AI Writes Three Times More and Says Less

Head-to-head testing on LM Arena turned up a result that will feel familiar to anyone who uses AI daily: one leading model produced roughly three times as much text as its peers, while scoring lower on linguistic complexity. More words, simpler words. Length had stopped tracking substance. If you have ever asked an AI a direct question and received four paragraphs that restate the question, hedge, list considerations you did not ask about, and then summarize themselves, you have met the problem. It is costing you more than you think.

Why AI pads

It helps to understand that this is not a bug someone forgot to fix. Models were trained to be helpful and thorough, and thoroughness is easy to imitate by adding words. During training, human raters tended to prefer longer answers, so length quietly got rewarded. And a model with no strong view about what matters hedges by covering everything. The result is an answer optimized to look complete rather than to be useful, which is a subtle but expensive difference. Padding is not the model failing. It is the model succeeding at the wrong target.

The three costs

Only one of these shows up on an invoice, which is exactly why the other two go unmanaged.

That last row is the one worth sitting with. Length reads as effort, and effort reads as authority. A padded answer is not just slower to read, it is more likely to be accepted without scrutiny, which is a quieter cousin of the workslop problem. And the reading time is a direct contributor to the AI management tax we measured recently.

Five instructions that actually work

Because padding is the default, you have to be explicit. Set a hard length ceiling with a number, not a vague request to be brief, since "concise" means nothing to a model that already thinks it is being concise. Ask for the conclusion first and the reasoning after, so you can stop reading once you have what you need. Tell it not to restate your question and not to summarize at the end. Ask for a shaped output, a table, a ranked list, a decision, because structure resists padding in a way prose does not. And when you find a phrasing that produces tight output, save it into your shared prompts, the same discipline as any good business prompt library, so nobody has to rediscover it.

Do not over-correct

Shorter is not automatically better, and forcing terseness on genuinely complex work causes its own failures. A model squeezed too hard will drop the caveat that mattered, or compress its reasoning into something you cannot verify. The target is not minimum words, it is no wasted words. Here is a test that takes ten seconds: read the output and ask which sentences you could delete without losing anything. If the answer is most of them, you have a verbosity problem worth fixing in your prompts. If cutting anything would lose real information, the length was earned. Judge output by what it tells you, not by how much of it there is.

Frequently Asked Questions