The Eloquence of Mediocrity
Let’s start the year with an article different from the usual, useless rigmarole of tear-jerking posts about resolutions and goals that nobody really cares about—the kind you’ve been seeing on social media since midnight on January 1st.
Since I was a boy, I have always been scared by the idea of being a “mediocre” person. Moderately good at school, in sports, aesthetically, in video games… in short, falling into that large and vague category of “yes, but…”. An absolutely anonymous person, whose absence is not even felt.
Being average is objectively frustrating (a bit like when coming home from school all happy for having taken a decent grade in a subject difficult for me, my mother started with an eloquent “Yes, but… what did X get?”), it is like seeing the sun appear behind a mountain, while you observe from the valley. You know that beyond there a “better” world awaits you, but how deep is the valley to get there…
Scared by the effort required to emerge, we often settle for completing our daily tasks. Whether it’s closing a ticket, getting rid of that annoying 5:30 call, of a client calling you for nonsense while you finally managed to concentrate… for many now the important thing is to get to the end of the day with as little effort as possible, and it matters little if these tours de force leave us nothing but tiredness, without a reward for having defeated “the monster”.
The advent of AI has brought this phenomenon to levels never seen before. Having made such a powerful tool available to everyone has led to its most natural consequence: many of us have simply turned off our brains.
Why bother learning to do something when I can take a file, feed it to ChatGPT (or whoever), and see my 2 hours of work done in 10 minutes? Among other things, the output is often of higher quality than what the person in charge would have produced (medium-low quality output), already enriched by comments (in case of code) and presented tremendously well on a lexical level.
A clamorous win-win, right?
No, absolutely not.
I make a premise: in this article I will voluntarily ignore all the part related to why it is crazy (or grounds for dismissal for just cause, you see) to shoot company documents or containing sensitive data to an external LLM. It is not important for the purpose of the post, so let’s go straight.
The Problem Is Not What You Think
The other day I was scrolling LinkedIn and I saw like the twentieth post starting with “In a rapidly evolving world…” followed by a series of perfectly formatted insights on how AI is revolutionizing everything.
And there I stopped. Because that post… I had already read it. A hundred times. Same different words, same cosmic emptiness inside. A mediocrity so lucid, so well articulated, so damn convincing as to seem almost intelligent.
And this is exactly the trap.
AI is not making us smarter. It is making us eloquently mediocre. And the worst thing? We don’t even realize it.
When Form Eats Substance
Ok, let’s take a step back. Ted Chiang (the one who wrote “Story of Your Life”, from which they made Arrival) defined ChatGPT “a blurry JPEG of the web”. A compressed, blurry, but statistically representative version of everything that has ever been written on the Internet.
Nice, right? Too bad almost everything on the Internet is mediocre. Indeed, worse.
There is this thing called Sturgeon’s Law which says that “90% of everything is crap”. And guess what these models were trained on? Exactly. On that 90%.
But here comes the brilliant part of the deception: these systems have a dialectical capacity that we humans generally do not possess. They take that 90% of mediocrity and give it back to you with the fluidity of a TED Talk, the structure of an academic paper, the tone of an industry expert.
Emily Bender called them “stochastic parrots”. They don’t understand shit. They predict the next token based on probability. Period. There is no understanding, there is no semantics, there is no intention.
Yet they work.
Like… take a junior who has to write a technical document. He gives it to ChatGPT. He gets an impeccable text: perfect grammar, logical structure, appropriate terminology. If you don’t know the subject, it looks written by a senior with 10 years of experience.
But if you know it… you see the cracks. The safe but empty generalizations. The assertions technically correct but saying nothing. The absence of that “click” that comes only from true understanding.
It is the difference between knowing music theory and knowing how to play an instrument. You can know all the scales in the world, but if you have never put your fingers on a keyboard…
The Loop of Death: When Mediocrity Feeds Itself
And now comes the part that should make us shit ourselves.
There is this phenomenon called “Model Collapse”. It happens when an LLM is trained on content generated by other LLMs. And it is exactly what is happening now, in real time.
Think of a photocopy of a photocopy. Each iteration loses quality, amplifies errors, flattens details. Now apply this mental image to the Internet.
Millions of articles generated by AI end up on Google. LinkedIn posts written by AI. Code on GitHub copied from Copilot. Answers on Stack Overflow generated by ChatGPT. And all this ends up in the training datasets of future models (see The Curse of Recursion and AI Collapse).
We are literally polluting the well from which we drink.
The model that will be released in a few years will have been trained on an ocean of mediocrity generated by AI. And it will produce even more refined outputs, even more convincing… even emptier.
Hallucinations: When AI Lies Without Knowing It
Let’s talk about hallucinations. Those times when your favorite tool cites a study that doesn’t exist, invents a date, gives you a completely wrong fact but presented with disarming confidence.
The first time it happened to me I was doing research on a security protocol. The AI cited a 2019 paper, with authors, title, everything. It seemed legit. I went to look for it… it didn’t exist. Nothing. Zero. Invented out of thin air.
And here comes the beauty: hallucinations are not a bug. They are a feature. Or rather, they are an inevitable consequence of how these systems work.
An LLM doesn’t have a database of “true facts” from which it draws. It only has a gigantic matrix of parameters that tells it which words make statistical sense after others. When you ask it to cite a study, it isn’t “remembering” anything. It is generating a string of text that looks like a citation.
If that citation really exists in the training, great. If it doesn’t exist, it invents it anyway. Because it has to complete the sequence. It’s like a student who studied poorly and at the exam shoots an answer to the prof that sounds plausible, hoping that he doesn’t try to dig deeper.
Small difference: the student knows he doesn’t know. The AI doesn’t even know it exists.
And to notice the hallucination? You have to know the subject. You have to have the expertise to understand that that citation is false, that that datum is invented, that that reasoning doesn’t track.
Without expertise, AI becomes an echo chamber that amplifies your ignorance masking it as knowledge.
Dunning-Kruger 2.0: The Perfect Illusion
There is a side effect of all this that is even more perverse: the amplified Dunning-Kruger effect. You surely know it, since 2006 onwards they have smeared it (-cit) in any semi-interesting post you have read.
Let’s simplify the concept: “Those who don’t know, don’t know that they don’t know”. And AI provides them with the perfect illusion of competence.
Like… a junior developer who uses Copilot to write code. The code compiles, seems to work, even has comments. The junior is convinced he has done a great job. He feels competent. Maybe he even shares it, like “look what cool thing I made”.
Then a senior arrives, takes a look, and sees: memory leaks everywhere, security issues, bad practices, fragile logic that collapses at the first edge case. But the junior doesn’t know. He can’t know. He doesn’t have the experience to see what is missing.
And this is the real revolution… not that machines think, but that we can stop doing it without realizing it.
AI lowers the barrier to entry, but does not eliminate the need for expertise. Indeed, it makes it even more critical. Because now you also have to know how to recognize when the AI is wrong. And this requires more competence, not less.
It is the paradox of democratization: making a powerful tool accessible to everyone does not make everyone an expert. It only makes it harder to distinguish experts from mediocre ones.
Centaurs, Not Parrots
Kasparov, after being defeated at chess by Deep Blue, did not give up. He invented the concept of “Centaur”: human + machine collaborating. And he proved that a grandmaster with a good computer beats everyone.
AI is an amplifier. It amplifies what you have. If you have competence, it amplifies it. If you don’t, it amplifies the void.
The difference between an expert and a beginner is not access to the tool. It’s knowing when to trust and when not to. It’s having the “nose” for bullshit. It’s that little voice that tells you “wait, this doesn’t add up”.
And that little voice develops only with true experience, not delegated.
The Useless Race for the Best Model
There are all these people obsessed with the latest SOTA (State-Of-The-Art, for those who don’t chew the acronym) model.
“Have you tried the latest GPT?” “No no, Claude Opus X is better for coding” “But have you seen the Gemini benchmarks?”
Leaving aside the uselessness of certain benchmarks, by now all top models are excellent. The differences are marginal for the average user. What really makes the difference is how you use them: how you structure prompts, how you validate outputs, how you integrate them into your workflows, how you maintain control.
The bottleneck is not the model. It’s you. Your understanding of the domain, your ability to ask the right questions, your critical spirit.
Chasing the latest model is like buying the most expensive camera equipment thinking it will make you a better photographer. Spoiler: it doesn’t work that way.
The Problem Is You
Let’s go back to the starting point. Is AI making us smarter or more mediocre?
It depends on you.
If you use AI to avoid thinking, you are condemning yourself to mediocrity. You are producing content that sounds good but is empty. You are contributing to the information pollution of the Internet. You are becoming part of that 90% of Sturgeon’s Law.
The real problem is not even the quality of the output. The real problem is that eloquent mediocrity is more dangerous than evident mediocrity. Because it camouflages itself. Because it convinces. Because it makes you believe you are competent when you are not.
And this is the real trap: not that machines become smarter, but that we can become stupider without realizing it. That we can delegate thinking and still feel productive. That we can stop learning and continue to produce “decent” output.
But “decent” is just another way of saying “mediocre”. And mediocrity, even when it is eloquent, remains mediocrity.
Luciano Floridi (see previous article on The Thinking Game) talks about the “divorce between Agency and Intelligence”. For millennia, if you wanted to do something intelligent, you had to be intelligent. Now no. Now you can act with efficiency without understanding, produce results without thought, convince without knowing.
And perhaps this is the real danger. Not Skynet taking control. But us who slowly, comfortably, inexorably… stop thinking.
In a world of parrots… be Centaurs.
Vittorio
References
- ChatGPT Is a Blurry JPEG of the Web - Ted Chiang (New Yorker)
- Sturgeon’s law - Wikipedia
- On the Dangers of Stochastic Parrots - Emily M. Bender et al.
- The Curse of Recursion: Training on Generated Data Makes Models Forget
- AI Collapse: The effect of synthetic data on generative models
- Unskilled and Unaware of It (Dunning-Kruger Effect)
- Garry Kasparov on AI, Chess, and the Future of Creativity