Check dates and freshness

Beginner · 5 min · Date claims and sources, and judge whether age matters.

research
verification

Every claim has a shelf life, and fluent text never looks old. A model’s knowledge stops at a training cutoff; search results keep years-old pages alive; a “2022 guide” still ranks today. Freshness checking is three moves: date the claim, date the source, and decide whether age matters for this claim.

Know which clock you are on

Model time — the training cutoff. Anything after it is invisible unless the tool retrieves live information.
Source time — when the page, report or dataset was published and last updated. The search result’s freshness and the source’s date are different things.
Subject time — how fast the subject moves. Tax rates, prices and product features change monthly; definitions, history and geography barely move.

Freshness only matters where the subject clock is fast. The skill is telling the three clocks apart — and matching your checking effort to the fastest one.

The freshness pass

Date the claim. Write what must be true, and as of when: “as of mid-2026…”. A claim with no date is a claim you cannot defend later.
Find the source’s date — the publication or last-update date, not when you happened to find it.
Check the subject clock. Fast-moving subject? Then an undated or old source fails regardless of how well-written it is. Slow-moving? An older source is fine.
Under-date, never over-date. “Current as of [month, year]” beats “latest” — you can keep the first promise; the second expires the moment you write it.

Weak prompt

Tell me about the latest AI regulations and current interest rates

Stronger prompt

Today is September 2026. For each regulation in force as of that date, give the rule, its jurisdiction, and the source with its publication or last-updated date. Flag anything you cannot date precisely.

Why the stronger one works

“Latest” and “current” carry no date — the model may answer from stale knowledge or a vague sense of now.
The strong prompt pins the present (“as of September 2026”), demands source dates, and flags undatable items instead of hiding them.
Undated answers become possible to review: every item either has a date or is explicitly flagged.

A bad example

Ship: a “state of the market” paragraph written from the model’s memory.

The sentences are confident, specific and — because the model’s world ends at its cutoff — possibly two years out of date. Nothing in the text signals it.

A better example

Ship: the same paragraph, except every fact carries a source with a date, and the section opens with “as of September 2026.”

Now staleness is visible at review time, and the next reader knows exactly when the picture was true.

Why it works

Staleness is invisible in good prose — that is precisely what makes it dangerous. Dating everything converts an invisible property into a checkable one: every claim either has a date or is a flagged gap. And the subject-clock rule keeps the effort proportional — you would not re-verify the speed of light, and you should not ship a tax number you cannot date. Cross-checking (previous lesson) and freshness checking are the two halves of grounding a claim in reality.

A draft says “67% of companies now use AI tools daily,” citing a source you cannot date. What is the disciplined next step?

Use it — the word “now” means the number is current Find the source’s publication or update date — and when the underlying data was collected; if it cannot be dated, it cannot be used Ask the model again — a second answer might sound fresher

Practice

Spot the stale claims

You are reviewing an AI-drafted briefing paragraph for a client newsletter about hybrid work and office costs. Nothing here is obviously false — but several items need a freshness check before the paragraph ships.

Select every item that needs a freshness check before publication. Leave the items whose age genuinely does not matter.

The AI answer

Hybrid work is now the default for tech teams. A 2021 industry survey reported that 71% of tech companies had adopted hybrid schedules, and the figure keeps climbing. Analysts expect the trend to accelerate through next year. Last year's office-rent benchmark put downtown vacancies at 18.4%. Remote work became widespread during the 2020 pandemic lockdowns, and its history is well documented. Austin, the city where the research team is based, sits in Texas. The briefing recommends budgeting using these figures.

Hint

Sort claims by their clock: survey numbers and forecasts move fast; history and geography barely move at all.

Relative time words — “now”, “keeps climbing”, “next year”, “last year” — are dates waiting to be pinned. Each one hides a freshness question.

Transfer

Reports: an “as of [date]” line at the top; source + date on every fast-moving number.
Briefs from AI: scan for “now”, “latest”, “recently”, “last year” — each one is a date waiting to be pinned.
Slow-clock facts (definitions, history, geography) need corroboration, not freshness.
Research with search tools: check the page’s update date before quoting it; an old page is fine only if the subject is slow.

Key takeaways

Three clocks: model cutoff, source date, subject speed. Match your checking to the fastest.
Under-date, never over-date: “as of March 2026” survives; “latest” expires on contact.
An undatable source is unusable for fast-moving claims — flag it, don’t ship it.

Next: Check the numbers — because figures need more than a date; they need recomputing. The training cutoff entry covers model time, and the primary source entry covers how to date a claim properly.