It is, I think, rational to prefer reading human-written texts over AI-generated texts of similar or better quality. I will outline four reasons for this (there are likely many more). None of these reasons will require LLM texts to be bad. This matters because LLM writing, although irritating at times, is probably already better than we give it credit for: we don’t notice all the times it is good enough to fool us into thinking it is human written. As Benjamin Todd from 80,000 hours put it on X: “Plastic surgery is most noticeable when it's bad, so it's more widespread and successful than it looks. Same with AI writing.” Furthermore, LLM writing will get much better very fast. So, any defense of human writing that relies solely on its superior output would be short-lived. Here’s why, regardless of how good AI writing becomes, you would often be justified in your taste for human-written content.
Publishers of LLM generated content face no costs, they do not need to select which ideas deserve their reader's time
First, attention is a scarce resource because life is short. And a text has this peculiarity: we discover whether it was worth reading only after reading it (i.e. we discover if it was worth reading only by spending the scarce resource in question). We must therefore choose what to read based only on available signals.
Until a few years ago, the existence of a text was in itself a signal. The fact that someone had paid days or weeks of his life to express an idea was proof that at least one person had judged that idea worth hours of a finite life. The cost of writing forced authors to select, from everything they might have said, the things that seemed to them most worth saying.
Today, publishers of LLM generated content do not need to select which ideas deserve the reader's time. They can publish everything (it costs them only a few seconds). Facing a block of text, the potential reader cannot be sure that anyone has judged it worth anyone's time; he cannot even be sure that someone has read it! This is why a human generated text is more likely to be worth his time, and it is rational for the reader to choose to allocate his attention to this kind of texts.
LLMs insert noise between the author's mental state and the reader
Second, we look in language for more than interesting sequences of words: we read to connect with other minds. Language developed because man is a social animal who needed to communicate. One of language’s functions is to make one's mental states accessible to others. Notice that when you read an argumentative text, you often feel something for its author: recognition, gratitude, the relief of not being alone in thinking what you think, or irritation, the itch to lay out the counterargument to the author, etc. Reading is (partly) a social act. (Of course, this doesn’t apply to all forms of texts, we probably don’t care if the weather report is AI generated).
And the trouble is that an AI text (even carefully prompted) never reflects exactly what its “author” meant to say. Between the intention and the result stands an LLM that fills in the implicit in its own way, chooses one framing of an argument over another, adds or removes a nuance, adds a stylistic effect that shifts the emphasis (and AIs love stylistic effects). Because LLMs are probabilistic in nature, you even know that had the “author” prompted his AI 10 seconds later, the text he would have asked you to read would have been different. So, by construction, the result is not his own version. (This does not apply to authors who proof-read their texts with LLMs and treat the model’s suggestions as editor's notes, going back to their draft to meticulously weigh every word. But the line is thin).
Of course, the output may be more brilliant than what the author would have written. But still, it is not what he had in mind. Noise has been inserted between the author's mental state and the reader. There are often a thousand adjacent arguments leading to the same conclusion, and only one of them is the author’s. More generally, word order, sentence structure, lexical nuance, stylistic emphasis are part of the reasoning: change those and you change the argument. (As an author, this is why I sometimes sacrifice a dose of fluency to stick more closely to what I wish to say – I have had countless arguments with editors over my use of parenthesis for example). This is why it would sometimes be better to read someone’s prompts rather than their model’s output. It would provide more precise information on what they actually think.
An AI copy-paste doesn't leave you feeling respected
When someone writes, he is by construction obliged to choose, at every point of every sentence, the word which he finds optimal for you to read. Of course, he does not always succeed in choosing the best words, but at least he has tried. On the other hand, when someone gives you an AI generated text, you know that he saw a bunch of words, and thought « good enough, I’ll make him read that ». An implicit contract binds an author to his reader; that contract is broken when the author spends less effort producing the text than the reader will spend reading it. Reading a good human text is spending a moment with someone who took care to receive you well. Nobody feels well received by an AI copy-paste.
We already have chatbots
Fourth, an author owes the reader what the reader could not have obtained alone, not something he could get in a few seconds from Claude. If a reader knows that he could regenerate an equivalent version of the text (perhaps more tailored to his taste and personality), the motivation to read legitimately evaporates. LLM texts can be interesting and informative, but we have chatbots for this.
It is fine to remember that we are social animals who crave connections with other members of our species, that a text is a medium of communication and not only a string of words, and that you cannot strip the human out of it without making it less appealing to a potential reader. The logic probably transfers beyond writing. The crowd that insists that “the only thing that counts is the quality of the output” should think about the future they are announcing: once machines surpass us at everything, nothing a human makes will be worth another human's time.
