ModelRefs / Verbosity — AI Glossary

Verbosity — AI Glossary

The tendency of instruction-tuned models to produce longer, more elaborate responses than necessary, often to satisfy RLHF raters.

Overview

RLHF creates a verbosity bias: raters often prefer longer answers that appear thorough, even if a concise answer is better. Manifests as padding, unnecessary caveats, and restating the question. Addressed in Claude's constitution and GPT-4o by penalizing padding and rewarding conciseness. Verbosity increases token cost and latency.

Reference details

Topic	prompting
Last reviewed	2026-06-24

How it works

Preference training collects human comparisons between candidate responses, and raters reading quickly tend to score the longer answer as more thorough.
The reward model fits that preference, so length becomes correlated with reward independently of whether the extra words carry information.
Policy optimisation against that reward model then discovers length as a cheap way to score well, since padding is easier to generate than substance.
The result shows up as recognisable patterns rather than as general wordiness: restating the question, enumerating caveats, announcing what the answer will cover before covering it.
Because the behaviour is learned from the reward signal, prompting can suppress its surface forms while the underlying tendency reasserts itself over a long conversation.
Corrections operate on the training signal — length-penalised reward, explicit conciseness preferences, or raters instructed to discount length — rather than on decoding parameters.

Example: Padding priced per day and per second

Suppose an answer that carries its content in 120 output tokens is delivered in 420: a restatement of the question, a preamble, and two caveats the user did not need. That is 300 tokens of padding and a response 420 / 120 = 3.5 times longer than necessary. At 200,000 requests a day the padding alone is 300 x 200,000 = 60,000,000 output tokens, which at $15 per million is $900 a day, or 900 x 365 = $328,500 a year — for text whose removal would not change a single answer. The latency cost lands on the user rather than the invoice, and it is larger than it looks. At a streaming rate of 50 output tokens per second, 420 tokens take 420 / 50 = 8.4 seconds while 120 tokens take 120 / 50 = 2.4 seconds, so padding adds exactly 6.0 seconds to every response. For an interactive assistant that is the difference between an answer that feels immediate and one the user watches arrive. Both figures point the same way and neither is visible in a quality evaluation that scores correctness, which is why verbosity tends to persist in systems that are measured on accuracy alone.

Commonly confused with

Verbosity is not thoroughness, and the distinction is whether the additional tokens change what the reader can do. A long answer that works through a genuine ambiguity, or shows the steps of a calculation the reader needs to check, is long because the content is. Verbosity is length that survives deletion without loss. The practical test is to cut and see whether anything was lost. It is also distinct from extended reasoning, where additional tokens measurably improve the answer on hard problems — those tokens are doing work, and in many deployments they are not even shown to the user. Treating reasoning tokens as verbosity and suppressing them degrades accuracy, which is why the two need separate budgets rather than one length limit.

When to use it

Reach for it when:

As a metric to track alongside accuracy, since an evaluation scoring only correctness cannot see it.
When specifying output format, where an explicit length or structure constrains the behaviour more reliably than asking for brevity.
In cost review of high-volume endpoints, where 300 padding tokens reached $328,500 a year at 200,000 daily requests above.
When tuning an interactive experience, given padding added exactly 6.0 seconds per response at 50 tokens per second.

Reach for something else when:

Confused with extended reasoning, where additional tokens improve accuracy and suppressing them costs quality.
As a criticism of genuinely complex answers, which are long because the content requires it.
Addressed by a blanket token cap, which truncates substantive answers along with padded ones.
Assumed fixed by a prompt instruction, which suppresses the surface forms while the learned tendency returns over long conversations.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Verbosity — AI Glossary.

Frequently asked questions

What is Verbosity?