Verbosity Bias
Length reads as rigor. It is produced by neither.
This page covers one specific way AI gets things wrong at work, and what to do about it.
It runs in order. What goes wrong, why it happens, where you'd notice it on an ordinary day, who takes the blame, roughly what it costs, and the check that catches it. Then one thing to try this week.
The dollar figures are estimates, not measurements. The assumptions behind each one are printed right there, so you can swap in numbers that fit your job. Anything actually measured carries an OBSERVED tag.
What is verbosity bias?
Preference tuning rewards responses raters judge as complete and helpful. Raters systematically prefer longer, more structured answers — a well-documented bias in human evaluation of model output.
So models produce length. Where substance is thin, the length is made of hedges, restatements, structural scaffolding, and elaboration of the obvious. Where substance is rich, the length is made of substance. Both look the same at a glance, and the glance is what most reviewers give it.
The downstream effect is the real cost: a fifteen-hundred-word response is not read the way a two-hundred-word one is. Volume suppresses scrutiny. The mode's danger isn't wasted words — it's that the words function as camouflage.
What do people assume?
That a longer, more detailed response reflects more thorough analysis. That a thin answer signals a thin basis and an extensive one signals depth.
Length is a stylistic output shaped by training preferences. It carries no information about the quality of the underlying reasoning.
Where does it show up at work?
A manager requests a competitive analysis. The output is 2,000 words, well-organized, with headers and bullets. Three of the eleven claims are substantive and sourced. The rest is inference presented in the same register.
Nobody separates the two, because separating them takes longer than reading it did.
Who carries the downside?
Vendor: none. Executive: makes a positioning decision on it. Manager: circulated it. You: produced it. When one of the eight unsupported claims is challenged, the whole document loses standing — including the three that were good.
What does it cost?
[MODELED — not reported]
ASSUMPTIONS Analytical documents / year: 60 Rate where unsupported claims pass review due to volume: 25% (~15 / year) Rate leading to a bad decision or credibility loss: 20% (~3 / year) Cost per incident: $5,000 – $50,000
Annualized exposure: ~$15,000 – $150,000
How do you control for it?
Claim extraction. Before circulating, list every discrete assertion in the document and mark each: sourced / inferred / assumed. Then either source the inferences or label them inline.
This is Paraphrase Hygiene (Part III, Concept 18) run on your own output.
CONTROL COST Analytical documents: 60 / year Claim extraction: 25 minutes each Annual: 25 hours Fully loaded rate: $70 / hour
Annualized control cost: $1,750
What should you do this week?
RECOMMENDATION
Cut your next AI-assisted document by 60% and label the epistemic status of everything that survives.
The short, labeled version is more defensible in a meeting than the long one, and it signals something the long one cannot: that you know which of your claims are load-bearing. Length is what people produce when they haven't decided what matters. Reviewers at senior levels read brevity as confidence, correctly.
Evidence
RESEARCH Length bias in human preference ratings of model output, and its transmission through RLHF.
ANALYSIS The "volume suppresses scrutiny" argument is the author's. Retain the label.
This is one of 24 failure modes. The book gives you all of them — plus the controls that catch each one and a 90-day plan to prove you ran them.
Preorder the Book