Using AI to write often feels like cheating. This is most obvious in education: a history teacher doesn’t assign an essay exploring the root causes of the French Revolution because they really want an eighth-graders’ opinions on Robespierre. That essay is a forcing function to revisit material, do research, and reap the benefits of putting ideas into words. AI can write a pretty good (better than an eighth-grader) essay on the topic instantly, without the hassle of research, revisiting materials, or editing.
But the benefit isn’t in the final essay, but in the process it took to write it. This is obvious in education, but also applies at work.
At a past company, we needed to revamp a part of our growth strategy (the team I work on). We had three AI-generated strategy documents from three people that were shared once and forgotten. None went through comments and revisions to create an official version that could act as shared context. The desire to “adopt AI” and have some version of a strategy negated any benefit a strategy document may have.
At work and elsewhere, some documents or other artifacts produce their benefit because they exist while others are valuable because they’re evidence of the process that created them. The former is a perfect target for LLMs while the latter causes more harm than hurt when done with AI.
I work in growth/marketing, so my examples are about writing, but this maps to every profession in upheaval due to LLMs, like software engineering, design, and much more.
The easiest way to explain when the effort is the point is in relationships.
Rory Sutherland believes one reason women love receiving flowers from men is because most men would never buy flowers for themselves. Flowers are proof he was thinking of making her happy. As Miley Cyrus would have us know, she could’ve bought herself those flowers. But the fact that he made that effort makes the flowers meaningful.
Flowers are one stereotypical example of a broader point. It may as well be a woman buying her husband flowers, or one partner planning a surprise dinner at a restaurant they dislike but their partner loves, or a friend remembering the wine they shared and buying that same bottle when they reunite.
The flowers, restaurant, or wine are symbols for the fact that one person spent time to make the other happy. They’re a byproduct, meaningful because of what they symbolize. This is when we imagine buying flowers you forgot your partner was allergic to: those flowers are suddenly evidence how little you care.
These are intuitive examples because interpersonal relationships don’t exist to complete deliverables, run workflows, and achieve objective results. Work is more transactional.
AI hands us the results of once-laborious processes and makes many processes much faster this is great because nobody misses taking meeting notes or scrolling through a recipe blog tracing the history of tomato sauce back to the Roman Empire before reading a basic recipe.
But using LLMs becomes a problem when we use them to generate outputs whose value lies in the process they prove.
What makes a strategy doc, project plan, or design spec valuable? Clear responsibilities, simple instructions, and well-defined milestones are nice characteristics, but the benefit lies in the process of a shared cognitive context.
Before LLMs, arriving at an official strategy document went through important processes:
The author wrote it from scratch, with the typical benefits of the writing process: understanding (and plugging) the gaps in one’s knowledge and sharpening one’s understanding by explaining it to others.
Review cycles surfaced where perspectives collided and required reconciling the differences into a shared perspective until there was a final version.
This process makes foundational documents useful because people understand not just what to do, but the underlying thinking. Their approach to tasks will reflect that and they can explain it to others. Meetings are smoother because everyone shares the same context.
Even if you could’ve generated the same doc to the letter with an LLM, the outcome would’ve been worse because nobody did the process it should represent. AI accelerated creating the document, but then creates friction in everyday work. The lack of a shared context causes more questions to come up and reviews necessitate starting from scratch. Outcomes are worse if people make faulty assumptions or take longer because the back and forth starts when tasks are assigned.
I’ve had someone tell me “I don’t know, it’s just a Claude draft” when asking a question about a document they sent me. This also diminishes trust and goodwill, damaging collaboration itself because it shifts the burden of work.
I previously described how getting someone flowers they’re allergic to expresses a lack of effort. I feel that way when I have to review AI slop, which frequently happens with freelancers. The better metaphor here is receiving plastic flowers presented with “I brought you the flowers”, which is technically true, but false in any meaningful way.
Editing AI slop is infuriating because it’s like trying to use water to clean a sandcastle. It melts away wherever you try to improve it. The knowledge a piece of writing is supposed to embody doesn’t exist, so you can’t sharpen it.
As a reviewer, I effectively write the first draft even though there were already words on the page, often thousands of them. LLM drafts are usually longer than human writing, so someone saving themselves two hours of writing by producing a document in 30 seconds makes me spend those two hours.
When this happens with internal documents, people usually have AI summarize or review it, which further corrodes the idea of a shared cognitive context. And if someones does a deep read with questions and comments, one person feels like they did the work other people cheated themselves out of.
A former colleague told me that if they see a Slack message longer than two sentences from their manager, they know it’ll be AI slop based on spurious information, so they try to find a way to let that conversation fizzle.
This is asymmetric workfare: the cost of executing a workflow is much higher than the cost of producing the impulse for it.
This affects the work process itself, which might be the most important thing to protect in any company: its culture. If people don’t trust one another and minimize their effort because their colleagues are doing the same, productivity falls apart, even if more people are using more AI to do more things.
None of us work in a vacuum, but AI slop would harm us even if we did.
My grandparents’ generation was a lost generation for physical exercise. Sports was a hobby like knitting, reading, or the local history society. Doctors cautioned against exercise and runners were questioned by police about their suspicious activity.
The Physical Activity Transition theory explains this: Physical activity used to be a byproduct of everyday life. Most jobs required physical labor and transportation required physical movement, which gave humans a high baseline of physical activity that conferred invisible health benefits.
When cars became ubiquitous and work became sedentary, those invisible health benefits turned into visible health deficiencies. Heart attacks, and strokes surged. The decline of ambient physical movement (together with processed foods) caused a massive, ongoing health crisis.
Today, we know physical movement is essential. We run, lift weights, or do yoga to recoup the exercise (and its benefits) we lost in the Physical Activity Transition.
I believe we’re experiencing the “mental activity transition”. Our workday required mentally straining activity’s now optional, eroding hidden benefits yet to surface.
This is a pre-AI trend supercharged by LLMs. Our attention spans have been decimated by an attention economy hypnotizing us with infinite feeds. Cal Newport’s Deep Work lamented fragmented workdays and lack of sustained, demanding tasks long before ChatGPT.
But up until LLMs became ubiquitous, writing was a mentally taxing process. Early in my career, I wrote mediocre articles about random products. Even if these weren’t exactly doctoral theses, I can still name products and facts about some of them. That’s because writing is such an intense process that cements things in the brain.
Generating AI slop and publishing doesn’t convey those same benefits, which also undermines one’s subject matter expertise. Overly relying on AI erodes the very skills we built our careers on and means we no longer cultivate the specific knowledge that makes our skills rare and valuable. This buys us a brief pause in which AI does the work and we leave work early, but deskills us and erodes the very things we built our careers on.
This is similar to walking to work before cars were ubiquitous: frequent little bits of exercise (whether physical or mental) are beneficial.
I’m no expert on how this affects our brain’s health as a physical organs and how declining cognitive capacity affects the substance of society, but I have my suspicions.
Any AI enthusiast had a million opportunities to interrupt. Many of these issues existed before AI, and there’s nothing negative in using AI to generate a first draft to polish.
While not a single letter of this essay originated as an AI token, I’m not against AI. I use LLMs all day, have built AI agents, and constantly find better ways to use AI.
AI automates a lot of drivel, and that’s a good thing. I don’t miss taking meeting notes. If Claude can summarize a 117-message Slack thread for me, great. And when I review articles from freelancers, I rely on Claude to flag niche technical errors I’d never catch.
You can can also use AI within workflows where the process creates the main benefit:
I’ve heard from people who use AI to create a slop first draft to see what they don’t want and iterate towards a good result. Or who manually create an extremely detailed outline with angle, arguments, and supporting evidence and have AI fill in the gaps for a first draft.
Sharing an explicit “AI slop” document with one’s own annotations to start a brainstorm is a valid way for teams to get somewhere.
These preserve the benefits about the process: Sharpening your angle, challenging your assumptions, and defining what you want to write in the former case. Preserving the shared cognitive context creation in the latter.
Some practical things I do:
Flag to what degree something was made by AI (Not every SEO article needs to be written from scratch, though it always requires multiple human passes) and proactively tell people to not waste their time if to their eyes, it’s still slop.
Actually write manually, all the time. I make it a point to write from scratch to preserve my own skills and continue deepening my domain knowledge. It also produces better results.
Use AI as a thought partner and force it to make me do the work by forcing me to describe exactly what reader a piece is for, what the angle should be, what arguments and evidence I want to present, and how to position our product in the middle of it. (This is a modified version of grill me)
I don’t ban AI from even the important processes and continue to use it liberally, but replacing what makes a workflow valuable in the first place erases any value it may have created.
Maybe I’m wrong, and AI will just continue to get better and we will merely puppeteer a swarm of autonomous AI agents looping prompts back and forth. Maybe we’ll be like ceramicists during the Industrial Revolution who could’ve forgotten their muscle memory on the pottery wheel and their specialized knowledge about clay and kilns without loss because moving forward, they’d just be pushing buttons in the factory. I hope not, and the fact that the most AI-forward companies are all betting on hiring human writers is evidence that an AI slop counter-swing is happening. If I’m wrong, it might be time to start a goat cheese farm.