When Fluency Is Free
Now that AI has become an almost mandatory part of most people’s writing, most text-based content that lives on the internet (e.g., blog posts) will be written by, or at the very least heavily helped by, an LLM.
I am concerned about the effects this shift will have on how people pick which written ideas to engage with, and the collapse of discoverability.
One could argue that LLM-written text will standardize writing, as with the help of LLMs people can quickly turn their half-baked thoughts into fluent ideas with a standard “LLM prose”. But for readers, this efficiency comes at a cost: content that is worthy of your time becomes indistinguishable from content that is not.
This is underscored by the phenomenon most of us have already experienced, where if a text is clearly AI generated, we will either skim it or skip it entirely, as it is expensive (w.r.t. time) to evaluate.
For most content on the internet, a prerequisite is that it is fluent. Historically, there was a sort of filter in what got posted to the internet—where fluency implied intentionality, which would act as a signal that something is worth reading. But with LLMs, fluency is free and so the signals are lost.
So what do people do? They shift to reputation systems—to known sources, authors, celebrities, etc. and subscribe to a person rather than a topic. This risks reversing the democratization and discoverability of written text on the internet. Sure, anyone can write, but the people who actually get traction are likely those who already had it to begin with.
Are you still speaking from the perspective of the people who use LLMs to quickly turn around their half-baked ideas, or from the perspective of readers who will be deluged more and more with LLM-baked content?
What do you mean by the internet—even before someone actually reads a post or article, a whole host of factors determine if it even gets a chance before the first paragraph: what personal or contextual relevance? Hyperbole (“10 best pictures of Fawns, you won’t believe no. 4”)? Ragebait?
On the flip side. Kant is notoriously difficult to interpret, Wes Cecil says it is easier to read Kant in English translations than his native German. Will such difficulty of prose become the new sign of intentionality?
Good catch, the perspective shifted in the paragraph. The question was meant to be spoken from the perspective of deluged readers. Edited for clarity.
Agreed, there is a host of factors that determine if something gets a chance at being read independent of the meaningfulness of the text itself. What I am trying to get at is that among content that does get a chance, fluency used to be a weak signal that the person who wrote it probably at least spent time trying to understand or write something meaningful, and so it may be worth a read. When everything looks coherent upon first glance due to LLMs polishing surface-level coherence, that signal gets covered in noise. So people spend less time engaging with unknown sources online, as it’s simply too expensive to evaluate all the content, and they resort to trust networks rather than independent discovery.
One place we may already be seeing this is cold email. Historically, receiving a personalized email from a stranger was weak evidence that another human had spent nontrivial effort thinking specifically about you. Since personalization was costly, it functioned as a signal. LLMs partially collapse this. Once plausible personalization can be generated at near-zero marginal cost, “I saw your recent work on X” no longer strongly implies genuine attention. Average cold email reply rates have dropped from around 8.5% in 2019 to 3.43% in 2026 (per the Instantly 2026 benchmark report), with industry analysts directly attributing part of the decline to low-effort AI-generated outreach. As the cost of generating plausible communication approaches zero, recipients adapt by allocating more attention through existing trust networks (friends, colleagues, reputation systems, established communities), and discoverability increasingly flows through connection graphs rather than unsolicited communication.
Difficulty is probably part of it, though I’m not sure exactly what the new signals will look like. If I had to guess, it would be anything structurally farther from what an LLM would output: relevant comic strips between paragraphs, forgoing em dashes, idiosyncratic structure. For cold emails, it may mean bypassing email entirely and showing up at a professor’s office hours without prior introduction.
I do suspect that genuinely difficult writing (actual hardness, not the appearance of it) will increasingly signal intentionality.