Are AI Companies Paid to Be Wordy? The Incentive Is Real, the Measurement Is Not

When an AI company charges by the token, every extra word earns money. That is the structural claim behind the question "are AI companies paid to be wordy?" — and it is not in dispute as a mechanism. But whether that incentive actually produces wordier answers is a measurement problem, and as of this edition, no one we can find has published a clean comparison of answer length and correctness under different pricing models.

The mechanism is simple. A provider that bills per token — roughly, per word — earns more when its model says "in summary, it's worth noting that the situation is somewhat complex" than when it says "it's complicated." That is not a conspiracy; it is arithmetic. However, the incentive is not uniform. Consumer chatbots are often flat subscription, where longer answers do not increase revenue directly. Enterprise contracts are frequently negotiated per seat, not per token. The padding incentive is strongest for metered API products, weaker for flat-rate products, and something else entirely under negotiated contracts. This split matters because any claim about "AI companies" as a block collapses under it.

The harder problem is measuring wordiness at all. Before anyone can say one model is wordier than another, the paper needs a definition. The obvious candidate is words per correct answer on a fixed benchmark, but that metric hides a variable: confidence policy. A model that answers "I don't know" in four words is shorter, but it also fails more questions if the benchmark demands a guess. If you count only words per correct answer, you are measuring verbosity and willingness-to-commit at the same time, and you cannot say which one moved. Fix the allowed error rate first, then count words. Otherwise the test is unfalsifiable.

The desk ran a web search for published comparisons of answer length and correctness under per-token versus flat pricing. It returned nothing usable across the lanes we checked. The closest public discussion we were handed is a Hacker News thread about a company called Vespper, which builds a document tool — a different company, a different product, and not evidence about AI billing. We cite it only to show where we looked. There is no academic study we can name that isolates the two variables, and we will not invent one.

Atelier is setting up exactly the test Mark described — comparing how quickly different models get to the right answer with fewer words — but it is a plan, not a shipped result. The pitch file on disk says so plainly. Atelier does not yet publish a benchmark for answer length versus correctness. Saying anything more would be a claim without a file behind it.

What would a good short answer look like? The same correct answer, minus the filler. The filler has a fingerprint: phrases like "in summary," "it's worth noting," "as previously mentioned," "to be clear," and restatements that add no information. But before the desk can measure filler, it has to agree on the list. That agreement is not a technical detail; it is the test. If we cannot say which phrases count as padding, we cannot measure wordiness, and the question stays rhetorical.

The kicker is not a verdict. It is a pre-registration. If Atelier runs this test, it should decide the benchmark, the error policy, and the filler lexicon before the first model answers — otherwise the result will confirm whatever the testers already believed.

Sources: https://news.ycombinator.com/item?id=49881505 (lead only, unrelated company); desk web search returned no citable study; the brain (Atelier’s own record)