Knowledge Cutoff
The model is fluent about a world that stopped existing on a date it may not disclose.
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 knowledge cutoff?
Training data has an end date. Everything after it is invisible unless the system has live retrieval, and even then retrieval is a separate subsystem that may or may not fire, may or may not find the update, and may or may not be weighted above the model's parametric memory when the two conflict.
Three compounding problems:
1. The model's stated cutoff can be wrong. It's a fact learned from training data, not read from a config file. Models have misreported their own cutoffs. 2. Coverage thins near the boundary. Events in the final months before cutoff are underrepresented — the internet hadn't finished writing about them yet. So the effective cutoff is earlier than the stated one. 3. Retrieval doesn't override belief. When a retrieved document contradicts a strongly-learned parameter, the model does not reliably prefer the document.
What do people assume?
That the model knows it doesn't know. That when asked about something recent, it will say so.
It often will. It also often won't — and the tell is absent in both cases, because a confidently outdated answer and a confidently current one are formatted identically.
Where does it show up at work?
A benefits coordinator asks an assistant to summarize contribution limits for open enrollment communications. The model returns a clean, correct-looking summary — of last year's limits.
The email goes to 900 employees. Some of them make elections based on it.
Who carries the downside?
Vendor: none. Executive: none, until it's a compliance matter. Manager: owns the correction and the apology. You: you sent the email. Nine hundred people have a copy with your name in the signature block.
What does it cost?
[MODELED — not reported]
ASSUMPTIONS Time-sensitive external comms: 12 / year Rate relying on stale AI output: 15% (~2 / year) Correction + remediation cost: $4,000 – $12,000 Regulated-domain incident: 1 in 5 chance / yr Regulated incident cost: $25,000 – $150,000 (filing correction, counsel, potential penalty)
Annualized exposure: ~$13,000 – $54,000
How do you control for it?
A date check on every time-sensitive claim. One question: what is the effective date of this fact, and have I confirmed it against a primary source published after that date?
CONTROL COST Time-sensitive outputs: 12 / year Verification time: 20 minutes each Annual: 4 hours Fully loaded rate: $65 / hour
Annualized control cost: $260
What should you do this week?
RECOMMENDATION
Build an Authoritative Source Register (Part V) for your function: the five to ten sources that are actually definitive for your domain — the regulator's own page, the vendor's own docs, the primary filing. Not a search result. Not a summary. The source of record.
Then make one rule: anything with a date in it gets checked against the register before it leaves the building.
You now own the source of truth for your team. That is a small amount of political capital that compounds, because everyone who needs a number now comes through you.
Evidence
RESEARCH Models misreporting their own training cutoff, and effective cutoffs preceding stated cutoffs due to thin recent coverage.
RESEARCH Retrieval failing to override parametric knowledge under conflict.
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