Is what you're reading about AI out of date?

Usually, yes, at least in part. Here is how to tell in about a minute, without knowing the subject better than the person who wrote it.

Last verified 11 August 2026 · This page contains no figures on purpose, so there is nothing on it to go stale

Why AI writing rots faster than other technical writing

Every field has stale articles. This one has a particular problem: the release cycle is measured in months, the prices move without notice, and the thing being described changes shape while people are still writing the first guides to it. An article about how to fix a tap ages gracefully. An article about which model to use ages like fish.

There's a second, nastier mechanism on top of that. A great deal of writing about AI is now produced with AI, and a model's knowledge stops at a fixed date. So a stale claim gets absorbed into training data, restated confidently in a new article with this year's date on it, absorbed again, and repeated. The date at the top gets fresher while the claim underneath does not. Freshness signals and accuracy have come apart, and most readers are still using the first as a proxy for the second. → Training cutoff

The tells

None of these prove a page is wrong. Each one raises the odds enough to be worth a check.

No date anywhere. The big one. A page that won't tell you when it was written is asking you to assume it's current, which is exactly the assumption that costs you. Undated is not neutral; treat it as old.

A copyright year in the footer, presented as a date. That number usually comes from a template and updates itself every January without a human reading a single word of the article. It tells you the site still exists. It tells you nothing about the page.

"As of this writing," "currently," "at the time of writing." A writer flagging that a fact is time-sensitive, and then not telling you when the writing happened. Where a date does appear elsewhere, these phrases are a useful map of exactly which sentences to distrust first.

Model names and version numbers. The fastest clock on the page. If the newest model you have heard of is missing, or a model you know was retired is described in the present tense, you are reading something written before that changed. This works even if you only vaguely follow the field.

Screenshots of an interface. Products get redesigned constantly. A screenshot showing menus that no longer exist dates the piece more honestly than its byline does.

Confident claims about what AI "can't do." The shortest shelf life of any sentence in this field. Capability claims are worth checking against something released since, always.

A tidy rule of thumb, quoted without a source. Round, memorable conversion figures spread faster than they get re-checked, and they keep circulating for years after the underlying systems changed underneath them. This site's Tokens page works through one that is repeated nearly everywhere and is now materially wrong.

An "updated" date with no note of what changed. Sometimes genuine. Sometimes a timestamp nudged to please a search engine. A page that says what it corrected is making a claim it can be held to; a bare date is not.

The sixty-second check

When something actually matters, and especially before you spend money or write code against it:

Find the date, and treat a missing one as a red flag rather than a formality.
Take the single load-bearing fact and go to the primary source. For prices and limits, that means the vendor's own pricing or documentation page, not another article. It is nearly always one search away, and it takes less time than reading the rest of the piece.
Check the primary source is itself dated. Vendor docs go stale too, though far less often, and the good ones publish a last-updated date.
Search the page for the newest thing you know about. If a recent, well-covered release is nowhere in a piece that claims to survey the field, you have your answer without evaluating a single claim.
If it's a tutorial, check that the identifiers still exist. Model IDs and API names get retired. Stale code is uniquely nasty because it looks completely correct and fails for reasons that seem unrelated.

The shortcut that covers most cases

Numbers, names and prices: verify against the vendor, every time. Explanations of how the technology works: much safer to trust from an older page, because the concepts move slowly. Most of the damage comes from treating those two categories as one thing.

Which claims rot fastest

Roughly, from fastest to slowest:

Prices. Can change with no notice, and promotional rates expire silently. The most dangerous category, because it looks like a hard fact.
Model names, versions and IDs. New ones land every few months, old ones get retired.
Capability claims. What a model can and can't do, and which one is "best."
Interfaces and workflows. Where the buttons are, what the settings are called.
Limits, like context sizes. Slower moving, and generally only in one direction.
Concepts. Stable for years. What a token is has not changed.

This ordering is why every hard number on this site is quarantined onto a single page, Model facts, and why the explainers teach the idea and link out for the figure. One page rots instead of a hundred, and the pages doing the actual teaching stay correct.

Can't I just ask an AI whether it's current?

Not usefully, and it's worth understanding why, because the failure is quiet rather than obvious.

A model without a search tool has no way to know what changed after its training data ended. Worse, it cannot reliably tell you where that boundary falls: for the model to know its own cutoff, that date would have had to be documented before the model was trained, which it obviously wasn't. Asked anyway, it will infer and answer with the same steady confidence it applies to everything else. → Hallucination

A model with a working search tool is a genuinely different proposition, because it can go and read the current page. That's the whole point of tool use. But it turns a recall problem into a reading problem rather than eliminating it, so the useful habit is to ask for the source and the date, then look at the source. If it can't produce one, you've learned something. → Tool use

What a page that's trying to be honest looks like

Four habits, in rough order of how much they tell you:

It's dated, per page. A date doesn't make a page current. It makes the page's currency checkable, which is the part that lets you calibrate.
Figures are cited to primary sources, with the date they were fetched.
Gaps are left visible rather than filled in. An empty cell is information. A plausible-looking wrong number is a liability, and plausible-looking is exactly what a language model produces when it doesn't know.
Corrections are stated, not silently patched. This is the strongest signal of the four, and the rarest, because it costs the writer something.

Those are this site's own rules, so treat the list as a description of what it is trying to do rather than a neutral survey. The changes log exists so you can check whether it actually does it, including the corrections. Apply the same test here that you'd apply anywhere else: this page is dated at the top, and if the date is old and the log below it is empty, discount accordingly.