<< See all articles

Resisting the mediocrity machine: in defence of expertise

We have been training AI models long before LLMs. This is how we learned that fancy algorithms are no substitute for careful research design.

We have been training AI models long before LLMs. This is how we learned that fancy algorithms are no substitute for careful research design.

The future belongs to the most tenacious among us. Those with enough persistence to corner a difficult problem and bring it down with a series of carefully dealt pinpricks.

This becomes increasingly clear as we progress further into shortcut culture. This over-indexing on expediency and output that turns us to machines for brute-force solutions rather than sitting with our problems. Yes, toil and trouble is uncomfortable. It is also, oftentimes, the point.

The longer and harder we outsource our thinking, the more we risk dividing ourselves into two camps: those who have – versus those who have not – retained their capacity for dealing with complexity.


Visions of language as co-worker

In 2019, I had my first experience with the kind of neural language models that form the basis of modern generative AI. It was during a machine learning course at the University of Copenhagen where the models of the day featured as nothing more than a sideshow curiosity. At the time, I had no idea how fascinated I would become with the technology. Least of all did I expect the intensity of the frustration that its widespread public dissemination would instil in me.

This was 3 years before OpenAI let slip the bots of war on an unsuspecting populace. Ironically, GPT-2 – the predecessor to ChatGPT – was withheld from the public for an extended period of time, due to concerns of abuse and the potential to spread rampant misinformation.¹

The generative language models that I worked with initially were rudimentary. They regurgitated garbled copies of their training data, and not much else. Nonetheless, as a computational social scientist, I became interested in using the mathematical foundations behind these neural models to help identify patterns in text. Over the coming years, I worked on developing tools to augment human analysis of text corpora so large that deep reading is impractical or downright impossible.

But even as the language models grow more convincing, their fundamental weaknesses remain. Language is nuanced, and even a Large Language Model (LLM) is basically just a compression of a text corpus, which means that details necessarily begin to fray at the edges. The model is similar to a grainy photocopy of an encyclopaedia. Some facts are luckily preserved; other entries are gibberish. Slap a servile, yet confident chat interface on top of that, and chaff starts looking an awful lot like wheat.


Sliding towards mediocrity

In statistics, the term regression to the mean (or regression towards mediocrity, as it was originally coined by Sir Francis Galton) describes the phenomenon that an unusually extreme sample from some population will tend to be followed by a more moderate one.² In terms of generative AI, we are inching towards mediocrity in at least two ways:

Specifically, these models tend to produce 'average' output. A bland remix of content ingested by the machine, presenting trivial and unoriginal ideas. Prose scrubbed of any idiosyncratic quirks of individual expression.

More generally speaking, the process of frictionless prompting relieves us of the need to reason. Folks without any particular expertise in a given field may feign a certain degree of proficiency without learning much. Conversely, those already competent may be tempted to settle for outcomes so half-baked that they constitute an affront to their actual expertise.


Democratising expertise: an oxymoron

The optimist's angle here would be that generative AI affords more equal access to specialised skills that were previously reserved for those somehow specially initiated through talent or training. Comprehension of computer science is no longer a prerequisite for building software, so the argument goes. Nor does lack of writing practice bar anyone from producing (procuring?) literary works.

By its very definition, however, I would argue that expertise cannot be democratised. It requires a certain level of discernment that is built from experience and obsession. Most new ideas that any of us have on this green Earth are so mundane or terrible that they deserve to perish on the drawing board. The costs (financial, cognitive, temporal) associated with putting an idea into practice acts as a filter that ensures our most absurd schemes are nipped in the bud.

Without such quality assurance, we are now faced with a deluge of slop. With AI agents entering the arena, the number of new mobile apps, for instance, has exploded, all while the number of apps with significant usage remains stagnant.³ The supply of mediocre digital content has never been greater.

Pretending that we are levelling the playing field is an insult to those experts who are left to justify their Fingerspitzengefühl against machine output – as if the quality of software architecture, legal assessments, or poetry are reducible to matters of preference.


With friends like these...

There is no shortage of actors with significant vested interests in portraying generative AI as an indispensable companion to anyone who wishes to escape irrelevance as the gears of industrialisation grind once more. Tech broligarchs keen to boost their pre-IPO valuations. Business leaders who fancy themselves trailblazers. Consultants on the prowl for shiny new trinkets to sell or for headcount to reduce. All eyeing profits by stoking the fear that one may be stranded on the wrong side of history.

In truth, I am hopeful for the future of AI technology. Especially on behalf of the diverse flavours of artificial intelligence research in both academia and industry which have nothing to do with chatbots or LLM agents. But a certain bitterness has taken root in me, seeing how the loudest proponents of generative AI dismiss its disadvantages in bad faith. The more we cling to those narratives, the more helpless we render ourselves.⁴

It is more lucrative to sell a dream than to have a nuanced conversation about pros and cons. And prompting oneself 80% of the way to a solution that nearly works does seem impressive – until you realise that those final 20% are both the most difficult and the most critical.

Generative AI solutions are becoming commoditised and actually ROI-positive use cases are trickling down into existing enterprise software. So, if you are a business executive, consider whether your organisation truly needs to be on the frontier of AI implementation – and think carefully about whose interests you are letting in the door if the answer is 'yes'.

What you probably will benefit from, AI or not, is a cadre of employees who can think for themselves – and those may soon be in shorter supply than you would like. So do not forget to cherish the nerds on your team who insist that thoughtfulness and conscientiousness is what drives value. You might still need them on the other side.


Notes

  1. OpenAI (February 14, 2019): https://openai.com/index/better-language-models/

  2. Sir Francis observed that taller parents, on average, had children shorter than themselves. He misinterpreted this as a biological fact rather than as a statistical artefact, but the name is cool, isn't it?

  3. Financial Times (June 5, 2026): https://www.ft.com/content/8e9ae7a4-7209-4e2c-aa36-f3af77d6ce1f?syn-25a6b1a6=1

  4. I suspect that this is equally true of the resignation that follows from doomer rhetoric as well.