Bloated: how chat made you fat

By Iain, 29 May 2026

The pitch for writing with a language model is that it saves you time: you describe the memo, the model produces it and 90 seconds later you have four pages (okay, maybe forty) instead of a blank document. Someone still has to read those pages though. The model did not remove that work. It just moved it downstream to your colleagues or suppliers, and on the way it produced more than any human would have ever bothered to write.

A person writing badly under time pressure tends to write too little. They run out of patience before they run out of points. A model has the opposite failure. It never runs out of patience and it has no deadline. So it gives you the paragraph. Then it gives you the caveat on the paragraph, the worked example beneath the caveat, and a closing section that restates everything above it in case you skipped to the end. Everything and the kitchen sink goes in. The model has no view on what you could safely leave out, because forming that view is a task it was never trained to do.

Why the machine pads

The padding is not a flaw in any one model. It is what the training rewards. Modern assistants are tuned on preference data, where a human assessor looks at two answers and picks the better one. Put a thorough answer next to a terse one and an assessor skimming for completeness will choose thorough almost every time. Length reads as a proxy for effort. Caveats read as care. The hedge is never marked wrong, because a sentence that commits to nothing cannot be contradicted.

When researchers at UT Austin and Princeton took RLHF apart in 2023, they found that most of the reward gain came down to a single variable, the answer getting longer, and that a reward linked to length alone reproduced most of the improvement the fine-tuning was supposed to deliver. Train on enough of those judgments and you get a system that has learned, correctly, that more, not less, is safer and better.

Brevity is evidence that someone decided

A short document is hard to write because brevity requires making difficult choices. To cut a paragraph you have to know it does not earn its place, and to know that you need a view about who the reader is and what they are specifically looking for. Every omission is a small act of judgment. This is why a tight memo is rare and faintly intimidating to receive. It is the output of someone having worked out exactly what to exclude.

A model dumping everything into the answer is telling you, plainly, that it has not done this. It has not decided, because deciding means risking being wrong about what mattered, and the safe move is to hand over the lot and let you sort it out. The discernment becomes the new work. What lands in your inbox looks finished but is in fact merely unedited, which is a worse state than a rough human draft, because at least the rough draft knows it is rough. What we are now seeing routinely is voluminous unedited model outputs being passed on verbatim, masquerading as the finished product.

None of this is new

The complaint is older than the technology. In 1940, with considerably more at stake than a status update, Winston Churchill sent the War Cabinet a memo titled “Brevity”, objecting that the papers everyone had to read were far too long, that this wasted time, and that the discipline of setting out the important points concisely was an aid to clear thinking in its own right. He had spotted the mechanism. Bad length is more than tiring to read. It is a sign the writer has not finished thinking.

The American military codified bottom line up front, the rule that the main point goes in the opening line and the background waits its turn, because a commander should not have to reach page four to find out what the memo wants or suggests. Barbara Minto handed consultants the pyramid, a single governing thought up top with the support stacked underneath. And Amazon banned slide decks from its decision meetings in favor of a maximum six-page narrative read in silence at the start of the meeting.

Every one of these techniques was a response to humans being verbose. The machine has not introduced a new disease. It has just automated an old one, stripped out the friction that used to limit it, and handed everyone a powerful pump to fire this bloat out into the ether.

If slop describes generic, low-effort generative output, perhaps bloat is its textual counterpart, slowing you down as you wade through a 20-page document that should have been two. Netted out across author and reader, any efficiency gains are lost.

The fix is a wall, not a wish

You will not get a leaner document by asking the model to be concise and hoping for the best. Concise is a direction, not a constraint, and the prose quickly drifts back to safe and long inside a paragraph. The lever that works is a structure imposed from outside. Decide the form before it writes a one-pager or a memo, and give it the ceiling as a hard number. Make it state the recommendation in the first line, so the bloat has nowhere to hide. Push the supporting detail into an appendix recipients are free to ignore.

The word count is beside the point. What a ceiling does is force the model into a decision it would otherwise dodge, because once space runs short something has to be left out, and leaving things out is the whole point.

It helps to remember the time you save generating a document is not free. It is borrowed from every person who has to read it, at interest, and the longer the distribution list the worse the rate of return. A model will always offer you seconds. It will always have another paragraph ready, another caveat, another section that could go in. The discipline, the entire discipline, is being willing to push the plate away while there is still food on it.