A long time ago, you might have heard from Jared Spool that everyone is a designer. But over 5 years before that bombshell, Ian Bogost put forward an even more maximalist thesis: everyone is everything. We are all hyperemployed; in addition to our official job, technology has heaped on countless minor tasks that we are obligated to perform to Get Things Done (a similar thesis was explored in book-length format 30 years prior to that by Ruth Schwartz Cowan in More Work For Mother, which has never been more relevant).

Silvio Lorusso’s piece on the history of design’s collision with Web 2.0 made that explicit connection, and then some:

Enabling this ruinous democratization of design was the notion of the template. It’s not that anyone could now do design; merely that everyone was now expected to fiddle around the edges of an artifact that was already designed by someone else.

I love Lorusso’s essay, and I think everyone in the business of user-facing software development should read it. I especially like the equivalence made between design and sending emails. Because, as you might have heard, there is now a machine that can generate “a design” as easily as an email. A machine that ingested millions of “designs” and billions of emails and turned them all into one big template that it can fill out for you. A machine that had no equivalent in 2017, when that essay was written, and is now everywhere.

But people have been noticing that while an LLM can send an email, it is not at all capable of doing it well. And the frustration with workslop has only intensified over the past year. We have long been begging our colleagues to give us their own thoughts in their own words instead of sharing the output of a chatbot that we could prompt just as easily as they did (if that output was what we wanted).

I’ve written about workslop on the Picnic before. Heck, I’ve also written about templates. Twice, even — once specifically about the templatization of design, and once about the templatization of thinking in general.

The insidious thing about templates is that they are frictionless to use. 99% of the work has already been done for you; you do little more than sign your name and then send it off into the world. But each email authored this way is a burden on the person (or people!) who must read it using their human brain and their human time. It is a gesture of profound disrespect. One may go as far as calling it cognitive littering.

An effective message is not formed by unidirectional template-slinging. It is a social process of telling your story and then hearing it told back to you to learn what was understood.

Unfortunately, we live in a world where everyone is an email writer, and the baseline quality of email writing is predictably atrocious. Who will notice if you phone it in and ask Gemini to generate the email that your boss is making you write? (I will, and your colleagues will, and we will make fun of you in the group chat).

But it gets worse.

You might have seen the term “world model” bouncing around. It’s a weighty-sounding phrase for a simple concept: executives don’t know what is going on, so they want to upload every scrap of data from the company into an LLM, and then ask it. This job used to be filled by a highly paid manager who pasted screenshots of dashboards into a PPT and took credit for the numbers, but now it seems AI is taking their jobs away.

This use case produces the greatest template of all: a template not for a generic report that reads the same as every other report, and not even a generic website that looks like every other website, but a simulacrum of a company that runs like every other company (which kills your margins because it makes your product a commodity).

Anyway, the world model won’t do anything for the same reason that a parade of Gannt charts didn’t do anything: boiling everything down into a big vat of stuff destroyed the context that made any of that stuff meaningful. It’s actually even worse now, because at least those charts were based on something real happening. In a world of LLM-generated documents, scraping everything indiscriminately will produce a thin broth that contains no real understanding.

Even if you write down (or, let’s face it, have an AI transcribe) a meeting or an interview into a document¹, knowledge does not automatically come along with that text. Knowledge is something that is produced in your brain as you transcribe the text, because transcription is an act of metacognition. It is synthesis. It is thinking.

Doing that thinking made you valuable, because you had insights no one else did. You were the only one with the entire corpus in your flesh-and-blood “context window” that we used to call institutional knowledge, and stitched it together with an agentic workflow we used to call “expertise” (the very expertise, by the way, that you need in order to check the output… you are checking the output, aren’t you?).

Maybe the AI aesthetic was cool, for a brief moment in time. The images felt novel. The writing may have seemed crisp and professional. But it’s not the early days of AI anymore; everyone can recognize it and everyone is tired of it.

There is even nostalgia for a moment that was not so long ago, which everyone complained about at the time:

Businesses are now going out of their way to prove they have nothing to do with AI. Some are even investing into painstaking analog methods; it would be trivial to generate a Stable Diffusion equivalent, but that is sort of the point. Even LinkedIn, which has a “generate slop” button right on its home page now has a second button to report when you see someone else posting slop.

Everyone has seen enough of this stuff to tell the difference between the genuine article and a synthetic imitation. And if you can’t, it doesn’t actually mean that you are somehow just that good at prompting.

— Pavel at the Product Picnic