What is a knowledge cutoff?
A knowledge cutoff is the point in time after which a model has seen no training data. Anything that changed after that date, such as a new law, a price, a software release or a news event, is invisible to the model unless it can search the web or you provide the information.
Every model has one, and models rarely state it clearly or notice when a question depends on it.
Why it causes confident mistakes
A model that does not know something is new will often answer from what it does know. The result reads as fluent and current, but describes the past. Typical cases:
- A regulation that was amended, with the model describing the old rule.
- A product or plan that was renamed, repriced or discontinued.
- A software version with changed syntax or settings.
- A recent event the model answers about as if nothing happened.
- A "best" or "latest" ranking that was true when the data was collected.
Why a panel does not remove it
Models with similar cut-off dates tend to miss the same changes, so they can agree on an outdated answer. This is a common source of false consensus. Models with live web access may be fresher, which is one reason a mixed panel can show useful disagreement: a model with search and a model without can answer differently simply because one can see newer pages.
How to check recency
- Ask what date the answer reflects. A model that cannot say is a warning.
- Look for a source with a date. An official page, changelog or notice that you can open beats a model's memory.
- Search the thing that could have changed. For a price, the vendor page. For a rule, the official publication.
- Distrust "latest" and "current." Treat any superlative about now as a claim to verify.
- Give the model the fresh information when you have it, and ask it to reason from that.
When it matters most
The cost of a stale answer grows with how fast the topic moves and how much is at stake: law and taxes, medicine, pricing and contracts, security advisories, software dependencies and anything involving a deadline.
Try it on your own question
Run the question you care about through six independent models and read where they agree and where they split. Ask 6 AIs on Satcove, free, no card required.
Frequently asked questions
How do I know a model's cutoff? Providers usually publish it, but models are not reliable at reporting it themselves. Check the provider documentation and test with a question you know changed recently.
Do models with web search avoid the problem? They reduce it for what they find, but search can return outdated or low-quality pages too. Check the cited source and its date.
Can agreement across models fix stale data? No. Models with similar cutoffs often share the same gaps. Verify time-sensitive claims against a primary source.