Obligatory disclaimer: I use LLMs extensively for my work. I also remain sceptical and critical.

There is no intelligence in AI

I am deeply reluctant, if not opposed, to using the terms AI or Artificial Intelligence in speaking about ChatGPT, Claude, et al. We are talking about an LLM, a Large Language Model, which are, at their core, essentially just token predictors. This means they are, to put it simply, moderately accurate software that predicts the next word in a sequence. There is no intelligence in the true sense of the word, no “reasoning”, even if the vendors advertise the features as such.

What’s being called “Reasoning” by Claude, or “Thinking” by ChatGPT, does not actually involve any real thinking. In fact, what happens in these modes is simply that the language model first engages in a lengthy monologue in the background before generating the output for the end user. So it isn’t really thinking; it’s merely talking to itself before talking to you. It’s cliché, but it bears repeating: These tools are talking instead of thinking.

Roughly speaking: All these systems do is find the next word that is most likely to follow. The fact that this works so well is an incredible testament to the power of our language, and mind blowing on its own. I am not hesitating to call it the most wondrous technological development of my lifetime. But it also means that every solution an LLM finds has been achieved by taking a detour through the linguistic level. For an LLM, two plus two equals four only because it is statistically very likely that the word “four” follows “two plus two.” It does not have a mathematical understanding of numbers and formulas, merely a fuzzy prediction of linguistic patterns.

That’s probably why LLMs are so good at taking over tasks that mostly involve talking. Management tasks, for example — things like coming up with buzzwords or writing pointless briefings and emails or drafting up PDFs nobody reads that regurgitate known information and will be outdated come next week. And these middle managers then see how easy it would be to automate parts of their own jobs, now mistakenly assuming this could be applied to every other field.

But maybe it can — if you’re willing to pay with a drastic reduction in quality. In a quote often attributed to John Ruskin, it’s said that “There is absolutely nothing in the world that some man cannot make a little worse and sell a little cheaper.” The tech world has seen that with the outsourcing and nearshoring craze of the 2010’s. We’re now seeing it, at a higher velocity than ever before, with the AI bubble. Nearly all forms of digital goods and services can now be offered at even lower cost for even worse quality. And indeed, end consumers, when confronted with a product that is marked as “AI-powered”, will automatically think: cheap and bad.

A recent study demonstrated this effect. Consumers, across all fields of goods and services, will consistently distrust and avoid products marked as containing “AI”.

Mental Malnourishment & Death of Authenticity

I have a confession to make: I am sick and tired of LLM outputs. It has a certain cadence to it that is impossible to not notice. Once you built an antipathy towards it, seeing it everywhere becomes grating.

Everything ChatGPT, Claude, Gemini, LLama, etc produce has been regurgitated over and over again, so often that any nutritional value it might’ve once held is long gone. AI companies crawl the internet, use the data to train their models, the models generate metric tons of output, which is posted to the internet just to be crawled and retrained on, repeat ad infinitum (et ad absurdum (et ad nauseam)).

It’s a cycle of self-digestion. Breaking this cycle has to become a priority for society as a whole; if we don’t start creating and valuing authentic and creative content any time soon, we will starve, intellectually and spiritually.

An anecdote comes to mind, one that might serve as a metaphor for what is currently happening: A family friend, working for UNO, told me about the modern face of malnutrition. In many regions, mostly in emerging economies, people paradoxically suffer from obesity and malnutrition at the same time. Let me repeat that: It has become disturbingly common for large segments of these populations to be overweight and malnourished at the same time. They suffer from a combined caloric surplus and a severe lack of nutrients. Simply put, this is because the “food” available to them barely deserves to be called such, offering nothing but calories and carbohydrates. But it’s highly dopaminergic, making use of addiction mechanisms.

Reading LLM output feels the same. It barely deserves the label of “content” anymore. There’s not a day in my professional life where I don’t receive an e-mail that is several pages of LLM output, but could’ve easily been two sentences of human authored text. Several pages of text, whose actual substances — the nutritional value — is no more than two sentences at best! The ironic part is that this short, two-sentence version of the e-mail exists, but only as the prompt. There is no added value to running your bullet points through ChatGPT. This just results in empty calories the recipient has to digest, so to speak, consuming valuable cognitive energy. Please just send me your imperfect bullet points instead! You’re not doing me, or anyone else, any favors by forcing us to wade through this mental clutter.

That’s why I’m begging you: Please make an effort to be more authentic. Post your photos the way they’ve been taken (color filters are fine). Type your posts and e-mails by hand, or dictate them, just the way you naturally speak and think. And if you want to write a blog post, just do it. Don’t second guess yourself, don’t seek reassurance from a machine.

We’ve now reached a point at which being authentic is a revolutionary act.

Thank you.