The question is not whether AI reduces work, but where the intellectual work moves.
There is plenty of AI slop on the internet, and much of it deserves the criticism it gets. But let’s not pretend AI invented slop. Humans were producing shallow, derivative, formulaic, and poorly researched content long before most of us knew what an LLM was.
What’s becoming more tiresome than the slop itself, though, is the growing commentary treating any use of AI-assisted writing as an intellectual failure. A certain kind of AI critic has become the new grammar Nazi, policing not only bad writing but also the means by which writing is produced.
Much of this criticism seems to rest on the shortsighted assumption that doing more of the writing manually necessarily makes the resulting work more intellectually valuable. The same assumption can shape how we judge reading, as though working through a text without assistance necessarily produces deeper understanding. Both judgments treat the difficulty involved in these processes as evidence of intellectual value.
But that treats all the layers of difficulty involved in reading and writing as if they were inherently valuable. Some are, of course, but others may simply be consequences of the interfaces through which we have learned to think and communicate.
Yes, reading and writing are interfaces. That may seem like an odd claim because most people think of interfaces as screens, controls, or other points of interaction with technology. But an interface is simply a means through which a person interacts with a system, and that system need not be computational.
But even if we associate interfaces with technology, written language is itself a technology, a system of symbols, conventions, and relationships developed to create, preserve, and communicate meaning. The idea is not new. Walter Ong famously described writing as a technology that restructures thought. Reading and writing are the interfaces through which we interact with that system.
If reading and writing are interfaces, then we can evaluate their friction much as we would the friction in any other interface. Some friction serves a purpose, while some simply makes the interaction more difficult.
Consider a confirmation that appears before a file is permanently deleted. It adds friction, but it serves a purpose by giving the person an opportunity to reconsider an action with consequences. Repeatedly entering the same information or navigating unnecessary screens also adds friction, but without comparable value. The question is what purpose that friction serves.
Reading and writing are no different. Interpreting a difficult argument requires effort, but so does spelling a word correctly, constructing a sentence, remembering a citation format, organizing information, and reconciling conflicting evidence. Some of that effort contributes to the capacity being developed or exercised, while some simply accompanies it.
For a child learning to read and write, struggling to spell a word may be part of the process of learning written language. For an experienced writer developing an argument, the same struggle may contribute very little to what they are trying to do.
Friction that contributes to developing or exercising a capacity can be understood as constitutive, while friction that simply accompanies it can be understood as incidental. What falls into each category depends on the person, the task, and the capacity involved.
Research on desirable difficulties, for example, has long shown that certain forms of effort can improve learning and retention. The important point is that difficulty has value because of what it contributes to learning, not simply because it is difficult.
Technologies have been reducing the friction of written language for a long time. Dictionaries reduce the need to recall spelling and definitions. Word processors let us move text around without rewriting a page, while spellcheck catches errors automatically. Search engines make information easier to find, and citation managers handle some of the tedious work of formatting sources.
But to be fair, generative AI is different from those technologies. It can get much closer to the thinking itself. It can interpret, organize, revise, synthesize, translate, and generate. While a citation manager can format a source for me, AI can tell me what it thinks the source means. Whether that interpretation is accurate is a different problem.
If AI corrects my spelling, I probably haven’t given up much of the intellectual work involved in developing an argument. If I give it two conflicting arguments and ask it to reconcile them, I may have given up much of the intellectual work the task requires.
It would be foolish to think that by using AI, the intellectual work simply disappears. AI can shift some of that work from production to judgment. I might ask for several interpretations and then compare them, question their assumptions, check their claims against the sources, reject some, and work out what I actually think. AI also requires decisions about what to delegate, what to trust, what to verify, and whether the generated language actually says what I mean.
Whether any of this is useful depends on what I am trying to do. In education, it also depends on what someone is trying to learn. Something one person has already mastered may be exactly what another person still needs to practice. AI could remove incidental friction for the first person while removing part of the learning for the second.
Reading and writing have been so central to education and intellectual life that we often treat competence with them as evidence of other capacities. Polished prose can indicate clear reasoning, but polished prose is not clear reasoning. A good summary can demonstrate understanding, but the summary is not the understanding. Correctly formatted sources are part of good research, but formatting them is not the same as evaluating them.
Until recently, these things were difficult to separate. If someone produced a clear synthesis of several difficult arguments, there was good reason to think that person understood them. Producing the synthesis usually required doing much of the thinking that the finished text appeared to demonstrate.
AI weakens that connection. Someone can now produce a convincing synthesis without necessarily understanding the material. That is a real problem, especially in education, because the finished text becomes less reliable as evidence of what the person who produced it can actually do. This problem is already leading researchers to rethink how student learning should be assessed when generative AI is involved, including whether the process of interacting with AI can provide evidence of reasoning that the finished text alone no longer provides.
But someone can also understand the material, develop a strong idea, and use AI to help express it. Knowing that AI was involved does not tell us which of those things happened.
I am therefore less interested in whether someone used AI than in what they used it to do. Did it help them express their thinking, or did it do the thinking they were supposed to be doing?
Bad AI-generated prose deserves criticism because it is bad prose. Claims that lack evidence deserve criticism. Fabricated citations deserve criticism. Using AI to avoid engaging with the material deserves criticism too.
But AI did not invent any of these problems. People produced bad prose, repeated unsupported claims, misrepresented sources, and wrote things they barely understood long before generative AI made doing so easier. AI may increase the scale of these problems and make some harder to detect, but its presence alone tells us very little about the intellectual value of the work.
There is a danger in allowing the worst uses of AI to define our response to the technology itself. We have a technology that, under the right conditions, can multiply our productivity and intellectual capabilities. Experimental research has already found substantial productivity gains from generative AI in professional writing tasks, including faster completion and higher-quality output.
Yet much of the conversation remains fixated on AI slop and preserving the ways we previously read and wrote. Criticism is necessary, but when criticism becomes a desire to preserve existing practices simply because they are familiar, it begins to look less like criticism and more like Luddism.
Treating AI assistance itself as evidence of intellectual failure risks defending the mechanics of reading and writing rather than asking what we want from those practices in the first place.
We do not need to preserve difficulty simply because it has traditionally accompanied reading and writing. We need to decide which capacities we actually care about, which forms of effort help develop them, and what exactly we are giving up when AI makes something easier.
Don’t miss out! Join my email list and receive the latest content.
Reading and writing are interfaces. And AI can reduce their friction. was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.
