Home / What a language model is, without the jargon
How Claude learned to answer, and up to when
Claude goes through three training stages before any conversation, a broad reading of the world, an adjustment into a useful assistant, then self correction grounded in written principles, and the resulting knowledge stops at a date that varies by model.
Claude learns nothing while it answers you. All of its knowledge comes from training completed before it was made available, and this training unfolds over several clearly distinct stages, each adding a different layer to what the model can do.
Three stages before any conversation
The first stage consists of reading a considerable amount of varied text to acquire a general language capacity, before it even knows how to hold a useful conversation. A second stage, called post training, turns this foundation into an assistant capable of answering an actual request rather than simply continuing a text. A third stage teaches the model to critique and revise its own answers based on a set of written principles, a reference text that sets out what counts as a useful and careful answer, then a second pass reinforces the choice of the least problematic answers by relying on these same principles rather than on human judgements collected one by one. The figure below reproduces these three stages in the order they occur.
A cutoff date that varies by model
Once these stages are complete, Claude's knowledge is fixed, it knows nothing of what happened after a certain date, called the knowledge cutoff. This date is not the same for every model in the family, the figure gives a precise example for one of them. Asking a question about a recent event with no context provided is like asking someone to comment on a news story they have never read.
Providing a fresh source
The fix is applied message by message, pasting the text of a recent document directly into the conversation before asking the question.
Here is the article published today by our association, pasted below:
[text of the article]
Answer only from this text, without relying on what you knew before.
This anchoring sentence does not repair the knowledge cutoff, it works around it for the single question asked. The principle seen in the previous lesson, the position of a piece of information changes its weight, also applies to a fresh source, it benefits from being placed close to the question rather than buried in the middle of a long message.
Three training stages, then a cutoff date specific to each model
An association treasurer asks his assistant to comment on the result of a professional election held the previous week, phrasing his question in a single, direct sentence.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: This situation establishes only that a direct question was asked about a recent event, the wording of the question not specifying whether a document about this event came with it.
What this does not establish: It does not establish that the assistant had information about this election, since nothing in the situation indicates that a recent source was provided in addition to the question.
The three most common miscalibrations
- Too broad This situation establishes that the assistant necessarily knows the result of this election, since Claude is trained on huge amounts of recent text.
- Too narrow This situation allows nothing to be affirmed, the treasurer's question being far too short to analyse.
- Beside the point This situation confirms that the treasurer's association uses a paid subscription giving access to the most recent models.
- Claude's knowledge comes solely from training completed before it was made available, in three distinct stages, reading the world, becoming a useful assistant, then self correcting on written principles.
- Each model in the family carries its own knowledge cutoff date, these dates are not identical from one model to another.
- Pasting a recent document directly into the message closes the gap between the knowledge cutoff and a question about current events.
- Self correction through written principles reinforces the choice of the least problematic answers without depending on human judgement collected answer by answer.
Choose a question you are asking yourself today about a recent event, write the message that pastes the corresponding source before the question, and send it to your assistant to compare its answer with and without this source.
These points depend on an interface or a rule that may have changed since this was written. Check them on your own screen before relying on them.
- The exact knowledge cutoff of your current model may differ from the example given, open the Transparency Hub page cited in the references to see the value in force for the model you are using.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Anthropic Transparency Hub consultée le 2026-09-02
- Claude's Constitution consultée le 2026-09-02