Difficulty With Negation
'Must not' and 'must' differ by one word. That word is frequently the whole control.
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 difficulty with negation?
Negation is a small lexical signal carrying a total semantic reversal. In distributional terms, "the report may be shared" and "the report may not be shared" are extremely similar strings, and the training signal for handling the difference correctly is weaker than the frequency of the surrounding pattern.
This compounds in three ways:
1. Scope ambiguity. What exactly does the negation cover? "Employees may not use personal devices to access customer data outside secure environments" — the negation's scope is genuinely unclear even to careful human readers. 2. Double negation and exceptions. "Except where prohibited" constructions invert meaning again. 3. Summarization loss. When condensing, negations are the most frequently dropped element, because they're syntactically small and carry no topical weight.
That third one is the operational killer. A summary of a policy is disproportionately likely to lose the prohibition — which is the only part of the policy that constrains behavior.
What do people assume?
That a system parsing a prohibition understands it as a prohibition — that "may not disclose" and "may disclose" are handled as opposites.
Negation is handled less reliably than affirmation, and the failure is silent.
Where does it show up at work?
A team asks an assistant to summarize a data handling policy into a one-page quick reference. The summary is accurate and readable.
The clause reading "may not be transferred outside the approved region without written authorization" becomes "may be transferred with authorization." Substantially similar. Operationally inverted — one describes a default prohibition, the other a default permission.
Who carries the downside?
Vendor: none. Executive: attests to compliance. Manager: distributed the quick reference. You: produced it. In a regulated transfer, this is a reportable event.
What does it cost?
[MODELED — not reported]
ASSUMPTIONS Policy / requirement summaries: 40 / year Rate w/ a dropped or inverted prohibition: 10% (~4 / year) Rate caught in review: 50% Bad summaries in circulation: ~2 / year Cost per instance: $10,000 – $200,000 (regulated data incident, remediation, penalty exposure)
Annualized exposure: ~$20,000 – $400,000
How do you control for it?
Explicit constraint extraction. Before accepting any summary of a policy, contract, or requirement, list the prohibitions separately — every "must not," "may not," "shall not," "except," "unless," "prohibited," "restricted."
Then verify each one against the source text individually. Prohibitions get their own pass, because they are the part that fails.
CONTROL COST Policy summaries: 40 / year Constraint extraction: 20 minutes each Annual: 13 hours Fully loaded rate: $80 / hour
Annualized control cost: $1,040
What should you do this week?
RECOMMENDATION
Take the policy your team relies on most and extract every prohibition into a standalone list. Compare it against whatever summary or quick-reference your team actually uses day to day.
If anything is missing or softened, you have found a live compliance gap and produced the artifact that closes it. This is Compliance Mapping (Part III, Concept 21) at individual scale, and it is the highest-stakes twenty minutes in Part II.
Evidence
RESEARCH Negation handling as a documented weakness in language models.
ANALYSIS The claim that summarization disproportionately drops negations is the author's and is the operationally critical assertion in this chapter. Verify empirically or label clearly — it should not carry a research label without support.
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