Cannot Truly Cite
A citation is an invitation to verify. It is not a certificate of truth.
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 cannot truly cite?
Two distinct paths, both failing:
Without retrieval, citations are generated. The model produces a plausible-looking reference because references are what appear in that position in that kind of text. It may be entirely fabricated.
With retrieval, a real document is fetched — and then the failure moves. The retrieved source may not actually support the specific claim attached to it. The model summarizes, infers, and extends, and the citation follows the topic rather than the assertion. So you get a real source, correctly formatted, attached to a statement that source does not make.
This second failure is more dangerous, because it survives the check most people perform. They click the link. The link works. The document is real and on-topic. They do not read it to confirm it says the specific thing.
What do people assume?
That a citation attached to a claim means the source was consulted and supports the claim — that retrieval plus citation equals verification.
Retrieval improves the starting position. It does not eliminate the need for a finish line.
Where does it show up at work?
A policy analyst produces a briefing citing an agency guidance document for the proposition that a category of activity is exempt. The document is real, current, and about that subject.
It describes a narrower exemption with conditions the briefing omits. The citation is genuine; the claim it supports is not.
Who carries the downside?
Vendor: none. Executive: relies on the briefing. Manager: approved it. You: wrote it, and your name is on a document that misstates a regulator's own guidance.
What does it cost?
[MODELED — not reported]
ASSUMPTIONS Externally-cited claims / year: 150 Rate where source doesn't support the specific assertion: 10% (~15 / year) Rate caught in review: 50% Bad citations published: ~7 / year Cost per instance: $4,000 – $60,000 (correction, credibility, regulatory exposure)
Annualized exposure: ~$28,000 – $420,000
How do you control for it?
The Source-to-Claim Check (Part III, Concept 17). Nine steps, but the load-bearing one is step 3: locate the exact supporting passage. Not the document. The sentence.
Then: does the scope apply, is the authority sufficient, is it current, are there exceptions, does anything newer contradict it.
CONTROL COST Externally-cited claims: 150 / year Source-to-claim check: 8 minutes each Annual: 20 hours Fully loaded rate: $75 / hour
Annualized control cost: $1,500
What should you do this week?
RECOMMENDATION
New personal rule, effective immediately: you do not cite what you have not opened to the sentence.
Not the abstract. Not the summary. Not the snippet in the search result. The sentence.
This is the single habit that most separates people whose analysis holds up from people whose analysis gets withdrawn. It costs eight minutes. It is also, not incidentally, the habit that would have prevented the most publicly humiliating AI failures on record.
Evidence
REPORTED Fabricated citations in professional filings.
RESEARCH Citation-support failures in retrieval-augmented generation — real sources attached to unsupported claims.
ANALYSIS "A citation is an invitation to verify, not a certificate of truth" — author's formulation, established in the app handoff. Use as the chapter's anchor line.
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