Learn AI /Training and Model Adaptation /Knowledge cutoff
Knowledge cutoff
- Starter
- Training and Model Adaptation
- Communications teams
- Policy officers
Also called: training cutoff, cut-off date.
Back to the milestone you came from
In five seconds
What it is
The approximate point up to which a model’s training-derived background is expected to be current, while newer events and facts may be absent unless added another way.
In plain words
It is the approximate date up to which the model’s training-derived background may be current. Newer facts may be missing unless added another way.
An everyday example
A teacher in Dakar gives students a printed world atlas published in 2024. In class, a student asks about a railway line opened in 2025. The atlas can still help with country names, borders shown at publication, and older transport routes, but it will not include the new line unless the teacher also brings a newer handout. If a student guesses the answer from older maps, the guess might sound sensible and still be wrong.
A printed atlas is fixed and visible, while a model stores patterns in numerical parameters rather than pages you can inspect directly. Also, an atlas does not generate new sentences, but a model does, so it can produce fluent guesses about post-cutoff events even when it has no dependable built-in record of them.
Remember thisA knowledge cutoff marks the approximate recency of the model’s training-derived background. It does not limit prompt length or prevent the model from using current information supplied through approved retrieval or context.
Understand it
Understand it
Why it matters
The cutoff affects whether an answer may be outdated. That matters when you judge risk, choose whether to trust a draft, and decide when you need current documents or human checking. It also helps you ask better questions of vendors: what date does built-in knowledge reach, and what approved method adds newer information? Without that clarity, people may assume a fluent answer is current when it is not.
What a knowledge cutoff does and does not mean
After viewing, the reader should be able to explain that the cutoff limits built-in training knowledge, not the amount of text a model can read now, and not whether current documents can be added separately.
01
1. Data before the date
Show a calendar marker such as 'cutoff: June 2024'. Only source material from before that point goes into training.
Draw a stack of documents on the left with small date tags before the cutoff line; place a bold vertical date marker beside them.
02
2. Training shapes the model
The model’s parameters are adjusted using those earlier sources. This forms its built-in background patterns.
Connect the document stack to a centre box labelled 'training', then to a box labelled 'model parameters'.
03
3. Ask about something newer
A user asks about an event after the cutoff, such as 'What changed in 2025?' The built-in knowledge may be incomplete or outdated.
Place a question bubble to the right with a date after the cutoff; connect it to the model box.
04
4. Two answer paths
Without added current sources, the reply may be uncertain or wrong. With approved current documents retrieved at use time, the answer can be grounded in newer material.
Split into two arrows: top arrow directly to an answer box marked 'may be outdated'; lower arrow passes through a box labelled 'current documents retrieved now' to an answer box marked 'uses newer sources'.
A left-to-right flow diagram with four boxes: older data collected before a marked date, training, the model, and a user question about a later event. A side branch shows optional current documents added at use time. Read top row left to right, then compare the answer path without retrieval and with retrieval.
Hover or tap each box to reveal a short note: 'training-time knowledge', 'user-time question', 'not the same as context window', and 'retrieval can add newer information'.
Describe this without the diagram
The diagram is a four-stage flow. First, documents dated before a clear cutoff line are collected. Second, those documents are used in training, which shapes the model’s parameters. Third, a user asks about an event dated after the cutoff. Fourth, the diagram splits into two paths: if no current documents are added, the answer may rely on older training patterns and may be outdated; if current documents are retrieved at use time, the answer can use newer information. The diagram also signals that this is different from the context window, which is not shown as a date limit.
With reduced motion: A static left-to-right flow with the same four stages and the two final branches visible at once, using labels and colour contrast instead of animation.
A model is trained on large collections of text and other data. Training usually uses selected material assembled up to an approximate date or period. This is often called the knowledge cutoff, training cutoff, or cut-off date.
What goes in is selected training data: books, websites, articles, code, and other sources mostly collected before the cutoff. During training, the model’s parameters, the internal numerical settings shaped by that process, are adjusted using the data. What comes out is a model that can generate useful text, but whose training-derived factual background may be outdated.
This does not mean the model cannot discuss later events. It may produce guesses, repeat patterns, or use information supplied in the prompt. But without a dependable external source, later claims should not be treated as reliable.
A context window is the input and prior output the model can use at one time. Retrieval-augmented generation can supply relevant passages at use time, including newer material. A model version is a named release whose documentation may state a cutoff. The cutoff itself concerns approximate temporal coverage of training.
How “A school atlas printed last year” maps onto the real thing
| In the story | In the AI system |
|---|---|
| atlas published in 2024 | training data collected up to a cutoff date |
| new railway line opened in 2025 | facts or events after the cutoff |
| teacher adds a newer handout | current information supplied through retrieval or the prompt |
| student guess based on older maps | model output that sounds plausible without being up to date |
A worked example
A desk officer at Headquarters asks an approved assistant to summarise changes in a revised draft circulated last week, for talking points needed before an afternoon first reading.
Starting from: User prompt: “Summarise the main changes in the revised draft circulated last week.” No draft or link is provided.
- The model’s stated knowledge cutoff predates circulation of the revised draft.
- The prompt does not contain the draft, and the assistant has no retrieval connection to the document system.
- The model may rely on an earlier version, infer plausible changes or invent details that sound consistent with previous text.
- The officer supplies the authoritative circulated draft through an approved method and asks the assistant to compare it with the earlier version, with references to the relevant paragraphs.
Result: The cutoff explains why the model cannot reliably know the revised text from its training background. Once the current draft is supplied at inference time, the assistant can work from that text, subject to access controls and human verification.
In a United Nations workplace
During a lessons-learned exercise after an assistant pilot, an ombudsman case officer at a shared services centre reviews a draft note for the file before close of business. The draft describes an intake workflow that was replaced after the model’s knowledge cutoff.
The newer workflow did not enter the model’s training-derived parameters merely because it was circulated internally. The officer must provide the current instruction through an approved source or correct the draft directly. A longer prompt alone does not make the model’s background more current.
Before you try thisDo not put names, contact details, allegations, case narratives or attachments into a tool that is not explicitly approved for ombudsman case material. For an approved tool, use de-identified extracts and the minimum necessary content under the office’s confidentiality requirements.
Where you may hear it
- A vendor demo during an AI procurement meeting
- A lessons-learned session after an assistant pilot
- An ICT governance board review of a model card
- A staff briefing before an approved assistant is rolled out
Avoid confusion
Avoid confusion
Often confused with
Context window
The maximum amount of input and prior output a language model can use at one time while producing the next part of a response.
Retrieval-augmented generation
A method that first finds relevant source passages and then asks a language model to answer using that material.
Model version
A particular identified state of a model or model-backed system, with documented changes, capabilities and checks for use.
Common misconceptions
Common misconception: If the cutoff is old, the model cannot say anything about newer events.
Correction: It can still generate text about newer events by guessing from older patterns or by using information supplied in the prompt or retrieved from another source. The issue is reliability, not silence.
Common misconception: Knowledge cutoff means the system stops reading after a certain date in the conversation.
Correction: That confuses cutoff with context window. The context window is the amount of text the system can process in one exchange. Knowledge cutoff is the latest date of the data used during training.
Common misconception: If a model mentions a cutoff date, every fact before that date is accurate.
Correction: No. Training data can contain errors, gaps, bias, or conflicting information. A cutoff only tells you the latest likely training period, not the quality of everything before it.
Common misconception: Connecting a model to the internet removes the cutoff.
Correction: It can reduce the problem for some questions by adding newer sources at use time, but the model’s built-in training background still comes from an earlier period. The quality then depends on the retrieved sources and system design.
Quick check
Which statement best describes a knowledge cutoff?
A knowledge cutoff is about the time range of training data. It is not the context window, not an account setting, and not a guarantee of accuracy. Even before the cutoff, errors can remain; after the cutoff, newer facts may need retrieval or human checking.
Go deeper
Go deeper
The simple version
It is the approximate date up to which the model’s training-derived background may be current. Newer facts may be missing unless added another way.
The technical version
During pretraining, a model is optimised on a corpus assembled up to an approximate time boundary. That boundary is commonly described as the knowledge cutoff. Information after that date is generally not represented in the model’s training-derived parameters. Newer information can be incorporated through additional training or fine-tuning, or supplied at inference time through retrieval or user-provided context. The term refers to temporal coverage of training data, not context capacity, conversation history, or automatic access to live external sources. Because models generate outputs from learned statistical patterns rather than a dated fact table, post-cutoff answers may still appear confident even when unsupported by current evidence.
Reviewed external resources
EncyclopaediaKnowledge cutoff (opens in a new tab)Wikipedia
This Wikipedia link gives a broad introductory overview of the term and suits readers who want a quick general starting point.
Intermediatelink checked 2026-07-31
CourseLLM Course (opens in a new tab)Hugging Face
This Hugging Face course link helps readers place the term within how large language models are trained and used, and suits beginners who want structured learning.
Deeperlink checked 2026-07-31
CourseShort Courses (opens in a new tab)DeepLearning.AI
This DeepLearning.AI course link offers short practical lessons around modern AI systems, and suits readers who prefer guided video-based learning.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Knowledge cutoff (opens in a new tab)YouTube search
This YouTube search results page helps readers find a range of video explanations of the term, and suits people who learn best by comparing presenters; it is a search page, not a single vetted source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Knowledge cutoff (opens in a new tab)Reddit search
This Reddit search results page helps readers see common questions and confusions people have about the term, and suits readers who want informal discussion; it is a search page, not a single authoritative source.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
A knowledge cutoff marks the approximate recency of the model’s training-derived background. It does not limit prompt length or prevent the model from using current information supplied through approved retrieval or context.