of functioning does more harm than the absence of the tool. And, perhaps most importantly: do not feel alone if they cannot applaud.
Where the Article Overstates Its Case
After rereading the text I noticed several instances of excess that I want to address. I acknowledge I exaggerated somewhat — but strategically, in tone and generalisations, not in the underlying facts.
The foregoing article is an opinion piece with a polemical style and strong rhetoric. Expressions such as "the emperor has no clothes," "collective illusion," and "a lie has short legs" are deliberately absolutist — they belong to the voice and the stakes of the text. But intellectual honesty requires naming the places where the formulation exceeds measurable reality. What follows is a balanced analysis, grounded in the article's text and in data available in 2025–2026.
1. The Most Significant Overstatement: "Total Stagnation at GPT-4 Level"
The article draws on Marcus's predictions and states that the AI ecosystem "has remained stuck at the level of GPT-4." The predictions were largely borne out — but not entirely.
What is correct: Pure scaling models have disappointed. GPT-5 was described by Marcus and other analysts as "overdue, overhyped, and underwhelming." The underlying problems — persistent hallucinations, the absence of a robust world-model, enormous energy costs — remain unresolved. AGI has not arrived. Large-scale corporate adoption has not materialised at the promised level.
What is more nuanced: Real progress has been made in reasoning (models of the o1/o3 type), in coding agents, in open-source models such as DeepSeek R1, and in limited but functional practical applications. It is not "nothing new" — it is closer to "not what was promised": a slowing of the promised revolution, with a paradigm shift towards agents and reasoning chains, not a total standstill.
Conclusion: The article presents the situation as an absolute standstill. The reality is a disappointment relative to the promises, with genuine progress in specific niches. The sceptics are right in their essence — the hype has outrun the delivery — but the word "stuck" is more dramatic than "slower and more limited than promised."
2. "Profitability Was Fictitious" — Minor Overstatement, Close to the Truth
The article states that the profitability of AI companies was fictitious, with the exception of hardware manufacturers.
The reality in 2026: Very close to the truth. OpenAI reports significant revenue, but projections of continuing losses are massive — estimates run to tens of billions cumulatively before any eventual profitability, perhaps 2029–2030. The majority of LLM companies are burning capital heavily on compute. Chip manufacturers — NVIDIA above all — are the only ones posting solid and consistent profits.
Conclusion: This observation is correct and confirmed by public financial data. It is not a significant overstatement.
3. The General Tone: "Everyone Is Applauding Blindly"
The comparisons with the Andersen fairy tale, with convenient collective illusions, with sidelined sceptics and omnipresent enthusiasts — this is the area of most obvious rhetorical exaggeration.
The reality: Not everyone is applauding blindly. The debates between Marcus and proponents of scaling laws are public, intense, and visible. Many engineers and analysts openly acknowledge the limitations — hallucinations, costs, lack of reliability. Scepticism is not marginal: it is present on Substack, on Twitter/X, in The New Yorker, in academic papers. The author overstates both his own isolation and that of sceptics for dramatic effect — he feels "alone," but scepticism has an active and visible community.
Conclusion: This is deliberate rhetorical exaggeration, not factual error. It works within the logic of the text — the moment of solitude followed by the discovery of the tribe has narrative power. But read as an exact sociological description, it overstates the marginalisation of sceptics.
4. Other Nuances
The blurry JPEG metaphor (Ted Chiang) — correct and well-calibrated, not an overstatement.
Friston is not only an academic researcher but also Chief Scientist at VERSES AI Inc. — a commercial company that promotes Active Inference as an alternative to deep learning. There is therefore a potential conflict of interest in Friston's case that the article does not flag.
The dilution of responsibility in public administration — highly relevant to the Romanian context, but generalised as a universal problem of AI. There are contexts in which AI assists without replacing human responsibility — the article does not distinguish between them.
The connection with the Philokalia and asceticism — culturally rich, consistent with the author's voice, but subjective and philosophical. It is not a claim that can be overstated — it is an interpretation, and as an interpretation it is legitimate within the frame of a personal essay.
The text's principal vulnerability is tonal absolutism — it presents a complex debate as a fairy tale with ignored heroes and a blind multitude, which simplifies a more nuanced reality.
Balanced Conclusion
The article overstates more in tone and absolutism — "total standstill," "everyone is blinded," inevitable collective illusion — than in concrete facts. Marcus's predictions, the economic problems of the AI industry, and its architectural limitations are broadly correct according to data from 2025–2026.
It is a healthy antidote to excessive hype. But like any sceptical manifesto, it simplifies the complex reality somewhat: slow but real progress in niches, a sceptical community more visible than the narrative of solitude suggests, the difference between "not what was promised" and "nothing new."
It is not propaganda. It is not hallucination. It is a well-documented personal confession, with a touch of Romanian dramatic instinct — and if read as opinion rather than neutral technical analysis, it is good. More responses from friends to follow.