The Dead Office Theory

The best-written email in your inbox today might have come from someone who couldn’t explain it to you on the phone. In writing, though, they’re doing great (sweetie). They might be several emails deep into a conversation with a colleague, making detailed arguments while playing passenger prince or princess in a 2017 Kia Sorento. And often, the person on the other end of their emails isn’t doing much writing of their own. There’s a decent chance neither of them is doing much of the reading, either.

One of them drops a meeting transcript and a Slack thread into Claude or ChatGPT and asks for a “clear and professional” email. He skims the draft, deletes an em dash or two because that’s how people can tell, and sends it along.

On the other end, the recipient opens a six-paragraph email full of neatly bolded intro sentences she has no particular desire to read. She drops it into her own LLM: “What even is he asking me for?” “What am I supposed to say here?” The model gives her the short version and a reply that suggests she has given the matter considerable thought. Off it goes.

The response comes back. He asks his LLM what she means and whether his thing is approved. He could read it himself, but he’s already got the chat open.

Both people are technically involved. They supplied the context and approved the drafts. Their names are on the emails, and if something goes wrong, nobody’s getting on a call with Claude to sort it out. But the machines are doing much of the writing, reading, and interpretation. The people are mostly moving things between windows.

Call it the Dead Office Theory. An office keeps producing work that looks as though someone thought it through, even when the people putting their names on it haven’t. The name borrows from the Dead Internet Theory, a maybe-not-so-fringe idea whose central observation has aged unusually well.

The what?

Take the more conspiratorial parts out of the Dead Internet Theory and you’re left with an uncomfortable idea: the internet can look full of people who aren’t actually there. Machines produce content and other machines supply the attention. Put too simply: way more of the internet than you think is just bots talking to each other.

In 2018, Max Read wrote in New York Magazine about YouTube employees’ fear of “the Inversion.” Bots could become so common that the systems meant to catch them would start treating their behavior as normal and flagging the humans as fake. You could be an actual person and still be doing it wrong. His larger concern was trust: there was money to be made manufacturing the appearance of human attention, whether anyone was actually paying attention or not.

Imperva’s 2025 Bad Bot Report put automated activity at 51 percent of the web traffic it measured in 2024. Some of those bots are doing perfectly useful things, like crawling websites. Useful or not, that’s still a lot of activity across the internet (which is, of course, a series of tubes) without anybody at a keyboard.

Generative AI makes it easier to hand over the conversation itself. One machine writes something and another reads it, figures out what it means, and drafts a response. A person can get involved at any point, but the exchange doesn’t necessarily need them to.

An email still arrives with somebody’s name on it. You reasonably assume that person knows what it says.

In the Dead Office, the bots are still talking to each other. We just put ourselves in the middle.

The office was ready

Office communication was formulaic to begin with. Status updates, project briefs, meeting recaps, follow-up emails, a deck that summarizes another deck. Things of that nature. There’s a whole genre of Instagram and TikTok creators who make their living reenacting the office (the place where you work, not the 2005 NBC smash hit The Office: An American Workplace). We get the jokes because we sit through the same meetings and hear the same phrases every day.

Having a standard format helped. If everyone presented things the same way, all that was left was to figure out whether the content was any good. The trouble started when people assumed the thinking had happened because the deck looked right. We’d seen plenty of good work arrive in that format and, after a while, we stopped looking quite so closely. Someone who knew what people expected to see could put together a deck that went over perfectly well (they got a nod in the meeting), even if they didn’t know much about the problem.

An LLM has read more meeting recaps than any content creator ever could. Of course it was going to learn how to do the voice.

The work inside the writing

Having to write about a topic used to be the fastest way to find out whether you actually understood it. Every English teacher you ever had said to lead with the point. So you’d sit down to write the point and discover there wasn’t one yet. What does that number mean? Why did the timeline move? What had the client really asked for? No idea. So you’d go open the spreadsheet, Slack somebody who wouldn’t judge you too hard for not knowing, or feverishly research your way out of having to ask the dumb question. Then you’d reread a thread and notice that two people had been agreeing with each other about completely different things for a week.

You’d spend an hour on a two-paragraph email and forty minutes of it was away from the Google Doc, deciding what was actually true. By the time you’d figured out which facts mattered and what you thought should happen next, the email mostly wrote itself. Because by then you knew what you were talking about.

That process was annoying. It was also a decent chunk of the job.

Now, you can paste six Slack messages, a meeting transcript, and a handful of half-formed notes into an LLM and ask for a concise client-facing explanation. A few seconds later, you have something clear and confident enough to send, but your understanding remains exactly where it was before you pasted anything. 

To be fair, not every email works like that. Plenty of times you already know the answer, and all that’s left is turning it into something a client can read. Having a machine do that part is usually kinda sorta fine. Nobody’s thinking got skipped.

Other times, you’re pretty sure you know the answer, and it’s the act of writing it out that catches you. You start explaining why the timeline moved and realize the reason you’ve been giving people doesn’t actually “hold water” (Gambini, Vincent, 1992).

The trouble is that all of these look the same from the outside. The same blank page, the same “explain to Brenda that’s not actually how computers work.” The LLM treats every one of them like you already understand. Sometimes that’s right. Other times, you’d have gotten halfway through writing it yourself and had the moment of “Wait. Shit. Hang on. That actually doesn’t make sense.”

How assumptions become facts

A person asks an LLM to turn some incomplete notes into an update. They give it a quick read (maybe) before sending it to someone who asks another LLM for the short version. Neither goes back to the notes. Whatever was missing at the start is now missing from something that sounds finished.

Reading the source material takes time. So does double-checking numbers. Math is hard. You get credit for shipping the deck. You rarely get credit for spending an hour discovering that the deck shouldn’t be shipped.

Imagine a go/no-go meeting for a website launch. An engineer says a Wednesday launch is possible if a vendor fixes a login problem by Monday. The AI-generated recap lists Wednesday as the target but leaves out the part about the vendor.

A project manager skims the recap and asks Claude to draft the client update. The client will hear that the launch is on track for Wednesday. By the next meeting, everyone is working toward that date, and nobody is talking about the login problem.

Ask the project manager and they’ll point out that nobody corrected the recap. The engineer, meanwhile, is fairly sure everyone heard the Monday condition, since everyone was sitting right there. None of this is much use to the client, who has an email saying Wednesday.

Nobody needed to invent a date. Wednesday started as one engineer’s “Yeah, maybe. If…” The “if” disappeared from the recap, and every person who used it afterward assumed they were working from something somebody had checked. Call it institutional hallucination.

And it didn’t take much. The recap was one paragraph. Someone who’d been in the meeting could have caught the missing condition. But they read it the way we tend to read recaps: quickly, assuming whoever made it got it right. By the time it reached the client, “someone reviewed this” meant little more than “this passed through someone’s inbox.”

Inside the Dead Office

A Dead Office is a busy place. Tickets close, meetings get summarized. Everyone has something to put in the status report.

After a while, most of the day is spent looking at things that describe work. The summary of the meeting. The draft of the update. The recap of the recap. And once the reams of generated content start to come in, feeding them to a model might be the only way to keep up. You’re still on the hook for the outcome, and your name is still on the recommendation, even if you never got far enough into the problem to have an opinion about it.

That didn’t happen all at once. Two years ago you still mostly wrote your own updates. Then last year you got Cowork and started having Claude draft them, and the job became reading the draft carefully and fixing what was wrong. Then the drafts got good enough that careful reading felt unnecessary, so you skimmed. Now you open the draft, see nothing alarming, and hit send.

What actually is the job at that point? A model writes a thing, you glance at it, another model on the other end reads it. You’ve relegated yourself to being the part in the middle that clicks the button.

You can usually tell when it’s happened to the person on the other end. The detailed, confident emails about things you know they had a passing understanding of a week ago. They’ve got a whole argument worked out, but you’re reading it thinking that if you got them on a call and asked a couple of basic questions, they’d have a hard time backing it up. They’ve sent something that makes sense on the page before they’ve really worked through it themselves.

From the outside, though, the exchange looks perfectly normal. And how often does someone actually press a colleague into talking through something they’ve already explained so confidently in writing? The email was sent and acknowledged. But somewhere in the thread, the person stepped out and the process kept going without them.

Some of it should disappear

A lot of office work is bullshit, and automating it is good. Nobody needs to rediscover the spiritual value of writing the weekly status update or reformatting meeting notes. If three facts are buried in 900 words, have the machine give you the three facts.

The awkward part is that the useful work was often mixed in with the tedious stuff. Putting together an update meant finding out whether the project was on track. You had to bother the developer about the fix before you had anything useful to tell the client. You probably didn’t enjoy either part, but by the time you finished, you knew something about the project that you hadn’t known before.

Now you can get the update without having that conversation. Sometimes you’ve saved yourself twenty minutes of rearranging information you already understood. Other times you’ve skipped the bit where you would’ve found out what was going on. Both feel like getting the update done faster, which makes it pretty easy to keep doing it. Whether it’s good or not is another question entirely.

Keeping a hand in the work

“Okay, Matt, that’s a lot of words. What’s your point?” First of all, thank you for reading this far.

In less than two years, we went from talking to each other to, at least in part, being middlemen for Claude, which is certainly an easier way to spend the day. Working out what we actually think can be a pain, and here’s a thing that will hand us an answer that sounds more certain than we feel. It’s hard to blame us for being happy to assume it knows better.

But if you, Mr. Strawman, are going to insist that this screed have a point, it’s this: if your name is on the recommendation, you should be able to explain it.

If you already know the answer and just need it typed up, let the machine type it up. (Maybe. Turning a thought into a fully formed message is a skill we should all have, but whomst among us.) If you’re still figuring out what you think, spend some time doing the research before you ask Claude to make it sound finished.

Try writing your conclusion in a sentence or two. Have you answered what the person reading it was really asking for? There’s probably an obvious follow-up you could answer before they have to ask.

It’s worth thinking about reasons you might be wrong, too. Those of us with anxiety disorders should find that part pretty familiar; playing out all the ways things could go wrong is already muscle memory.

If you’re struggling to explain your conclusion, ask the model to help you work through it. But don’t stop just because it’s produced something that sounds confident. Otherwise, there’s not much reassurance in calling the person who clicks send the “human in the loop.”

And please have some mercy on the people who have to read what you generate. Being able to produce ten decks doesn’t mean there’s a good reason to have produced ten decks. Your teammates still have the same amount of time and attention they had before we discovered we could generate a report on everything. Bury them in enough of it and the only way they’re getting through the day is by feeding content into an LLM. Now their bot is reading your bot’s work. I guess that’s the way the whole durned human comedy keeps perpetuatin’ itself down through the generations. Westward the wagons, across the sands of time until we…ah, look at me. I’m ramblin’ again.

Okay let’s wrap this up: a lot of what we’re calling time saved is work we skipped. Those annoying hours were where we found out what was going on and decided what to do about it. Claude can give us something that looks ready to send before we’ve figured out either.

We still have to spend time with the problem. If your name’s on the recommendation, you need to understand it well enough to disagree with it, or at least notice it answers a question nobody asked. Otherwise all we’re doing is throwing a bot’s answer over the fence, and sooner or later somebody’s going to figure out it’s cheaper to leave us out of the exchange. The way to stay in the thread is to be the person who knows enough to ask, “Say man, what question are we even answering here?”

So use it for the dumb stuff, like reformatting the notes, pulling three facts out of 900 words, or turning an answer you already know into something another person can read. This is the time worth saving. Don’t let AI just Leeroy Jenkins its way through the office because it can. The hour it takes to find out what’s going on was never the hour that needed saving.

1. Tread Carefully