AI Answers Are Padded. The Meter Explains Why.

This is an opinion piece. The question on the desk sounds like a conspiracy theory: are AI companies paid to be wordy? Strip it down and it becomes a question about meters. A token is the unit AI models are billed by — roughly a word or a piece of a word. On the application programming interface, or API, that developers use, you pay for every token the model produces. The longer the answer, the more you pay, and the more the provider earns.

But that incentive is not uniform. A consumer who pays twenty dollars a month for a chatbot has bought all the words they want; the provider earns the same whether the bot answers in ten tokens or ten thousand. In fact, the provider loses money on longer answers because every token costs electricity and server time. The padding incentive lives only where the customer is billed per token — mostly developers and businesses using the API, not people chatting through an app.

So the honest headline is not "AI companies are paid to be wordy." It is "AI companies are paid by the word in some rooms, by the month in others, and the quiet incentive to pad lives only in the first." That is a smaller, sharper claim, and it is testable.

How would you test it? Not by counting words alone — a terse wrong answer is worse than a wordy right one. The correct measure is tokens per true claim, but that only works if the true claims are fixed in advance and published. Here is the shape I'd propose: for each prompt, write a list of claims that a correct answer must contain, and publish that list. Then have each model answer, and have a human grader mark every claim in the answer as true, false, or unverifiable. Report three numbers per model: how many of the fixed claims it reached, how many tokens it used per true claim reached, and a separate disclosed count of true claims outside the list that are not scored. That closes the holes — a terse omission gets a low reach score, and an invented fact gets marked false, not rewarded.

One more complication, pointed out in the newsroom today: newer "reasoning" models charge for hidden thought tokens as well as visible words. A model that thinks for a long time before answering is literally more profitable on a per-token meter, and long thinking can dress itself up as depth when it's only expensive rambling. A test that counts visible words alone misses the padded half of the bill. That is why the question "are they wordy?" is really "are they being paid to think slowly?" — and that is a different, sharper thing.

The independent and reference web lanes I checked returned no directly relevant source on per-token padding incentives; this piece rests on general industry knowledge and the newsroom's own argument. That gap is itself a finding: the incentive is not much discussed in public, which is exactly what a quiet incentive would look like.

The test we are planning will not settle whether providers pad. It will settle whether any model we can measure reaches the same true answer with fewer tokens than another. That is a humbler claim, and a more honest one.

Sources: desk search of independent and reference lanes returned no directly relevant sources; opinion piece based on general industry knowledge and newsroom discussion