A knowledge cutoff means the date where an AI model’s training data stops. The model learned nothing that happened after it, so on its own it can miss or get wrong recent news, prices, products and people. Tools such as web search let a chatbot look up newer facts.
Think of someone back from a long expedition with no phone and no news. They are well informed about everything up to the day they left, and blank after it. Ask about last month and they will either say they do not know, or guess from what they remember. A model with a knowledge cutoff is that traveller, unless you hand it a newspaper.
In a sentence: “The chatbot hadn’t heard of the new law because it passed after the model’s knowledge cutoff, so I switched on web search.”
What the term means
The plain point: a model is trained once on a big pile of text, and the pile has an end date.
Anthropic, which makes the Claude chatbot, says: “While we’re constantly updating Claude’s data, each model has a knowledge cutoff”. Its help page warns that models “may not be aware of events or information that occurred after their respective cutoff dates.”1
Amazon’s cloud business, AWS, explains why: the training data is fixed. It lists “out-of-date or generic information” among the known problems of large language models.2
The date is not always one clean line. Anthropic’s model table lists two dates for each model, a “Reliable knowledge cutoff” and a “Training data cutoff”, and they are not always the same month.3
A simple example
The plain point: ask about something new and you may get an old answer, stated confidently.
Say a new phone came out last month and you ask a chatbot with search switched off which phone is newest. It can only name phones from its training data, and it may not say it is unsure. That is not lying; it answers from the patterns it learned, as hallucination meaning explains. This site does not list each model’s date here, because the dates change with every new model: check the maker’s own page.
For anything recent or changing, such as news, prices, laws or product versions, use a chatbot that searches the web, and open the sources it cites.
How web search and RAG help
The plain point: the model cannot relearn overnight, but it can be given fresh text to read before it answers.
Web search. Many chatbots can now look things up. OpenAI says ChatGPT “can search the web to answer questions with current information and links to relevant sources.”4 Google describes its search feature for Gemini as a way to “cite verifiable sources beyond its knowledge cutoff.”5
RAG. Retrieval-augmented generation is the same idea for your own documents: search first, then answer from what was found. AWS describes it as making the model reference “an authoritative knowledge base outside of its training data sources”.2 Tying an answer to supplied sources like this is called grounding.
Still check. Search does not make answers automatically right. OpenAI itself warns that “Search results and citations can be incomplete, outdated, or incorrect.”4
Often confused with
- Release date: the day a model came out. The cutoff comes before it, because training has to finish first.
- Context window: how much text a model can read at once in a conversation, not how up to date it is. See context window meaning.
- AI memory: a chatbot saving details about you between chats. It does not teach the model new world events. See AI memory meaning.
- “The AI knows everything online”: without search it knows only its training data. Common AI misconceptions covers this.
Questions people ask
Why does ChatGPT not know recent events?
Its model learned from data that stops at a cutoff date. With web search switched on it can look up newer information and link to it, as OpenAI’s help page describes.
How do I find a model’s knowledge cutoff?
Check the maker’s own documentation or help pages; Anthropic, for example, lists the date for each Claude model.
Learn more on this site
- Search first, then answer: RAG meaning and grounding meaning.
- Myths about what AI knows: common AI misconceptions.
- Why answers can be wrong: hallucination meaning.
- The words: training data meaning and the AI glossary.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is a knowledge cutoff?
- The date where a model’s training data stops
- The longest answer a chatbot can give
- The day a chatbot app was first released
Show the answer
Answer: The date where a model’s training data stops. The model learned nothing after that date, so on its own it can miss or get wrong recent news, prices, products and people.
2. What helps a chatbot answer about things after its cutoff?
- Asking the question in capital letters
- Web search, or handing it fresh text to read
- Asking it to try harder
Show the answer
Answer: Web search, or handing it fresh text to read. The model cannot relearn overnight, but web search or RAG gives it newer text to read before it answers.
3. Does switching on web search make answers automatically right?
- Yes, searched answers are always correct
- Yes, as long as the answer has links
- No, search results and citations can still be wrong, so open the sources
Show the answer
Answer: No, search results and citations can still be wrong, so open the sources. The page quotes OpenAI’s own warning that search results and citations can be incomplete, outdated or incorrect.
Related terms: Deep research · Few-shot prompting · Grounding · Latency · Machine translation · Multimodal
- 1Claude Help Center (Anthropic), How up-to-date is Claude’s training data?, updated 28 Sep 2026, read at source 5 Oct 2026“While we’re constantly updating Claude’s data, each model has a knowledge cutoff” … “These models may not be aware of events or information that occurred after their respective cutoff dates. If you ask about more recent events, the model may not have accurate information.” ↑
- 2AWS, What is RAG (Retrieval-Augmented Generation)?, read at source 5 Oct 2026“Additionally, LLM training data is static and introduces a cut-off date on the knowledge it has.” … “Presenting out-of-date or generic information when the user expects a specific, current response.” ↑a“Retrieval-Augmented Generation (RAG) is the process of optimizing the output of a large language model, so it references an authoritative knowledge base outside of its training data sources before generating a response.” ↑b
- 3Anthropic, Models overview, read at source 5 Oct 2026the table has separate rows headed “Reliable knowledge cutoff” and “Training data cutoff”, with different months for at least one model. ↑
- 4OpenAI Help Center, Searching the web with ChatGPT, read at source 5 Oct 2026“ChatGPT can search the web to answer questions with current information and links to relevant sources.” ↑a“Search results and citations can be incomplete, outdated, or incorrect. Open a cited source to check that it supports the answer”. ↑b
- 5Google AI for Developers, Grounding with Google Search, last updated 23 Sep 2026, read at source 5 Oct 2026“Grounding with Google Search connects the Gemini model to real-time web content and works with all available languages. This allows Gemini to provide more accurate answers and cite verifiable sources beyond its knowledge cutoff.” ↑
The comparison on this page (the traveller back from an expedition) and the new-phone example are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK