THE GAP
Cutoff first, release later
In one benchmark study, a model released in August 2025 had a developer-stated cutoff of November 2024: about nine months of news it never saw.
See the numbers
WHAT THE RESEARCH FOUND
Released today,
trained earlier.
A new AI model is not a fresh one.
Language models learn from text gathered up to a cutoff date, then spend months being tested and tuned before release. Anything that happened after the cutoff is simply not in what they learned.
CUTOFF before release day
Check a question- Learned up to a date
- Released months later
- Stated date is a guide
- Ask how fresh the answer must be
THE SHORT ANSWER
Every model has
a last day of learning.
The release date tells you when you can use it.
The cutoff tells you what it could have read.
THE MIX
Old pages in new data
Researchers found web crawls are full of older copies. In one training set, over 80% of Wikipedia documents came from versions before 2023.
Why it happens
THE RISK
Fluent but out of date
When guidance changes, a model trained before the change can repeat the old version confidently. A 2025 study measured this with revised guidelines.
Check a question
FOUR HABITS TO TRY
Use the model,
check the date.
None of these needs technical skill. They are about noticing when an answer depends on something that may have changed.
Find the stated cutoff
Developers often publish a knowledge cutoff date. Note it, and treat it as an upper bound, not a promise.
Don't rely on the model's own answer
Asked about their own cutoff, models in the LLMLagBench tests named dates a year or more off.
Ask whether it searched
Some tools fetch live web pages; many models do not. If no source is shown, assume the answer comes from training data.
Confirm time-sensitive facts
For rules, prices, results and guidance, check the official page and its date before acting on it.
TWO NUMBERS WORTH KNOWING
The stated date
is only a start.
One figure on the gap before release,
one on how old training data can be.
≈ 9 months
Stated cutoff to release, one model
The benchmark asked models 1,713 questions about dated news events. One model released in August 2025 had a stated cutoff of November 2024, and asked directly, it named a date a year earlier still.
CUTOFF TO RELEASE · LLMLAGBENCH
80%+
Wikipedia pages from older versions
Researchers found that in one open training set, over 80% of Wikipedia documents came from versions before 2023, because web crawls keep old copies.
OLD DATA · DATED DATA STUDY
The Dated Data study calls the date a model's knowledge really reflects its effective cutoff, and found it often differs from the date developers report, and can differ from topic to topic.
A QUICK FRESHNESS CHECK
How fresh does
the answer need to be?
Pick the kind of question you are asking. The bar shows how much a knowledge cutoff can matter for it.
Based on the UK Government AI Playbook and published research on knowledge cutoffs. General information, not professional advice.
WHY THE DATE IS FUZZY
Old pages hide
in new crawls.
Much training text comes from regular web crawls. The Dated Data study found a 2023 crawl still holds copies of Wikipedia pages dating back to 2016, so the data leans towards older versions even when new dumps are added.
Removing duplicates can shift it further. Researchers also found that the stated cutoff and the effective cutoff can differ by topic and source, which is why freshness benchmarks probe models with questions about dated events.
Questions
Questions people ask
What is a knowledge cutoff?
It is the date up to which a model's training data was gathered. The Dated Data study notes that released models are often paired with a claimed cutoff, but what the model actually knows can reflect an earlier date.
Why is the cutoff before the release date?
After the data is gathered, a model is trained, tested and tuned before launch. In the LLMLagBench study, one model released in August 2025 had a stated cutoff of November 2024.
Can I just ask the model when its cutoff is?
Not reliably. In LLMLagBench tests, one model named a cutoff more than a year earlier than the benchmark's estimate, and another was about two years off.
Does a bigger model fix this?
No. The 2021 'Mind the Gap' study found models did worse at predicting text from after their training period, worse with time, and that size alone did not solve it. Updating models with new information helped.
How do some AI tools give current answers?
Some tools search the web or documents and pass the results to the model, an approach called retrieval augmented generation. The UK Government AI Playbook notes that many models still have no real-time access.
Sources
- Cheng et al. (2024): Dated Data, Tracing Knowledge Cutoffs in Large Language Models
- Pęzik et al. (2025): LLMLagBench, Identifying Temporal Training Boundaries in LLMs
- Lazaridou et al. (2021): Mind the Gap, temporal generalisation in language models (NeurIPS)
- Cacioli et al. (2025): Do Knowledge Cutoffs Drive Clinical Accuracy? (NeurIPS workshop)
- GOV.UK: Artificial Intelligence Playbook for the UK Government
Updated