{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"the-cut-off-and-the-clock","title":"It has no clock, and its knowledge thins before it stops","url":"https://addaly.com/learn/ai-actually-explained/the-cut-off-and-the-clock","position":35,"of":92,"minutes":8,"answer":"A model's knowledge fades over the year before its cut-off rather than stopping at a wall, and it has no clock and no reliable knowledge of its own cut-off.","keyIdea":"A model's knowledge fades over the year before its cut-off rather than stopping at a wall, and it has no clock and no reliable knowledge of its own cut-off, so time-sensitive answers need an explicit date and a dated source."},"body":{"format":"markdown","text":"## Two separate blind spots\n\n**It does not know what day it is.** There is no clock in a file of numbers. When an assistant tells you today's date correctly, the date was written into the hidden system prompt by the product. Strip that away (a raw model, a local install, an API call without a date) and it will guess, usually landing somewhere near the end of its training.\n\nThis produces small errors constantly and large ones occasionally. \"How long until the deadline\", \"is this certificate expired\", \"what is the current financial year\" all depend on a date it may not have. Give it the date explicitly whenever time matters.\n\n**It has a cut-off.** The training text was collected up to a point, and nothing after that point exists for it. Ask about an event from last month and, without a search tool, you get either a refusal or a confident answer about the world as it was.\n\n## The part nobody tells you: knowledge thins before the cut-off\n\nThis is the useful, non-obvious fact in this lesson.\n\nA stated cut-off of, say, early 2025 does not mean solid knowledge right up to early 2025. It means text was collected until then. But the internet writes about events slowly. An event in December produces a few news reports in December, analysis in January, encyclopedia updates over months, and the bulk of the eventual writing (the summaries, the retrospectives, the forum discussions, the textbook paragraphs) over the following years.\n\nSo the model's knowledge does not stop at a wall. It **fades over the final year or so**, from confident and well-supported, through thin and error-prone, to absent. The most dangerous region is that fade: recent enough that the model has something to say, thin enough that what it says is unreliable, and not obviously beyond the cut-off, so nobody thinks to check.\n\n> Figure: Knowledge fades before the cut-off; it does not stop at a wall. Line chart of Depth of written coverage, % of settled against Months before the stated cut-off: How much had been written by the time the text was collected (36, 100) (24, 97) (18, 91) (12, 74) (6, 45) (3, 22) (0, 8). The web writes about an event slowly: a few reports at the time, analysis over months, the bulk of the eventual writing over years. So the final year before a stated cut-off is thin, and it is the most dangerous region: recent enough that the model has something to say, not obviously beyond the date, so nobody checks.\n\nA practical rule: treat roughly the last twelve months before a stated cut-off as poorly supported, whatever the cut-off says.\n\n## It often misreports its own cut-off\n\nAsk a model when its training data ends and the answer is unreliable in both directions.\n\nThe model has no introspective access to this. Whatever it says is either something written into the system prompt by the product, or a guess assembled from what it has read about model releases. Models routinely state cut-offs that are wrong by months, and sometimes describe themselves as an earlier version entirely, because text about that earlier version was abundant in training.\n\nSo: get the cut-off from the provider's documentation, not from the model.\n\n## How to test the edge yourself\n\nRather than trusting a stated date, probe it. Ask about several events whose dates you know, spread over the eighteen months before the claimed cut-off, and watch where the answers become vague, hedged or wrong. That boundary is more informative than any published figure, and it takes five minutes.\n\nDo it once for the model you use most. You will probably find the reliable edge is earlier than advertised.\n\n## What search fixes and what it does not\n\nAn assistant with web access can answer about yesterday, and that closes the gap for current facts.\n\nTwo cautions carried over from the previous module. Some products search only when they judge it necessary, and that judgement is itself unreliable: a question about a changed rule may not look time-sensitive to the router. And retrieval brings back a page, which may itself be out of date; a top search result is frequently an article from three years ago that still ranks well.\n\nSo when currency matters, do two things: ask explicitly for a search, and check the date on the source rather than on the answer. A confident sentence with a link to an undated page is not a current fact.\n\n## Where this bites hardest\n\nAnything that changes on a schedule: tax rates and thresholds, benefit rules, visa requirements, exam syllabi, software APIs, drug guidance, interest rates, public holidays, who currently holds an office.\n\nFor all of these the model's answer is a snapshot of a past state, delivered in the present tense, with no marker distinguishing \"this is stable\" from \"this changed in April\". The absence of that marker is the whole problem, and no amount of careful prompting supplies it."},"check":{"question":"A model with a stated cut-off of early 2025 answers a question about a rule change from late 2024 confidently and incorrectly. What best explains this?","options":[{"text":"The provider's stated cut-off is inaccurate, and the true cut-off is earlier than advertised.","correct":false,"tempting":"Possible in some cases, but treats the fade as a labelling error rather than as the expected shape of coverage."},{"text":"Very little had been written about the change by the time the data was collected, so the model has thin, unreliable coverage while still being inside the stated window.","correct":true},{"text":"The model confused the change with a similar earlier one, which is a retrieval error rather than a coverage problem.","correct":false,"tempting":"Describes what the wrong answer looks like without explaining why coverage near the cut-off makes it likely."},{"text":"Rule changes are deliberately excluded from training data for legal reasons.","correct":false,"tempting":"Invents an exclusion policy; no such filtering of regulatory news exists."}]},"prev":{"slug":"counting-and-arithmetic","title":"Why it cannot count, and what to do instead","url":"https://addaly.com/learn/ai-actually-explained/counting-and-arithmetic"},"next":{"slug":"bias-in-the-output","title":"Bias, and the difference between fixing it and hiding it","url":"https://addaly.com/learn/ai-actually-explained/bias-in-the-output"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'It has no clock, and its knowledge thins before it stops', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/the-cut-off-and-the-clock (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"It has no clock, and its knowledge thins before it stops","courseTitle":"AI, Actually Explained","module":{"index":4,"slug":"why-it-goes-wrong","title":"Why it gets things wrong"},"answer":null,"keyIdea":"A model’s knowledge fades over the year before its cut-off rather than stopping at a wall, and it has no clock and no reliable knowledge of its own cut-off, so time-sensitive answers need an explicit date and a dated source.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"line","title":"Knowledge fades before the cut-off; it does not stop at a wall","series":[{"label":"How much had been written by the time the text was collected","points":[[36,100],[24,97],[18,91],[12,74],[6,45],[3,22],[0,8]]}],"xLabel":"Months before the stated cut-off","yLabel":"Depth of written coverage, % of settled","caption":"The web writes about an event slowly: a few reports at the time, analysis over months, the bulk of the eventual writing over years. So the final year before a stated cut-off is thin, and it is the most dangerous region: recent enough that the model has something to say, not obviously beyond the date, so nobody checks."},"alt":"Figure: Knowledge fades before the cut-off; it does not stop at a wall. Line chart of Depth of written coverage, % of settled against Months before the stated cut-off: How much had been written by the time the text was collected (36, 100) (24, 97) (18, 91) (12, 74) (6, 45) (3, 22) (0, 8). The web writes about an event slowly: a few reports at the time, analysis over months, the bulk of the eventual writing over years. So the final year before a stated cut-off is thin, and it is the most dangerous region: recent enough that the model has something to say, not obviously beyond the date, so nobody checks."}},"blocks":[{"t":"h2","id":"two-separate-blind-spots","runs":[{"text":"Two separate blind spots"}]},{"t":"p","runs":[{"text":"It does not know what day it is.","b":true},{"text":" There is no clock in a file of numbers. When an assistant tells you today’s date correctly, the date was written into the hidden system prompt by the product. Strip that away (a raw model, a local install, an API call without a date) and it will guess, usually landing somewhere near the end of its training."}]},{"t":"p","runs":[{"text":"This produces small errors constantly and large ones occasionally. “How long until the deadline”, “is this certificate expired”, “what is the current financial year” all depend on a date it may not have. Give it the date explicitly whenever time matters."}]},{"t":"p","runs":[{"text":"It has a cut-off.","b":true},{"text":" The training text was collected up to a point, and nothing after that point exists for it. Ask about an event from last month and, without a search tool, you get either a refusal or a confident answer about the world as it was."}]},{"t":"h2","id":"the-part-nobody-tells-you-knowledge-thins-before-the-cut-off","runs":[{"text":"The part nobody tells you: knowledge thins before the cut-off"}]},{"t":"p","runs":[{"text":"This is the useful, non-obvious fact in this lesson."}]},{"t":"p","runs":[{"text":"A stated cut-off of, say, early 2025 does not mean solid knowledge right up to early 2025. It means text was collected until then. But the internet writes about events slowly. An event in December produces a few news reports in December, analysis in January, encyclopedia updates over months, and the bulk of the eventual writing (the summaries, the retrospectives, the forum discussions, the textbook paragraphs) over the following years."}]},{"t":"p","runs":[{"text":"So the model’s knowledge does not stop at a wall. It "},{"text":"fades over the final year or so","b":true},{"text":", from confident and well-supported, through thin and error-prone, to absent. The most dangerous region is that fade: recent enough that the model has something to say, thin enough that what it says is unreliable, and not obviously beyond the cut-off, so nobody thinks to check."}]},{"t":"p","runs":[{"text":"A practical rule: treat roughly the last twelve months before a stated cut-off as poorly supported, whatever the cut-off says."}]},{"t":"h2","id":"it-often-misreports-its-own-cut-off","runs":[{"text":"It often misreports its own cut-off"}]},{"t":"p","runs":[{"text":"Ask a model when its training data ends and the answer is unreliable in both directions."}]},{"t":"p","runs":[{"text":"The model has no introspective access to this. Whatever it says is either something written into the system prompt by the product, or a guess assembled from what it has read about model releases. Models routinely state cut-offs that are wrong by months, and sometimes describe themselves as an earlier version entirely, because text about that earlier version was abundant in training."}]},{"t":"p","runs":[{"text":"So: get the cut-off from the provider’s documentation, not from the model."}]},{"t":"h2","id":"how-to-test-the-edge-yourself","runs":[{"text":"How to test the edge yourself"}]},{"t":"p","runs":[{"text":"Rather than trusting a stated date, probe it. Ask about several events whose dates you know, spread over the eighteen months before the claimed cut-off, and watch where the answers become vague, hedged or wrong. That boundary is more informative than any published figure, and it takes five minutes."}]},{"t":"p","runs":[{"text":"Do it once for the model you use most. You will probably find the reliable edge is earlier than advertised."}]},{"t":"h2","id":"what-search-fixes-and-what-it-does-not","runs":[{"text":"What search fixes and what it does not"}]},{"t":"p","runs":[{"text":"An assistant with web access can answer about yesterday, and that closes the gap for current facts."}]},{"t":"p","runs":[{"text":"Two cautions carried over from the previous module. Some products search only when they judge it necessary, and that judgement is itself unreliable: a question about a changed rule may not look time-sensitive to the router. And retrieval brings back a page, which may itself be out of date; a top search result is frequently an article from three years ago that still ranks well."}]},{"t":"p","runs":[{"text":"So when currency matters, do two things: ask explicitly for a search, and check the date on the source rather than on the answer. A confident sentence with a link to an undated page is not a current fact."}]},{"t":"h2","id":"where-this-bites-hardest","runs":[{"text":"Where this bites hardest"}]},{"t":"p","runs":[{"text":"Anything that changes on a schedule: tax rates and thresholds, benefit rules, visa requirements, exam syllabi, software APIs, drug guidance, interest rates, public holidays, who currently holds an office."}]},{"t":"p","runs":[{"text":"For all of these the model’s answer is a snapshot of a past state, delivered in the present tense, with no marker distinguishing “this is stable” from “this changed in April”. The absence of that marker is the whole problem, and no amount of careful prompting supplies it."}]}],"check":{"question":[{"text":"A model with a stated cut-off of early 2025 answers a question about a rule change from late 2024 confidently and incorrectly. What best explains this?"}],"options":[{"text":[{"text":"The provider’s stated cut-off is inaccurate, and the true cut-off is earlier than advertised."}],"correct":false,"tempting":[{"text":"Possible in some cases, but treats the fade as a labelling error rather than as the expected shape of coverage."}]},{"text":[{"text":"Very little had been written about the change by the time the data was collected, so the model has thin, unreliable coverage while still being inside the stated window."}],"correct":true,"tempting":null},{"text":[{"text":"The model confused the change with a similar earlier one, which is a retrieval error rather than a coverage problem."}],"correct":false,"tempting":[{"text":"Describes what the wrong answer looks like without explaining why coverage near the cut-off makes it likely."}]},{"text":[{"text":"Rule changes are deliberately excluded from training data for legal reasons."}],"correct":false,"tempting":[{"text":"Invents an exclusion policy; no such filtering of regulatory news exists."}]}],"order":[0,3,2,1],"why":null},"prev":{"slug":"counting-and-arithmetic","title":"Why it cannot count, and what to do instead","minutes":8},"next":{"slug":"bias-in-the-output","title":"Bias, and the difference between fixing it and hiding it","minutes":10}}}