Knowledge Cutoff
The point where a model's training material stopped
- A knowledge cutoff is the point where the training material was gathered and frozen. Anything after that isn't inside the model.
- The cutoff exists because training is one huge push done all at once, not something that can be updated daily.
- The months right before the cutoff are especially thin — not enough has been written about them yet.
- If a model seems to know something after its cutoff, it's one of three things: search was turned on, you fed it in yourself, or it made it up.
- A model doesn't know today's date or even its own cutoff date on its own.
Contents
1The analogy
Flip through last year's paper planner all you like — March of this year never shows up. The pages only go as far as the day it was printed, and nothing exists past that. That's not the planner being lazy; it's just where things stood when it was made.
Only what was already fixed at print time made it onto the page, too. A holiday announced later isn't marked on any date. Plans for late December might have a lot of blank space, because they weren't finalized yet when the planner went to print.
There's still a way around it, though. Write a newly scheduled event on a sticky note and stick it on that date. It shows up as long as the note stays there, and disappears once you peel it off — the planner itself never changed.
2In detail
Why a cutoff exists at all
Building a model isn't a drip-feed process. An enormous amount of text gets gathered in one place and trained on all at once, over months. To start, collecting material has to stop on some specific day, and that day becomes the cutoff.
Work doesn't end there, either. Once training finishes, a stage of adjusting how it answers and checking for safety issues follows. That's why the day people actually start using it lands well after the cutoff — from a user's seat, you're already looking at a world several months old the moment it launches.
You might ask why not just retrain it. It's expensive and slow, and a freshly retrained model can end up worse in some ways than the one before it, so it doesn't happen often. The cutoff keeps slipping forward a little at a time — it never disappears.
The stretch right before the cutoff is thin
It's tempting to assume everything up to the cutoff is known equally well, but that's not how it works. Once something happens, writing about it keeps piling up for a long time afterward — recaps, analysis, arguments, summaries, arriving over months or years.
So something that happened a few weeks before the cutoff only made it in through a handful of short news items. Something from years earlier is layered in through thousands of pieces of writing. Within the same cutoff, the more recent something is, the shakier and shallower the answer tends to be.
That's why asking about something right up near the cutoff tends to get a wishy-washy answer. Not quite unknown, and not quite properly known either.
When it seems to know what happened afterward
If a model smoothly describes something that happened after its cutoff, it's one of three things. First, search might be turned on. Most services these days can look something up on the web when needed and answer using what they found — usually with a source link attached.
Second, you fed it in yourself. Paste in a document or upload a file, and its contents weren't inside the model, but they're readable within that conversation. It's the sticky-note case — open a new conversation and it's gone again.
Third, it made it up. Rather than saying it doesn't know, a model has a tendency to keep writing something plausible instead. If it names dates and people smoothly for something after the cutoff with no source attached, that's the one to suspect.
It doesn't even know its own cutoff for sure
Ask it when its cutoff is and it'll answer, but that answer is hard to trust. That answer, too, either came from what it read during training or from a note the people who built it wrote in — and it often lands earlier than the truth.
Today's date works the same way. There's no clock inside a model. It only seems to know because the service tacks today's date onto every request behind the scenes. Without that note, it defaults to whatever period shows up most often in its training material.
That's why a question with "now," "this year," or "the latest" in it is risky — which point in time the answer is anchored to gets fuzzy. Writing the year out explicitly is a lot safer.
How to handle a question that depends on the date
For anything that changes — prices, laws and policy, product specs, someone's current title — it's worth thinking about the cutoff first. Turning search on, or pasting in current material directly, is the safer move for these.
Once you get an answer, check whether a source is attached. No source usually means it's either inside-the-cutoff knowledge or invented. Even with a source attached, it's worth double-checking the date on that document — an old page can get presented as if it were current.
Plenty of questions have nothing to do with the cutoff, too. Editing a sentence, translating, summarizing, explaining code — anything where the method stays the same regardless of what's changed in the world barely feels the cutoff at all.
3More precisely
A knowledge cutoff is the point where gathering material for pretraining stopped. It isn't even a single clean date. Different data sources were collected at different times, so some bundles reach more recent than others, and some reach much further back. A little fresh material sometimes slips in during the later stage that adjusts how the model answers, so the boundary is closer to a band than a line.
The analogy breaks down in a place worth naming. A planner has no March page at all, so flipping past it makes the gap obvious immediately. A model has no visible blank spot. Sentences flow just as smoothly whether the ground under them exists or not, and that's the most dangerous part of dealing with a cutoff. Turning search on also isn't buying a new planner — it's sticking on a note. The knowledge inside the model stays exactly the same; what got found only lives inside that one conversation. One more thing: you can count the blank pages in a planner just by looking, but there's no way from the outside to tell what a model doesn't know.
4Try it yourself
5Common misconceptions
It's easy to think an AI watches the internet live and answers from that, but actually it only looks outward when a search feature is specifically turned on.
It's easy to think everything inside the cutoff is known equally well, but actually anything closer to the cutoff has less written about it yet, and comes out shakier.
It's easy to think asking about its cutoff gets you an accurate answer, but actually the model has no way to know its own cutoff for certain, and can state the wrong date.
7One-line summary
In shortA knowledge cutoff is the point where a model's training material stopped, and if it describes what happened after that smoothly, either search was turned on, you fed it the material, or it made it up.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02