--- title: "What Is Artificial Intelligence? Definition and Core Principles" date: 2026-05-27 description: "What is artificial intelligence, really? A machine-based system designed to operate with varying levels of autonomy — one that has already moved into our daily lives without knocking." author: "Petru Cojocaru" categories: - name: "AI - a tool" url: "https://cojocarupetru.info/ai-a-tool.md" --- # What Is Artificial Intelligence? Definition and Core Principles There are moments when you realise the world you live in no longer resembles the one you grew up understanding. It isn't a dramatic shift — no announcement, no fanfare. [![What Is Artificial Intelligence? Definition and Core Principles](https://cojocarupetru.info/images/AI-o-unealta/resized_ce-este-ai.jpg "What Is Artificial Intelligence? Definition and Core Principles")](https://cojocarupetru.info/index.php?option=com_content&view=article&id=380:what-is-artificial-intelligence-definition-and-core-principles&catid=41:ai-a-tool&lang=en-GB&Itemid=511)It's slow, almost insidious: the kind you notice when your son stops looking for answers in books and turns instead to a dialogue box on a screen, and that dialogue box answers him more quickly, more articulately, than you ever could. That is artificial intelligence in our actual lives. Not an academic abstraction. A presence that has already set up its desk in the middle of our everyday existence — uninvited, unannounced, entirely at home. What follows is my attempt to explain what this technology actually is. Not as a Stanford engineer might explain it, nor as a DeepMind researcher would. But as an ordinary person who has come to live alongside it, trying to understand what, exactly, he is cohabiting with. --- **What Is Artificial Intelligence, Really?** The European Union, through its 2024 Artificial Intelligence Act — the first major legal framework in the world dedicated to this technology — defines an AI system as *"a machine-based system designed to operate with varying levels of autonomy, that may exhibit adaptiveness after deployment and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments."* It is a cold, juridical definition. Useful for legislators. For anyone who wants to understand the thing itself, I would put it more plainly: artificial intelligence is the capacity of a computing system to imitate human cognitive processes — to learn from experience, to recognise patterns, to make decisions, to generate content — without being explicitly programmed for every individual situation it might encounter. We are not talking about a single technology. We are talking about a vast field that encompasses machine learning, neural networks, natural language processing, computer vision, recommendation systems, and, more recently, generative models capable of producing text, images, music, or software code at a quality that, ten years ago, would have been unimaginable. --- ## Weak AI and Strong AI — A Distinction Worth Making ### Weak AI (Narrow AI) Everything that exists today, everything we actually use — from Siri and Google Assistant to the algorithm deciding what you should watch next on Netflix, to the system flagging fraud on your bank card — is what specialists call narrow AI, or weak AI. The word *weak* carries no pejorative weight. It simply means these systems are designed and trained for specific, well-defined tasks, and cannot generalise their knowledge beyond the domain they were built for. A system that recognises faces in photographs cannot, through that same architecture, play chess. One that translates text cannot drive a car. Every AI system in existence today excels within its perimeter and remains entirely blind beyond it. Understanding this structural limitation matters enormously, because so much of the public anxiety around AI stems from a confusion between what actually exists — powerful but narrow tools — and what remains, for now, purely theoretical. ### Strong AI — AGI (Artificial General Intelligence) Strong AI, or Artificial General Intelligence (AGI), is the concept of a system capable of understanding, learning, and applying knowledge across any cognitive domain, at human level or beyond. Such a system could move between tasks without retraining, reason within entirely new contexts, exhibit something resembling creativity, and — in the more radical versions of the concept — develop a form of self-awareness. The essential thing to understand is that AGI does not yet exist. This is not a matter of months, or even a few years — though estimates vary considerably. Demis Hassabis, CEO of Google DeepMind, places *"genuine human-level AGI"* on a horizon of several years, emphasising that current systems, however impressive, do not replicate human thought in its full complexity. Shane Legg, DeepMind's co-founder and chief AGI scientist, has publicly estimated a 50 per cent probability of minimal AGI by 2028. Eric Schmidt, Google's former CEO, suggested in April 2025 that we are three to five years away. And 76 per cent of the researchers surveyed in the AAAI 2025 panel believe that simply scaling up current approaches will not be sufficient to reach AGI. The honest scientific consensus is this: AGI may arrive in the coming decades. It may never arrive in the form imagined. And it is certainly not *near* in the sense the press so often implies. What we have now are extraordinary tools — not the artificial minds of science fiction. --- ## How It Actually Works ### Machine Learning The fundamental principle of machine learning is elegant in its simplicity. Rather than explicitly programming a computer with rules for every conceivable situation, you give it examples — thousands, millions, sometimes billions — and let it discover the patterns itself. The system *learns* from data, not from instructions. There are three broad types: **Supervised learning** — the system receives paired inputs and correct answers (images labelled "dog" or "cat," for instance) and learns to classify new images correctly. Applications: facial recognition, spam detection, image-based medical diagnosis. **Unsupervised learning** — the system receives unlabelled data and identifies structures and groupings on its own. Applications: customer segmentation in marketing, anomaly detection in cybersecurity networks. **Reinforcement learning** — the system learns through trial and error, receiving rewards for correct actions and penalties for wrong ones. This approach underpinned DeepMind's AlphaGo, which defeated the world's best Go players, and it also underpins the training of today's most advanced generative language models. ### Neural Networks and Deep Learning Artificial neural networks are computational architectures loosely inspired by the organisation of the human brain — interconnected nodes arranged in layers, processing information through successive waves of mathematical transformations. Deep learning is the variant with many hidden layers, capable of extracting abstract features from raw data without any human guidance. This architecture has transformed fields including speech recognition, machine translation, and image generation, and more recently has given rise to Large Language Models — LLMs — the technology underlying ChatGPT, Gemini, Claude, and their peers. A notable recent benchmark: in 2025, Google DeepMind's Gemini model, running in Deep Think mode, achieved a gold-medal performance at the International Mathematical Olympiad — solving five of six problems within the official time limit, in natural language, without recourse to formal symbolic reasoning tools. It is a significant signpost for the reasoning capacities of current systems. ### Generative Models — The Paradigm Shift of 2022–2025 Where classical machine learning analysed and classified, generative models *create*. ChatGPT (OpenAI), Gemini (Google), Claude (Anthropic), Llama (Meta) — these are all large language models trained on vast quantities of text, capable of generating coherent content, answering complex questions, writing code, summarising documents, translating, or sustaining extended conversation. This is the paradigm shift we are living through in real time. AI has moved from a tool that *processes* information to one that *generates* content — with profound implications for education, work, creativity, and, inevitably, the labour market. --- ## Applications in Daily Life ### Voice Assistants and Conversational Interfaces Siri, Google Assistant, Alexa — the voice interfaces millions use each day — are narrow AI applications: speech recognition, natural language processing, integration with external services. Their limits are visible enough. They function well within defined tasks and miss entirely anything involving context, nuance, or irony. Far more capable are today's LLM-based conversational interfaces, accessible through apps or APIs, which can sustain extended conversations, analyse documents, assist with writing, or help with programming. The difference from the previous generation of voice assistants is structural, not merely incremental. ### Search Engines and AI Overviews Google has integrated generative AI deeply into its search product through its *AI Overview* feature, which synthesises answers directly on the results page. Bing, integrated with OpenAI's models, works similarly. This evolution fundamentally alters the relationship between user and information: rather than navigating towards sources, the reader receives a ready-made synthesis — with all the advantages and risks that entails. Oversimplification. Unverified sources. Loss of context. ### Recommendation Systems Netflix, Spotify, YouTube, TikTok, Amazon — the recommendation engines of these platforms are, perhaps, the form of AI with the greatest direct impact on human behaviour at a global scale. They do not merely suggest what you might enjoy; over time, they *shape* what you will prefer, creating feedback loops that can amplify initial inclinations towards more extreme forms. Understanding this mechanism is essential for anyone who wants to understand how attention actually functions in the digital age. ### AI in Medicine, Science, and Industry Beyond consumer applications, AI is producing deep transformations in high-stakes fields. DeepMind's AlphaFold 2 predicted the three-dimensional structure of more than 200 million proteins — a foundational biological problem that had remained unsolved for decades. Diagnostic systems based on medical imaging now match or exceed the accuracy of specialist physicians in detecting certain cancers. Predictive models are being used to manage electricity grids, optimise supply chains, and enable early warning of natural disasters. --- ## Ethical Challenges — Beyond the Clichés Any serious discussion of AI must acknowledge that this technology arrives with a set of structural problems — some of them soluble, some of them constitutional. **Algorithmic opacity** — AI systems, particularly deep neural networks, are in essence black boxes. Even their creators cannot always explain why a system reached a particular decision. This is a grave problem when the decisions in question carry real consequences: bank lending, recruitment, criminal sentencing. **Systemic bias** — systems trained on historical data reproduce, and often amplify, the prejudices embedded in that data. A recruitment system trained on a company's past hiring decisions will "learn" that men are stronger candidates, if that was the statistical reality of its past. Correcting for bias is an active area of research with no complete solution as yet. **Data privacy** — AI models are trained on enormous quantities of data, often gathered without the explicit consent of the individuals concerned. The tension between model performance and personal data protection is one of the great legal and ethical battles of this decade. **Copyright and intellectual property** — generative models were trained on texts, images, and other copyright-protected material. Who owns a work generated by AI? Who bears responsibility for the implicit plagiarism? Courts in the United States and Europe are only just beginning to build jurisprudence in this area. **Labour market impact** — estimates vary, but the emerging consensus is that AI will profoundly transform broad categories of routine cognitive work. Whether it will create more jobs than it eliminates, and at what pace this transition will unfold, remains unclear. What is certain is that the pressure to retrain will be real and significant. **Concentration of power** — Sam Altman, CEO of OpenAI, has publicly warned that advanced AI infrastructure may become concentrated in the hands of a small number of actors, with the risk of unprecedented power imbalances. Investment figures confirm this trajectory: OpenAI reached a valuation of $300 billion, Anthropic $183 billion, and the AGI market is estimated to grow from $2.5–6.2 billion in 2024 to $38–63 billion by 2032. --- ## Conclusion Artificial intelligence is neither the apocalyptic monster of the most pessimistic scenarios nor the universal solution promised by technology's most fervent evangelists. It is a family of extraordinarily powerful tools — with real capabilities and real limits — developing at the fastest pace in the history of technology. What matters, in my view, is not becoming an expert in neural network architectures or following every announcement from specialist conferences. What matters is understanding the basic mechanisms: how these systems learn, what they can and cannot do, whose interests lie behind them. Without that, you risk becoming the passive user of tools you do not understand — shaped by them before you have had the chance to evaluate them. My son understands this instinctively, even though he has never once thought to read an article about machine learning. His generation grew up alongside these tools and handles them with an ease that those of us who remember the world before them will never quite replicate. What they sometimes lack — and what I am trying to offer in these texts — is perspective: the understanding that behind the dialogue box there are human choices, economic interests, technical limitations, and ethical responsibilities. That it is not magic. That it is not the enemy either. That it is, for now, a tool. A tool worth knowing well. --- ## 20 Questions Worth Holding On To --- **1. What is artificial intelligence, in essence — beyond the technical definitions?** It is the capacity of a computing system to imitate human cognitive processes — to learn from experience, to recognise patterns, to make decisions, to generate content — without being explicitly programmed for every situation it might face. It is not a single technology but a vast field encompassing machine learning, neural networks, generative models, and natural language processing. Concretely: it is what answers when you ask something on Google, what suggests what to listen to on Spotify, what presented a radiograph with a suspicious nodule to a physician before the physician had a chance to look. --- **2. Why does the term "weak AI" not mean AI is unintelligent?** *Weak* is not a value judgement — it is a structural description. A narrow AI system is designed for a specific task and executes that task at a level which frequently surpasses human performance. AlphaGo defeated the world's best Go players. Google's skin-cancer detection system matches the accuracy of an experienced dermatologist. But neither of these systems can do anything beyond what it was trained to do. *Weak* means narrow, not inferior — and that distinction matters enormously if we are to avoid false expectations in either direction. --- **3. What is the fundamental difference between weak AI and AGI?** Weak AI exists and functions today — it excels within its perimeter and remains entirely blind beyond it. AGI (*Artificial General Intelligence*) is a theoretical concept: a system that could understand, learn, and apply knowledge across any cognitive domain, at human level or above, without specialised retraining for each new task. The difference is not one of degree but of kind. All AI in use today — including ChatGPT, Gemini, and Claude — is narrow AI, however impressive it appears in conversation. --- **4. Does AGI exist at present?** No. And it is worth saying so plainly, without apologising for the simplicity. AGI does not yet exist in any functional form. There are language models that simulate general conversation, systems that reason mathematically at an olympiad level, robots learning to move through the world — but no existing system can transfer knowledge between arbitrary domains, possesses self-awareness, or can adapt to entirely new contexts without retraining. The press frequently conflates spectacular performance in narrow tasks with general intelligence. They are not the same thing. --- **5. When might AGI arrive — and whom should we believe in the avalanche of predictions?** Estimates vary so dramatically that the variation itself should make us cautious about any precise forecast. Shane Legg (Google DeepMind) puts the probability of minimal AGI by 2028 at fifty per cent. Demis Hassabis (CEO, DeepMind) places *"genuine human-level AGI"* on a multi-year horizon, pointing to gaps in creativity and robust understanding. Eric Schmidt, Google's former CEO, suggested in April 2025 that we are three to five years away. Perhaps most significantly, 76 per cent of researchers surveyed at the 2025 AAAI panel believe that simply scaling up current approaches will be insufficient to reach AGI. The honest answer is: we do not know. And anyone who tells you otherwise is either extrapolating from an incomplete model or has something to sell. --- **6. How does an AI system "learn" — in terms anyone can follow?** Imagine you want to teach a child to recognise cats without ever explaining what a cat is. You show them ten thousand photographs containing cats and ten thousand without, tell them each time whether they guessed correctly, and let them adjust their own internal rules until they rarely get it wrong. Machine learning works in precisely this way — in place of a child, you have an algorithm; in place of photographs, you have data; in place of "you guessed correctly," you have a mathematical error function that the system iteratively minimises. There is no understanding in the human sense of the word. There is statistical optimisation. That distinction matters enormously if we are not to anthropomorphise these systems in ways that mislead us. --- **7. What are the three types of machine learning, and why does the distinction matter?** **Supervised learning** — the system learns from labelled examples (images with correct answers, texts with predetermined classifications). Applications: facial recognition, spam filtering, image-based medical diagnosis. **Unsupervised learning** — the system receives raw, unlabelled data and discovers hidden structures on its own. Applications: behavioural segmentation of customers, anomaly detection in cybersecurity. **Reinforcement learning** — the system learns through trial and error, receiving rewards and penalties. It underpinned AlphaGo and underpins the training of today's generative models. The distinction matters because each type addresses different problems and carries different risks of bias and error. --- **8. What are neural networks, and why aren't they what the name implies?** Artificial neural networks are *loosely* inspired by the organisation of the human brain — interconnected nodes arranged in layers, processing information through successive mathematical transformations. But the resemblance to the biological brain ends there. There are no neurons, no synapses, nothing organic. There are matrices of numbers and algebraic operations. The name is a metaphor that took hold — not an accurate description. The danger is that the metaphor instinctively triggers the wrong associations: that the system *thinks* or *feels* in something approaching the human sense. It does not think. It calculates. The difference is not semantic. It is fundamental. --- **9. What is deep learning, and why did it change everything after 2012?** Deep learning is the neural-network variant with many hidden layers — dozens, hundreds — capable of extracting abstract features from raw data without anyone defining in advance what to look for. Before 2012, an image-recognition system required engineers to manually specify the relevant features: edges, shapes, textures. Deep learning eliminated that step: the system discovers for itself what matters. The consequence was a revolution in speech recognition, image analysis, and machine translation — and, in recent years, in the large language models that have transformed the way we interact with computers. --- **10. What are Large Language Models, and why do they represent a genuine break with what came before?** A Large Language Model (LLM) is a model trained on an immense quantity of text — in practice, a significant portion of everything written on the internet — that has learned to predict, with remarkable statistical accuracy, which word follows another. From this apparently modest capacity something surprising emerges: the ability to sustain coherent conversations, to summarise, to translate, to write code, to analyse documents. The break with earlier AI is this: we have moved from systems that *process* information to systems that *generate* content. ChatGPT, Gemini, Claude — all are LLMs. And all, however fluent, remain statistical systems. They do not understand meaning in the way a human does. They compute probabilities. --- **11. Why are recommendation systems the form of AI with the greatest psychological impact?** Because they do not merely suggest what you might enjoy — they *shape* what you will come to prefer. The algorithms of Netflix, TikTok, Spotify, and YouTube are not neutral selection tools; they are engagement-optimisation systems that create feedback loops: they show you what you like, you consume more of that kind of content, the algorithm confirms and amplifies the preference, the cycle repeats. Over time, the consequence can be a narrowing of taste, a radicalisation of political preference, a dependency on constant stimulation. This is not conspiracy — it is mathematics. The system does exactly what it was optimised to do: maximise time spent on the platform. The problem is that it works. --- **12. What did AlphaFold 2 solve, and why should it matter to someone who is not a biologist?** DeepMind's AlphaFold 2 predicted the three-dimensional structure of more than 200 million proteins — effectively the entire known diversity of the proteome. This had been considered a foundational and intractable problem in molecular biology for decades: a protein's structure determines its function, and its function determines everything, from the mechanism of a disease to the way a drug works. AI resolved in a few years what classical research would have taken generations to map. For someone who is not a biologist: it is proof that AI is not just about chatbots and playlists. It is a research instrument that accelerates science in domains where the stakes are life and death. --- **13. What is algorithmic opacity, and why is it a serious problem rather than merely a technical one?** AI systems based on deep neural networks are, in essence, black boxes. Not even their creators can always explain why a system arrived at a particular decision. When we are talking about music recommendations, this opacity is inconvenient but not dangerous. When we are talking about bank lending decisions, criminal sentencing, employment screening, or medical diagnoses, it becomes a grave ethical and legal problem. How do you contest a decision whose reasoning cannot be articulated? The right to explanation is one of the central principles of the EU AI Act (2024) precisely for this reason — not out of philosophical refinement, but out of practical necessity. --- **14. How does bias arise in AI systems, and why is it so difficult to correct?** Bias is not an intentionally introduced error — it is a structural consequence of how these systems learn. A recruitment system trained on a company's historical hiring decisions will "learn" the statistical patterns embedded in those decisions, including implicit discrimination on grounds of gender, ethnicity, or age. The system is not discriminating maliciously; it is reproducing the mathematics of the past. Correction is difficult because it involves a fundamental tension: more diverse data corrects bias, but gathering it raises fresh questions of privacy and representativeness. There is no simple solution. There is, for now, only degrees of amelioration. --- **15. Who owns the copyright to a text generated by AI?** The honest answer, in May 2026, is: it depends on jurisdiction, and nowhere is it settled definitively. US courts have so far held that a text generated solely by AI, without significant human creative input, cannot benefit from copyright protection — copyright presupposes a human author. In Europe, the AI Act and existing copyright directives are still being interpreted and implemented. In parallel, active litigation proceeds against OpenAI, Google, and Anthropic from publishers and creators whose work was used in model training without explicit consent. This is a field in motion, with major consequences for journalism, literary creation, and the creative industries at large. --- **16. Will AI destroy more jobs than it creates?** Nobody knows with certainty — and anyone who tells you otherwise is either extrapolating from an incomplete model or has an interest in the answer they give. What we do know: AI will profoundly transform broad categories of routine cognitive work — standard drafting, data processing, scripted customer support, translation, mid-level coding. Some of these jobs will disappear or contract significantly. New ones will emerge — though whether at the same pace, and for the same people, remains unclear. What is certain is that the pressure to retrain will be real, unequally distributed, and insufficiently prepared for by current educational systems. --- **17. What does the concentration of power in AI mean for an ordinary person?** It means that the infrastructure required to train and operate frontier AI models — specialised computing hardware, data, energy, rare expertise — is concentrating in the hands of a small number of private American and, in part, Chinese companies. OpenAI at a $300 billion valuation, Anthropic at $183 billion, Google DeepMind with access to Alphabet's entire infrastructure: these are not ordinary companies. They make decisions about what these systems can and cannot do, what data is used, what values are embedded in their behaviour. Sam Altman himself has warned that this concentration could generate power imbalances without precedent. For an ordinary person: it means that the tools shaping your access to information, work, and services are being designed elsewhere, with different priorities, and that the influence you have over them is very nearly nil. --- **18. What is the EU AI Act, and why is it relevant to everyday users?** The European Union's Artificial Intelligence Act (2024) is the first major legal framework in the world to regulate AI according to the risk it poses. High-risk systems — those influencing bank lending, employment decisions, social scoring, critical infrastructure, or medical diagnosis — are subject to strict requirements around transparency, testing, and human oversight. Prohibited systems include generalised social scoring and subliminal manipulation. The relevance for an ordinary person: if a bank, employer, or public institution uses AI in decisions that affect you, you have legal rights — to an explanation, to challenge the decision, to human oversight. It is real protection, even if enforcement in practice will be slow. --- **19. What is the difference between a classic voice assistant (Siri, Alexa) and a modern LLM (ChatGPT, Claude)?** The difference is structural, not a matter of degree — LLMs are not simply better versions of voice assistants. Siri and Alexa are narrow AI systems: they recognise voice commands, access predefined databases, execute scripted actions. They function well in anticipated scenarios and fail at everything else. LLMs are models trained on vast quantities of text, capable of general conversation, document analysis, multi-step reasoning, complex content generation — with their own particular limitations: hallucinations, lack of persistent memory, inability to act in the physical world without external tools. Asking Alexa to set a five-minute timer and asking Claude to analyse a contract and identify unfair terms are fundamentally different acts, not variations on the same gesture. --- **20. What should we take away from AI — not as technologists, but as people living alongside it?** That it is a tool, not an oracle. That behind every dialogue box there are human choices — about design, data, embedded values, economic interests — not neutrality or objectivity. That current systems are extraordinarily capable at what they were trained to do and entirely unprepared for what they did not anticipate. That bias, opacity, and the concentration of power are real problems, not paranoia. That AGI does not exist and is not *near* in any dramatic sense of the word. And that the greatest risk is not that AI will become conscious and subjugate us — it is that we will become passive users of systems we do not understand, shaped by them before we have had the chance to evaluate them. Basic understanding is, in itself, a form of freedom. --- --- ## Author's Note *On limits, tensions, comparisons, and rhetorical tools* --- I wrote this piece several times in my head before I put it on paper — or, more precisely, on screen, which is an irony I feel without forcing it. I rewrote it mentally on mornings when my son asked me something at breakfast, and I watched him frame the question with a pattern he had absorbed from conversations with chatbots — brief, clear, stripped of the meandering context I inherited from books and value precisely because it is not efficient. Efficiency has never been the sole criterion of good thinking. I tell him this, and I feel as though I am speaking to someone who already lives in a different language. **On the limits of the piece** The first thing I should say is that this article is not objective. Not in the sense that it contains false information — I have checked what I have claimed — but in the sense that any choice to include something and omit something else is already a position. I chose to write about AGI with calibrated scepticism rather than the kind of enthusiasm that sells well. I chose to include the EU AI Act not as a juridical footnote but as a signal that society has officially acknowledged this technology needs limits — which strikes me as essential information, not administrative noise. I chose not to include speculations about artificial consciousness, not because the question fails to interest me philosophically, but because at the level of an introductory piece, speculation without empirical grounding does more harm than good. The article's limits are, in large part, deliberate choices. There is another kind of limit: temporality. The figures I have cited — company valuations, AGI predictions, Gemini's performance at the Mathematics Olympiad — are accurate in May 2026 and may be partly outdated within months. This field moves at a pace that makes any text age faster than a book on organic chemistry or medieval history would. I accepted this perishability as a condition of honesty. It is better to cite concrete, verifiable data that can be updated than to take refuge in safe but hollow generalities. **On the tensions I felt whilst writing** The first tension, the most persistent, was between accessibility and precision. If I say a neural network "works like the human brain," I am telling a convenient lie. If I explain that it involves algebraic transformations in high-dimensional spaces, I am telling the truth and losing half my readers. I chose a third path — the metaphor stated, then immediately corrected — which is, I think, the most intellectually honest solution, even if it is the most rhetorically uncomfortable one. "Loosely inspired by the human brain": that single word, *loosely*, carries the full weight of the distinction. The second tension was between neutrality and position. I have a point of view about AI — not an apocalyptic one, not an evangelising one, but one that tries to understand the tool without deifying it or demonising it. But a text claiming perfect neutrality on a subject with this many economic and political stakes is either naïve or dishonest. I preferred transparency about perspective. "A tool worth knowing well" is a position, not an objective conclusion. One I own. The third tension — more personal — is between what I know and what I do not. I have read enough to write with competence about machine learning, AGI, and ethical challenges. I have no formal training in computer science or advanced mathematics. There are limits to my understanding that I have sometimes masked through careful phrasing and at other times exposed through the choice of sources. I try not to write beyond what I genuinely understand — which means there are angles of this subject I have touched superficially or not at all: transformer architectures, information theory applied to language models, the mathematics of loss functions. I did not set out to write a technical treatise. I set out to write something that might help an intelligent but uninitiated person understand the world they are living in a little better. **On comparisons** I used comparisons as instruments of clarification, not decoration. Several of them I consider essential to understanding the piece: *Narrow AI versus AGI* is the structural comparison on which the entire text rests. Without it, all public discourse about AI remains an unresolved confusion — we speak simultaneously about Siri and about the Terminator as though they were the same kind of thing. They are not. The distinction is not academic; it is practical. It determines how you evaluate a tool today, how you vote on legislation, how you bring up a child. *Machine learning versus classical programming* is the second structural comparison. "Rather than programming rules, you offer examples and let the system discover the pattern for itself" — that is an epistemological shift, not merely a technical one. Its consequence — that the system can arrive at solutions its creator did not anticipate and cannot explain — is precisely what makes algorithmic opacity a real problem rather than a philosopher's indulgence. *Classic voice assistants versus LLMs* is a comparison I felt compelled to make explicit because general audiences tend to treat them as versions of the same thing. They are not. The structural difference between asking Siri to set a five-minute timer and asking Claude to identify unfair clauses in a contract is not one of degree. It is one of kind. *AI as tool versus AI as mind* is the implicit comparison that runs through the entire text, never stated directly as such but present in every choice of phrasing. I deliberately avoided verbs that anthropomorphise: the system does not *think*, does not *understand*, does not *feel*. It calculates, optimises, generates. This is not humility before machines — it is honesty before the reader. **On the rhetorical tools I used** *The personal narrative anchor.* The piece does not begin with a definition — it begins with my son. Not because the subject is a domestic one, but because abstractions about technology have a tendency to slide past people without touching them. The personal and concrete is an anchoring device — it places the reader in relation to the subject before the subject is technically articulated. The risk is that it can seem sentimental or contrived. I accepted that risk because the alternative — opening with the EU AI Act's definition — would have been correct and dead on arrival. *A tone calibrated between the accessible and the rigorous.* I avoided jargon where it was unnecessary and kept it where omitting it would have created confusion. "Large Language Model" remains in the text as a technical term because there is a community that uses it and that needs to recognise what we are discussing. "Algebraic transformations in high-dimensional spaces" does not appear, because it adds nothing to the understanding of an uninitiated reader. *The long sentence as a thinking instrument.* I do not write in short sentences — not because I could not, but because a complex idea truncated to three words becomes a slogan, not a thought. The long sentence, with its subordinate clauses, its hesitations, its returns to the original idea, follows the actual course of thinking rather than its presentable form. It is a stylistic choice with a cost — it demands something of the reader — but also with a gain: it does not simplify to the point of distortion. *The immediate correction of the metaphor.* Each time I used a simplifying metaphor — the human brain, the black box, the tool — I followed it immediately with a qualification or a nuance. "Loosely inspired by the human brain." "Black boxes — even to their creators." "A tool — worth knowing well." The metaphor opens the door; the nuance keeps it from slamming into the wall. *The named attribution.* I chose not to make any claim about AGI without attributing it to a named source — Hassabis, Legg, Schmidt, AAAI 2025. Not because these are academic texts, but because in the field of AI, anonymous predictions travel at the speed of myth and become, through repetition, dictionary truths. Naming the source is an act of intellectual hygiene: it tells the reader they can verify, they can contest, they need not take it on faith. **A final, more personal note** I do not know whether my son will ever read this. Probably not — he moves at a different pace, in a different format. But I write it with him in mind, because the only thing I believe I can give a young person in this era is not a safe career path or practical advice about what to study. It is a way of looking — critical without cynicism, open without naivety, capable of distinguishing tool from magic and useful from essential. If this text reaches someone and makes them look at the dialogue box they sometimes spend hours inside — just slightly differently, conscious that behind it there are human choices, interests, and limits — then it was worth writing. The rest is statistical optimisation. --- *Petru Cojocaru**cojocarupetru.info — May 2026* --- ## Bibliography and Sources - **European Union.** (2024). *EU AI Act — Regulation (EU) 2024/1689*. Official Journal of the European Union. [eur-lex.europa.eu](https://eur-lex.europa.eu) - **IBM Think.** (2025). *What Is Artificial Intelligence (AI)?*[ibm.com/think/topics/artificial-intelligence](https://www.ibm.com/think/topics/artificial-intelligence) - **IBM Think.** (2025). *What Is Strong AI?*[ibm.com/think/topics/strong-ai](https://www.ibm.com/think/topics/strong-ai) - **Springer / AI & Society.** (2024). *Strong and weak AI narratives: an analytical framework.*[link.springer.com](https://link.springer.com/article/10.1007/s00146-024-02087-8) - **Stanford Online.** (2023). *Getting Beyond the Hype: A Guide to AI's Potential.*[online.stanford.edu](https://online.stanford.edu/getting-beyond-hype-guide-ais-potential) - **AIM Multiple.** (2025). *AGI/Singularity: 9,800 Predictions Analyzed.*[aimultiple.com](https://aimultiple.com/artificial-general-intelligence-singularity-timing) - **Tech AI Magazine.** (2025). *The AGI Race 2025: OpenAI, DeepMind & Anthropic.*[techaimag.com](https://www.techaimag.com/artificial-general-intelligence/agi-race-openai-deepmind-anthropic-2025) - **ArticleSledge.** (2026). *What is AGI? Complete Guide to Artificial General Intelligence.*[articsledge.com](https://www.articsledge.com/post/artificial-general-intelligence-agi) - **IntuitionLabs.** (2025). *Latest AI Research: GPT-5, Agents & Trends.*[intuitionlabs.ai](https://intuitionlabs.ai/articles/latest-ai-research-trends-2025) - **Russell, S. & Norvig, P.** (2020). *Artificial Intelligence: A Modern Approach* (4th ed.). Pearson. - **LeCun, Y., Bengio, Y. & Hinton, G.** (2015). Deep Learning. *Nature*, 521, 436–444. - **Jobin, A., Ienca, M. & Vayena, E.** (2019). The global landscape of AI ethics guidelines. *Nature Machine Intelligence*, 1, 389–399. --- *Source: [cojocarupetru.info](https://cojocarupetru.info/) | Updated: May 2026* ---