Lost in Compression

I have stopped reading my colleagues’ documents.

Not all of them, not yet, but enough that reading requires a shameful triage: a glance at the author, the topic, the structure, sometimes the first paragraph, and then a quiet calculation about whether anyone really wrote this thing and whether anyone else is going to read it carefully enough that I need bother. Most days, the answer is no.

What looks like laziness is rational.

Alberto Brandolini formulated what he called the bullshit asymmetry principle: the amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it. The time to produce a plausibly competent document has fallen to seconds. The time to review that document carefully and to verify its claims has not. The asymmetry is now structural, because the bottleneck is no longer who can produce but who can read carefully enough to spot the garbage.

What I am observing, and what I suspect many have observed too, is the predictable equilibrium that follows. People stop reading. Instead they skim the executive summary and trust that someone else has done the verification. The document circulates because circulating it is the job, yet nobody upstream is fully sure what is in it, and nobody downstream is going to find out.

This is how the workplace commons is slowly being poisoned, and the people doing it are mostly not malicious. They are tired and the tools they have been handed have made it ridiculously easy to produce things they would not have bothered to write before.

The rule: expand inputs, do not compress outputs

The rule to sensible AI usage within an organization is as follows: use AI to expand your inputs. Never use it to compress your outputs.

When you study a topic or market, work through counterarguments to your own ideas, or discover a research paper you would not have known about otherwise, AI widens the input space your judgement operates on. I have written things that won awards because an LLM helped me find a connection I would not have made on my own; the connection was real, the prose was mine, and the judgement about what mattered was also mine. That is the version of the tool that helps, and the one I have no intention of giving up.

When pointed outwards (e.g. a Slack message you owe a colleague, the RFC from which a decision has to be made, or the code someone else will have to maintain), it does the opposite. It compresses the specifics of you, of them, and of this moment into the statistical middle of how people in roles like yours typically address situations like this particular one.

Industrialized regression-to-the-mean

Large language models (LLMs) are regression-to-the-mean machines. Trained to produce the most probable next token given a context, they fluently output the statistical centre of mass of their training corpora. Consequently, LLMs are homogenizing human language and reasoning at scale. An LLM is like a blurry JPEG of the web. The compression is good enough that the artefact still looks sharp, but everything specific has been sanded off to be so very average.

In literature, AIs can help make each short story better yet reduce the diversity of all short stories. That pattern has been replicated in business, particularly brainstorming sessions: AIs can boost individual contributions, but people tend to converge on the same ideas when using the same model without any coordination. Or in the words of one of the researchers:

If you rely on ChatGPT as your only creative advisor, you’ll soon run out of ideas[.]Prof. Christian Terwiesch

As such, your colleagues’ documents are individually fluent, collectively interchangeable, and devoid of the particular intelligence and style of the person whose name is at the top. Withdrawal is therefore rational: the reader who no longer bothers is responding to the fact that production and quality have both circled down the drain.

When in Rome…

When you use the outputs of an LLM as inputs for training, the model “collapses”, which means its data distribution narrows. Rare events vanish first, then less-rare ones, and eventually the model is generating noise, though confidently. This is what happens when you photocopy a photocopy of a notarized photocopy: it looks legit as long as no one actually tries to read it.

Business communication is nowadays run through a sequence of lossy codecs that produce verbose mediocrity. Person A has a thought and lets an AI compress it into a Slack message. Since the text produced is much larger than the thought itself, they believe the AI improved upon their idea, but it is just dressed up in more fluff. Person B summarizes it and reads the gist of what A meant to say. Subsequently, B compresses their reply, and A reads a summary of B’s response. It is a back-and-forth between two humans with four compression steps, in which the likelihood anybody understood anything decreases geometrically. And one day soon, A and B are no longer part of the conversation: the AIs are conversing on their behalf, and no human thought is required at any step.

LLMs in communication automate the appearance of efficiency at the cost of more noise that drowns out the actual signal and erodes rapport. People who suspect their conversation partners are using AI-generated replies rated their interlocutors as less cooperative and more dominant.

Vibe coding, in which the person whose username is in the git commit message but who cannot actually reason about the code, follows the same pattern. For throwaway scripts and prototypes, the cost of compression is practically zero and the practice is fine. But when team members need to debug, extend, and trust the code the cost is real. It is merely deferred to the next on-call engineer or the customer who hits the edge case the AI averaged out.

When pointed inwards at what you have to think about, an LLM can help. But when pointed outwards at what you owe to another human, it removes the specifics that made the contribution yours. The first can make you smarter, whereas the second will make you less worth talking to and the things you ship less worth maintaining.

The smell of LLMs in the morning

Because the compression is uniform, so is its signature. LLM-generated text has a peculiar tell that, once detected, rapidly changes people’s perceptions of the person who published it. You can observe it on Hacker News: a comment that smells as if it has grazed past an LLM gets dismissed on those very grounds, with no engagement on the content, sometimes deservedly and sometimes not. This is perhaps a transitional cultural moment, in which being suspected of having outsourced the writing is sufficient to discredit the author or an idea. I am guilty of it too: I have rolled my mind’s eyes at executives who blurt out random ideas backed by words such as “substrate” or “scaffolding” or “load bearing” or “tapestry”, when in the years before they never once had such words in their vocabulary. Such is, unfortunately, the “collapse” of reputation.

What is at stake

Some corporate communication does not deserve the time to be written by hand. Status updates that can be summarized from Slack, Jira, and the like, for instance, ought to be generated automatically. An output, in the sense the rule means, contains something that did not exist before you came up with it. A status update summarizes data that is already there; it never needed a mind, and it still does not.

There is, however, a category of communication, especially with the people we are responsible to, about the things that actually matter, where compression is exactly the wrong move, and where the cost of treating it like throughput optimization erodes the trust needed to sustain relationships.

The documents I am no longer reading are mostly ones that did not need to be written at all. AI is not solely responsible for the slop; it is simply making it clear how much of what we were producing before was already slop. AI may have produced more good documents too, but a document you cannot find is the same as one that was never created.

The danger is that the noise LLMs generate is making the signal indistinguishable, and rational readers respond the only way they can, by withdrawing attention from all of it. The good documents go unread alongside the bad ones. The latter have always outnumbered the former, but nowadays they bury the exceptions.

My triage is correct in aggregate and wrong in the particular. Most days I am right to ignore what slides across my desk; when I am wrong, I will not know it. Which makes me a lossy codec too.