In a short novel I wrote about 20 years ago, there’s a scene where Eric K. (the protagonist, a private detective) and his best friend Patrick Z. (a hacker), are approaching the investigation from different angles. There’s a bit of banter between them, then Patrick ends the conversation dismissively, hinting that there’s work to do and they’re just wasting time. “Off you go,” he says to Eric, “As they say, You’re a detective, go detect. I’ll stay here with my machines. [a beat] They’re computers, they’ll compute.”
I was reminded of this exchange when thinking about the current phenomenon of so-called ‘AI psychosis’, where people, believing the many lies of the ‘AI’ industry, think that now computers and other computing devices are capable of thinking, reasoning, understanding, remembering and learning from our ‘conversations’ and their training, and so on and so forth. They aren’t.
There is no intelligence involved behind the blanket term that ‘AI’ has become today. Large Language Models are a useful tool within specific applications, but that’s all there is to it. It involves statistical analysis, it involves prediction. But outside those applications where it can produce useful results, it’s a gimmick, it’s a magic trick. The ‘AI’ industry is doing exactly this — selling you a tool that can be used to perform a magic trick while telling you that what you’re getting is real magic. That magic exists.
Nothing has changed in the nature of computers since their invention. The difference between the first mechanical computers (e.g. Charles Babbage’s machines) and what we have today on our desks is computing power. Nothing else. Computers have become increasingly fast at performing calculations, their hardware has become increasingly refined and sophisticated, their use cases have multiplied, but nothing in their fundamental structure has changed. Today’s computers do not think. Despite what the ‘AI’ industry and ‘AI’ fanatics may tell you, there is nothing ‘neural’ about ‘artificial intelligence’ or computers today. It’s snake-oil salesman marketing terminology designed to misdirect people. It’s like telling you that that mug with Homer Simpson’s face on it is alive and has feelings, while actually it has a thermo-sensitive coating that makes Homer Simpson change expression when you pour hot liquids into the mug.
The ‘AI’ industry has essentially capitalised on that famous Arthur C. Clarke’s quote, which is the third of his three laws: Any sufficiently advanced technology is indistinguishable from magic.
I’ll say that the computing speed of today’s devices is fast enough to trick some people into thinking that such devices are doing more than just performing calculations. But this ‘understanding’ and ‘learning’ and ‘communicating’ are just a faster and sometimes more efficient autocorrect. Yes, the autocorrect you have on your phones and tablets and personal computers.
When you’re typing a message and you see your phone’s OS correctly suggesting the next word after the one you’re typing, that’s not intelligence. The phone doesn’t ‘get’ you. The phone doesn’t ‘know’ you. The system is making a prediction by calculating probabilities. The data sample it’s using for this prediction is everything you’ve typed on the phone since you bought it. Fun anecdote: one of the vintage iPhones in my collection is a second-hand model I purchased from an Austrian seller on eBay. Throughout its previous life, this iPhone was set up in German by a user who only spoke German. I switched the language to English, and the first times I was typing messages and other things, the predictive text engine was essentially shooting random stuff, and ‘guessed right’ only in obvious scenarios, like suggesting ‘the’ after I typed ‘into’.
Now, some will say, But as you kept typing stuff in English, the iPhone got better at understanding you and at suggesting the right words. But that is not learning in the same way a human learns. That is ‘learning’ in the sense that pieces of information have been progressively stored in the device’s memory. When I type “Hey, do you want to meet for lunch…” and the iPhone autocorrect proposes ‘today’, ‘later’, and ‘now’, these are guesses that are calculated every time, from scratch, by going through all the similar sentences I may have written in the past. It’s not knowledge. In fact, the iPhone doesn’t even know what ‘lunch’ — or any other word in that sentence — means. It has only registered that, up to now, 349 previous sentences I wrote ending with ‘meet for lunch’ had the word ‘today’ after ‘lunch’, 188 sentences had the word ‘later’ after ‘lunch’, and 70 sentences had the word ‘now’ after ‘lunch’.
By the way, Large Language Models don’t even see those as words, but numbers, tokens.
When you ask ChatGPT or Claude or your fave chatbot something — whether a simple query or what sounds like a more sophisticated one — these tools are not sentient, they’re not ‘built different’. They process the data you passed to them, which is analysed, computed, and spat out in a friendly, anthropomorphised language. These tools are designed to sound like real, responsive, agreeable, trustworthy assistants. But the emphasis should not be placed on trustworthy — it must be placed on sound like. Because they’re far from infallible, and their output should always be fact-checked. This is the main reason I don’t waste one second of my time with this stuff and insist on doing my searches and research myself. I want to look at the sources, I want to cross-check and do my due diligence. I don’t need shortcuts — I’m not in a hurry. And if you’re doing important and meaningful work, you ought not to be in a hurry, either.
In a conversation with friends on this subject, one put this in a way that looks pretty straightforward to me: ‘AI’ isn’t a surgeon (or any other professional); it’s an actor pretending to be one. If you’re rushed to the emergency room, you don’t want to be checked out by George Clooney or Noah Wyle, no matter the amount of medical terms they have memorised and they’re able to assemble on the spot — you want a real doctor.
Going past the trick of using anthropomorphised terms to describe LLMs’ modus operandi (like ‘thinking’, ‘reasoning’, ‘understanding’, ‘remembering’, ‘learning’, ‘hallucinating’, and so on), another unfortunate approach of ‘AI’ advocates and apologists is that to demonstrate how ‘AI’ behaves in a human way, they reduce every complexity of human processes and every complex facet of human identity to the cartoonish and vague approximation provided by these ‘AI’ tools. They will tell you that no, an LLM really learns because what is human learning if not an accumulation of data and a probabilistic extraction of such data whenever necessary. (Hint: it doesn’t work that way). What is a human brain if not a very large neural network, they say. (Hint: it’s not. It’s so, so, so much more. Just because you put the term ‘neural’ doesn’t mean that every node has the same functional ability of a human neuron. In fact, you could say that it takes a very deep neural network to barely do something a single biological neuron can do. As Brooks Brown says in this video, “That thing we named neuron [in a neural network] is to an actual neuron what a stick figure is to the human body”).
Another approach I profoundly dislike is the sheer abuse of the term ‘AI’ even when it’s really unwarranted. ‘AI’ advocates will tell you that there have been several advancements in medical research (or other fields) thanks to ‘AI’, so ‘AI’ is good and it’s the future and similar bullshit. It wasn’t ‘AI’. There wasn’t a machine that woke up one day and autonomously decided to absorb the most pertinent and up-to-date cancer research and made a discovery based on deduction and intelligence. All the data was fed to computers that made calculations also utilising LLMs and analysed the data in a relational manner and made educated (i.e. by processing an absurd amount of information) projections. It wasn’t artificial intelligence, it was computers that computed. It’s always been computers computing. We have more refined tools today, but that’s it. The ‘AI’ industry likes this kind of misdirection because it puts their tools in a far better light. It builds trust in the snake oil they’re selling. And by the way, I’m not saying that LLMs aren’t useful per se. But they’re just a tool, not the Second Coming of Computer Intelligence, and certainly not the Prelude to the Singularity that will produce, one day, out of nothing, a sentient machine. This is the snake-oil part. Use the tool, if you like, but think critically and don’t fall for the narrative. And for the love of all that’s good, don’t outsource your thinking to the tool. Do your fact-checking, do your due diligence.
And that’s really it. Computers perform calculations using potentially immense amounts of data, at exceedingly high speeds. They compute. Any other verb is basically an illusion. Or manipulation, however we want to call it.
Personal recommendation
The number of YouTube creators I have chosen to support can be counted on one hand. Brooks Brown is one of them. On his channel (mainly about philosophy, but also technology, with entertaining ‘After Hours’ streams) he has started a new series of 15 videos called The Lies of AI, where he demystifies ‘AI’ much more articulately than yours truly. At the time of writing he has published seven of these videos, roughly 10–12 minutes long each. This is very good stuff, and I strongly suggest you subscribe to his channel and watch these. Here’s a link to the first video of this series.