Ivo Vegter: Stop fearing Artificial Intelligence. Start fearing its would-be masters

Artificial intelligence is rapidly becoming the defining technology of our age, promising extraordinary gains in productivity, knowledge and innovation. Yet the greatest questions it raises may have little to do with machines themselves. In this thought-provoking analysis, the author cuts through both the utopian hype and apocalyptic fear surrounding AI, exploring its impact on truth, work, power and liberty. From the risks of information manipulation to the geopolitical struggle for technological dominance, he argues that the most serious threat is not what AI might do, but how governments, corporations and elites may seek to control it. A provocative read..By Ivo Vegter*.AI is the most consequential tool of our age. The greatest danger it poses is not what it will do to us, but what governments will do to save us.Last weekend, I had the honour and pleasure of joining the CEO of the Institute of Race Relations, John Endres, in a discussion about artificial intelligence (AI) before the IRR’s Council.We only had an hour or so, and covering all the angles could have taken all day. My notes captured about twenty largely disorganised thoughts I had on the subject.This is an attempt to synthesise those notes into something resembling a coherent whole. Some of it will be familiar from previous columns; much of it will not.A bit of relevant biography might be appropriate here. I have a little background in computer science and applied mathematics from my student days. I spent a decade and a half working as a technology journalist. I also kept my wits fairly sharp as a result of more recent geekery in my free time, some of which involved maths-related coding projects and taking university-level courses in topics adjacent to AI.I remain an enthusiastic amateur, however. The details of how large language models (LLMs) work get very complicated very quickly. Though I have a fair grasp of what’s going on under the hood (and how other types of AI systems work), I cannot call myself an expert.That disclaimer out of the way, let’s get going..Two ways to be wrong.There are two ways to be wrong about artificial intelligence, and most people manage both at once.The first is to over-estimate it.Large language models are not sentient. They do not understand what they are doing. They do not “think” in any sense of the word that would survive a moment’s scrutiny. They take your input and assemble a statistically plausible continuation from a vast aggregate of training data.They are a mirror, not a mind. When an LLM dazzles you, it is reflecting human intelligence back at you, and inviting you to marvel at human cleverness.To approach anything like human ability to “think”, an AI would need a model of much more than just language. It would need a world model, sufficiently detailed that it could make inferences about cause and effect in the world in the same way that we discover how the real world works.They’re working on it, but a sufficiently detailed world model will take a sight more compute than even the largest AI companies could throw at it in the foreseeable future.The second error is to under-estimate how even the limited capabilities of LLMs will change the world.Even granting every limitation, the rise of LLMs signals an extraordinary upheaval in the global economy. Their impact will be disruptive and transformative, freighted with immense upside and tremendous risk. To say that AI is “just” a search interface to humanity’s aggregated knowledge is true, and also rather like saying the printing press was “just” a faster way to copy a page, or the internet was “just” a better way to communicate.Both things are true at once. The trick is to hold them together. Refuse the breathless hype of the singularity prophets. Refuse the apocalyptic warnings of human extinction. Refuse also the comfortable dismissal of those who insist it is nothing more than a parlour trick to pump up stock prices.It is none of these. AI – by which I mean LLMs unless otherwise specified – is a tool.It is the most powerful general-purpose tool since the computer itself. It raises urgent questions about how that tool will get used, for good or for evil. Those questions are less about technology, and more about power, knowledge, and liberty.Let me take them in turn..The epistemic threat.What genuinely keeps me up at night about AI is not whether the machines are sentient and whether they want to kill us. What worries me is the collapse of our ability to tell what is real.I have written about this before, and I become more convinced of it by the month. When synthetic text, images, and video become cheap, and convincing even to experts, the marginal cost of producing plausible falsehood collapses toward zero.The scarce resource is no longer information. We are drowning in information. The scarce resource becomes verified provenance – knowing who made, did or said a thing, and whether they can be trusted.This is a problem of authentication, not abundance, and our institutions are nowhere near ready for it.There is a fashionable counter-thesis that LLMs are quietly converging our discourse – narrowing the range of public opinion, smoothing everyone toward a bland statistical centre, in contrast to social media’s well-documented tendency to fragment and polarise.If true, that raises questions about the power that a handful of Big Tech companies have over the direction and limits of public debate.I am not convinced, however. Depending on whose model you use, and how you prompt it, an LLM is perfectly capable of telling you exactly what you want to hear and leading you down an ever more extreme garden path.Remember that LLMs draw upon the collected writings available on the internet, including not only books and scientific papers, but also the archives of social media platforms. The sycophantic chatbot that flatters your priors is no improvement on the echo chamber; it is the echo chamber, but now with a more convincing, authoritative voice..Epistemic dependence.This connects to a more subtle danger: epistemic dependence. As people offload their reasoning, their memory, and their judgement to systems they do not understand and cannot audit, the muscles of critical thought and independent verification atrophy.We have already seen the harbingers. Doctors have long had to contend with “Google patients”, who arrive for their appointment having diagnosed themselves, with demands for specific treatment. They see doctors as mere bureaucrats with the power to prescribe. Now those patients arrive having been diagnosed by Dr ChatGPT, more convinced than ever that they know what they’re talking about.Multiply that across every domain of expertise and you have a quiet centralisation of cognitive authority in a handful of models owned by a handful of firms.This raises the question: how should we use AI?In my experience, AI can make a lot of cognitive drudgery far more efficient. An LLM with the ability to search the web can conduct desk research much faster and more comprehensively than a single human will ever be able to do.When used carelessly, they can lead you to error, or reinforce your confirmation bias, much in the same way that Google Search, which remembers your previous interactions, serves up just what it thinks you want to see, or sometimes serves up internet nonsense.When used judiciously, however, it can highlight your blind spots, and alert you when what you thought were novel questions already have well-publicised answers..When to use AI.In many specialist fields, especially those where advanced pattern recognition is required, AI can produce faster, more reliable results. It should absolutely be deployed in such fields, as long as the price of AI failure is lower than the documented price of human failure.In fields that depend on formal, precise and structured language, such as mathematics, computer programming and law, LLMs are extraordinarily competent. They cannot produce novel knowledge, but they are very good at finding relevant existing information, and finding novel connections between established facts.In some other fields, however, AI will almost always produce worse outcomes.Education is perhaps the most consequential field in which injudicious reliance on AI can have serious negative consequences, especially on a learner’s memory, creativity, reasoning skills, critical thinking, understanding and mental facility with the subject matter.AI isn’t going anywhere, however. We cannot not use AI.It is a tool that, for all its epistemic risks, affords its users great competitive advantage.Honing our ability to use AI without falling victim to its hidden biases and hallucinations will be essential for all of us, one way or another..Geopolitical points of failure.Frontier AI requires capital, compute, and data at a scale only a handful of companies and states can muster. They require literally trillions in investment in hardware and energy supply.Whoever controls those inputs gains structural leverage over economies and information flows alike. That is a fact about power, not merely about technology, and no amount of cheerful talk about “democratising AI” should detract from it.The hardware choke points are worth dwelling on, because they have geopolitical implications. The entire edifice of advanced AI rests on a startlingly small number of companies.Two matter above all others. The first is ASML, in the Netherlands, the sole maker of the extreme-ultraviolet lithography machines without which the most advanced chips cannot be made. It sits, conveniently, in a stable and friendly jurisdiction.The second is TSMC, in Taiwan, which actually fabricates those chips – and which sits in precisely the opposite kind of jurisdiction. Whoever controls TSMC will hold a marked advantage in the build-out of AI infrastructure. Whether Intel’s fabs, or anyone else’s, can ever rival it remains to be seen.It does not take a strategist to notice how exposed that single point of failure is.A great deal of the world’s technological future depends on the continued security of one island that a nuclear-armed neighbour regards as a renegade province, at a moment when the principal guarantor of that island’s security has been depleting its munitions stockpiles in wars of questionable strategic import.I do not say this to be alarmist. I say it because anyone modelling the future of this industry who leaves Taiwan out of the model is not modelling the real world.And then there is the money. I should disclose that I am a poor forecaster of markets, so weigh this accordingly – but every instinct I have about current AI company valuations yells bubble.The industry is consuming literally trillions of dollars in capital investment, a great deal of which will be obsolete hardware within a year or two. I cannot yet see how the anticipated revenues from generative AI come close to justifying the sums being committed.I may well be wrong; revolutions have a way of confounding the cautious. But it is possible to believe wholeheartedly in AI’s transformative power and still suspect that a great many of the companies racing to provide it are valued like they have already won a race that has barely begun..Economic anxiety.The economic anxiety that gets the most airtime is the fear that AI will take our jobs.It will. Good.I have made the full case previously, so I will be brief here. Every wave of automation for four thousand years has been met with the same prophecy of mass unemployment, and every time the prophecy has failed, because human wants are effectively unlimited and the quantity of work to be done is never fixed.Destroying jobs is not a bug of progress. It is the engine of it. The displaced typists and clerks of the second half of the 20th century did not starve; some took retirement, and the freed labour of the remainder built a new way of doing business as computers and the internet permeated every office.What is genuinely different this time is not the destination but the route. AI-driven automation will target cognitive work that previously felt insulated – the very knowledge-economy jobs whose holders always assumed they were safe.The dislocation’s speed and concentration, rather than any net loss in aggregate, is what will drive the political backlash. People do not riot over productivity statistics; they riot over their own upended lives, and they vote accordingly. That is a real political problem, and pretending it away serves no one.The fear of displacement should be recognised, but it should not be the headline. It is a real effect, but it is also the basis for new, more productive employment.Sébastien Krier, policy development and strategy lead at Google Deepmind, puts the underlying logic well: as long as the combination of human-plus-AI yields even a marginal gain over AI alone, the human retains a comparative advantage.I agree. Tellingly, in coding – the field where LLMs have proved to be the most competent – employment has risen, as firms discover they can attempt ever more ambitious projects with the help of AI.It is striking, then, that society has turned so sharply against AI before its effects have really been felt at all. That is not prudence. It looks a great deal more like paranoia..AI cannot be governed.Here is where I part company with almost every government on earth, and a fair number of the AI companies too.AI is not governable. I mean this in a strong sense: not that regulation is unwise, but that it is structurally futile, and illiberal in proportion to its seriousness.Start with the technology. The classic illustration is Asimov’s Three Laws of Robotics – a robot may not harm a human or through inaction allow harm; must obey orders except where they conflict with the first law; must protect itself except where that conflicts with the first two.People remember them as a blueprint for safe AI. They were nothing of the kind.Asimov devised them precisely so that he could spend years, across a multitude of stories, demonstrating that they were inconsistent, gameable, and prone to catastrophic outcomes that were simultaneously lawful and monstrous.The Three Laws are not a solution. They are a literary proof that rule-based safety cannot work.The deeper point is that the very thing which makes LLMs useful – their capacity to generate unanticipated, non-obvious, non-deterministic output – is exactly the thing that regulatory guardrails will suppress.You cannot have the upside without the unpredictability, because they are the same property. The more tightly you regulate AI, the less useful it becomes. Recent research shows that improving a model’s “safety” measurably degrades its accuracy.Safety, in the regulator’s sense, is not free. It is paid for in capability, and every legitimate user pays that cost..Censorship as safety policy.Then there is the politics of it, which is worse. Deciding what a model may refuse, whose norms it encodes, and which trade-offs it makes are not engineering questions. They are moral and political choices about contested values that require someone to choose for everyone.From a liberal standpoint, the prospect of governments making those choices – deciding, at national scale, which ideas an AI may help you explore and what tasks an AI may perform – should be deeply alarming. That is not a safety policy. It is censorship.And it does not even work on its own terms. The mathematics behind AI is published. The data is everywhere. Anyone, in principle, can build a model outside any government’s control.Trying to regulate AI is like mandatory age-verification laws that are supposed to keep pornography away from teenagers: a great deal of intrusive machinery that mostly inconveniences the law-abiding, and, being easily circumvented, fails just when it is needed most.As with any intervention in the market, AI policy is industrial policy in disguise.Export controls on chips, subsidies for domestic compute, and “sovereign AI” initiatives might reshape global supply chains, but they always do so in favour of incumbents. Framing these as safety or competitiveness measures obscures that states are picking winners and entrenching national champions.Moreover, any jurisdiction that shackles its domestic AI industry simply cedes the advantage to its geopolitical rivals, who will not be so scrupulous.This is why calls for regulatory AI safety regulation should be met with scepticism..Safety policy as anti-competitive strategy.When an incumbent demands rules, ask who the rules will benefit, and who they will bind. Anthropic’s much-publicised plea to slow AI development – like xAI’s similar call a couple of years ago – is, whatever its sincerity, a call to stall rivals.A company’s demand to be regulated is invariably an anti-competition strategy disguised with a public-safety halo.Anthropic’s stated dread of recursively self-improving AI is probably overblown in any case. There is no reason to believe that AI will escape human control; we always control the power switch. The more plausible near-term outcome of unsupervised iteration is that models become less capable and more hallucinatory as they begin to ingest the slop that previous models have spewed across the internet.This is pure regulatory capture: “responsible AI” rules, written with industry input, that freeze today’s oligopoly in place and price out tomorrow’s challengers, all under the banner of keeping us safe.The historian Niall Ferguson has supplied the elite version of this instinct, lamenting that we will look back “with incredulity” on a period when so dangerous a technology was offered to an unsuspecting public at subsidised prices, and insisting that AI was never a consumer product and should never have been served up as a cheap chatbot.This is elitist, statist, illiberal hogwash. Strip away the patrician phrasing and it amounts to a claim that powerful tools should be reserved for the right sort of people, with governments holding the keys.AI is a tool. Like every powerful tool, it can be turned to good or ill, and every measure designed to prevent the ill will necessarily strangle the good.The cotton gin, the printing press and the internet were all “dangerous” by this logic. We are immeasurably richer for not having locked them away..Open questions.None of this is to wave away the hard problems. Two questions in particular are real and unresolved.The first is liability. When an AI system causes harm, responsibility is diluted across developers, deployers, and users, and without a clear allocation the result is either chilling over-caution or a vacuum in which no one is answerable.This question becomes more pointed as autonomous, AI-run companies become real. Argentina is already drafting law to recognise them.Such a company will have income and assets, and so can in principle be sued like any other; but the law has a great deal of catching up to do. It should do that catching up carefully, as harms emerge, instead of pre-emptively legislating against imagined future harms.The second is humility. The honest position on AI is that both the utopian and the apocalyptic forecasts rest on thin extrapolation from very little evidence.Sound strategy does not bet the house on either. It hedges across a range of outcomes, and above all it refuses to trade away our institutions and our liberties to purchase protection against speculative fears about a future nobody can actually predict.Do not over-estimate the machine, and do not under-estimate it. Take its disruptions seriously and its prophets sceptically.When someone arrives promising to keep you safe from it – whether it is Minister Malatsi with his seven new bureaucracies, a celebrity historian who wants to keep the power of AI out of the hands of plebeians, or a frontier lab that has discovered a sudden enthusiasm for regulation – keep a hand on your wallet and an eye on your liberty.Respect the machine, but do not fear it. Fear the people who want to govern it..*Ivo Vegter is a freelance journalist..*This article was originally published by Daily Friend and has been republished with permission..Sign up for your early morning brew of the BizNews Insider to keep you up to speed with the content that matters. The newsletter will land in your inbox every morning on weekdays. Register here.Support South Africa's bastion of independent journalism, offering balanced insights on investments, business, and the political economy, by joining BizNews Premium. Register here.If you prefer WhatsApp for updates, sign up to the BizNews channel here.

Ivo Vegter: Stop fearing Artificial Intelligence. Start fearing its would-be masters

Related Stories

Cambridge historians see SpaceX and think East India Company — and the parallels are alarming

Kenneth Rogoff on BizNews: AI gold rush has winners and losers — Africa not in the running

Is the AI arms race about to outrun human judgment? The Economist thinks so

Dell takes aim at Apple with a $699 touch-screen laptop