# Knowledge cutoff

> The date after which an AI model has seen no data, so it knows nothing newer.

- Level: Basic
- Topics: AI
- URL: https://tidbit.legios.com.ar/knowledge-cutoff/
- In Spanish: [Fecha de corte](https://tidbit.legios.com.ar/es/knowledge-cutoff/)

A model's **knowledge cutoff** is the point in time where its training data ends. An LLM learns from a huge snapshot of text collected up to a certain month; whatever happened after that is simply not in the model. Ask it about last week's news or the newest version of an app and it either says it does not know or, worse, guesses.

The gap cannot be avoided. Collecting data, training and testing a model takes months, so on the day a model is released its cutoff is already in the past, and it stays frozen while people keep using the model for a year or more. Providers usually publish the cutoff in their documentation. It is also blurrier than it sounds: the last months before the cutoff are thinly covered, because the internet had not finished writing about them yet.

The usual fix is not retraining but handing the model fresh information when you ask: a web search tool, RAG over your own documents, or simply pasting the relevant text into the prompt. When a chatbot gets a recent fact right, that is usually how. When it confidently states something outdated, the cutoff is a likely reason, so it is worth checking anything that may have changed.

## An analogy

Someone coming back from six months at a research base in Antarctica. They know plenty, but everything after the day they left is a blank, and if you ask about the news they might fill the gap with what seemed likely back then.

## Keep in mind

- Training data stops at a date; the model knows nothing after it.
- The cutoff is always months before the model's release.
- Search tools, RAG or pasted text give the model fresh facts.
- Outdated answers stated with confidence are a classic symptom.

## Go deeper

- [Knowledge cutoff on Wikipedia](https://en.wikipedia.org/wiki/Knowledge_cutoff)
- [Example: training data cutoff of each Claude model (Anthropic docs)](https://docs.anthropic.com/en/docs/about-claude/models/overview)

## Related terms

- [Training and inference](https://tidbit.legios.com.ar/training-and-inference/): Training builds the model once; inference is using it, millions of times a day.
- [LLM](https://tidbit.legios.com.ar/llm/): A neural network trained on huge amounts of text to predict the next word.
- [Internet](https://tidbit.legios.com.ar/internet/): A worldwide network of networks that lets any connected device reach any other.
- [RAG](https://tidbit.legios.com.ar/rag/): Let the model look things up in your documents before it answers.
- [Prompt](https://tidbit.legios.com.ar/prompt/): Everything you send to an AI model: the question, the instructions, the context.
- [Chatbot](https://tidbit.legios.com.ar/chatbot/): A program you talk to by typing or speaking, as if it were a person.