# Knowledge cutoff notice > The answer is stated in the present tense about a world the model last saw a year ago. - Canonical: https://patterns.konigi.com/ai-assistants/knowledge-cutoff-notice - Group: Grounding and disclosure - Level: implementation - Status: published - Updated: September 12, 2026 - Also called: training date, staleness warning, may be outdated --- ## Problem Someone asks who runs a company, what a product costs, or whether a library still supports a feature. The answer comes back fluent, specific and in the present tense, describing a state of the world that stopped being true some time after the model finished training and before the question was asked. ## Solution Surface the boundary at the moment it becomes relevant rather than as a permanent disclaimer. A footer saying the model may produce inaccurate information is read once and never again. A line on a specific answer saying the information predates a stated date is read every time, because it arrived attached to something the viewer cares about. The trigger is the design problem. Flagging every answer is noise; flagging none is the status quo. The workable heuristic keys on the question rather than the answer: anything asking about current state, prices, people in roles, versions, availability, or events is time-sensitive by construction, and those are detectable. So is a question containing a recent year. Once a model can search, the notice changes shape rather than disappearing. The useful distinction becomes whether this particular answer was checked against live sources or produced from training data. Those two look identical on screen and differ completely in reliability, and a viewer can't tell them apart without help. The presence of a source list is the strongest available signal, which makes the absence of one meaningful in a product that usually shows them. Precision matters more than it seems. A cutoff is a property of a specific model, and products offering several are offering several cutoffs. Providers publish these, and a product that lets people choose a model and then hides the date has left the most decision-relevant property out of the picker. HAX guideline 2 asks systems to make clear how well they can do what they do, and a training cutoff is one of the few limits that can be stated exactly rather than estimated. That precision is worth using, because "may be outdated" carries almost no information while a date lets someone reason about whether it matters to their question. ## Use when The model answers from training data about things that change, and especially where a product also has a search mode that may or may not have run. ## Don't use when The answer is drawn from supplied material or retrieved sources, or the task is generation rather than recall. A cutoff notice on a rewritten paragraph is noise attached to an operation the cutoff doesn't affect. ## Trade-offs An honest notice lowers confidence in answers that are often correct, and the products that ship it most prominently look less capable than ones that stay quiet. Triggering on time-sensitive questions requires classifying the question, which adds latency and misclassifies. A permanent disclaimer is cheap, universally ignored, and mostly serves the people who wrote it. Stating a date also invites a reasonable follow-up the product usually can't answer: what the model knows about the weeks just before it. Coverage thins out well ahead of the stated boundary. ## Checklist - Does the notice appear on time-sensitive answers rather than all of them? - Is a specific date given rather than a vague warning? - Does the date track the model actually used? - Can the viewer tell whether this answer was checked against live sources? - Is the absence of sources meaningful in this product? - Does the model picker show cutoffs alongside the names? - Is there a one-click path from the notice to a search-backed answer? - Does the notice survive copying or sharing the answer? - How does it read on the tenth occurrence in one session? - Is the notice suppressed when the answer came entirely from supplied material? ## Compare **Claude** publishes per-model training cutoffs in its documentation, which makes the property checkable rather than folkloric, and leaves surfacing it to the product built on top. **ChatGPT** mostly resolves the question by searching when a query looks current. A disclosure problem becomes a retrieval one, and the viewer is left to infer from the presence of sources. **Perplexity** sidesteps the pattern almost entirely by retrieving for nearly everything, so its equivalent risk is a stale source rather than a stale model. **GitHub Copilot** shows the same limit in its sharpest form, suggesting code against the library versions present in its training data, so a stale cutoff surfaces as a deprecated API call rather than as a wrong fact. ## Related Citation chip and source list are the strongest evidence that a given answer went beyond training data. Mode switch is what a viewer reaches for once the notice tells them the answer needs checking. Refusal is the stronger response when the gap is too large to caveat. Model picker is where the cutoff belongs and is usually missing. ## Related patterns - [Citation chip](https://patterns.konigi.com/ai-assistants/citation-chip) — A well-sourced paragraph and a fabricated one read identically until the viewer can see where each claim came from. - [Source list](https://patterns.konigi.com/ai-assistants/source-list) — The answer drew on eight pages, and the viewer needs to judge those pages before trusting it. - [Mode switch](https://patterns.konigi.com/ai-assistants/mode-switch) — The same question wants a fast answer some days and a researched one others. - [Refusal](https://patterns.konigi.com/ai-assistants/refusal) — The assistant will not do the thing, and the viewer needs to know whether to rephrase or give up. - [Model picker](https://patterns.konigi.com/ai-assistants/model-picker) — Several models sit behind one box, and the difference between them stays invisible until the answer arrives. ## Sources - [Anthropic, Models overview](https://docs.anthropic.com/en/docs/about-claude/models/overview) - [Microsoft HAX Toolkit, Guideline 2, Make clear how well the system can do what it can do](https://www.microsoft.com/en-us/haxtoolkit/guideline/make-clear-how-well-the-system-can-do-what-it-can-do/) - [Google PAIR, Explainability and Trust](https://pair.withgoogle.com/chapter/explainability-trust/) --- Screenshots on patterns.konigi.com are reproduced for commentary and criticism. Product names and marks belong to their owners.