The model knows an old version of you
Structure and markup refreshed for current answer engines; the original analysis is preserved.
This month OpenAI shipped newer models — function calling, a bigger context window, lower prices — with a knowledge cutoff still stuck in September 2021. The lesson hides in that gap: a newer model is not newer knowledge. What a model has memorized about your company can be nearly two years out of date, delivered with complete confidence, and you should be watching what it says.
the short answer
A model answers from a memory frozen at its training cutoff — right now, September 2021. So anything after that (a rebrand, a pivot, a new product, your current pricing) is not in there, and the model will describe the old you as confidently as the current you. This month’s newer models added capability but did not move the cutoff: a newer model is not newer knowledge. Live browsing helps but is early and sometimes off. You cannot edit the memory — so measure what it says, keep your record accurate, and expect corrections to land in the next snapshot, not this one.
key takeaways
- This month’s new models added function calling and a bigger context window — but kept a knowledge cutoff of September 2021. A newer model is not newer knowledge.
- What a model has memorized about your company can be years out of date: a rebrand, a pivot, a new product, current pricing — anything after the cutoff is simply not in its memory.
- It will describe the old you as confidently as the current you. The risk is not obvious errors; it is a dated snapshot delivered with full assurance.
- Live web browsing is the right long-term fix but, as of now, it is early and sometimes switched off — so you cannot assume an answer came from a fresh lookup rather than stale memory.
- You cannot edit the model’s memory, but you can shape it: monitor what it says, keep your current record accurate and well marked up, and treat corroboration as paying off in the next snapshot, not this one.
what the model knows, and when it stops
The shaded band is what the model can answer from memory; it ends at the cutoff. Everything in the gap after it — your rebrand, your newest product, your real pricing — is invisible to a memory-based answer, even from a model released this month. That gap is the subject of this note.
Why a confident answer can still be stale
The reason this slips past people is that staleness does not announce itself. A wrong fact you can catch; a dated fact delivered fluently you often cannot, because nothing in the answer flags its age. When a model tells a prospect what your company does, in clean and assured prose, the prospect has no way to know whether that description reflects you today or you at the cutoff two years ago — and neither does the model, which is simply reading from memory. If you have changed in any way that matters since September 2021, the assured description may be quietly describing a company that no longer exists, and the more polished the prose, the more credible the outdated version sounds.
It is worth being fair about what this is and is not. It is not the model malfunctioning; a frozen snapshot is how parametric memory works, and reporting it faithfully is the model doing its job. It is also not permanent: the next training run will move the cutoff forward, and live retrieval, where it works, can pull current information at query time. The point is narrower and practical: at any given moment, the memory-based picture a model holds of your company has a date on it, that date can be well in the past, and a capability update like this month’s does nothing to refresh it. Knowing that turns a hidden risk into a thing you can watch and manage.
The shape of it, in three parts
A model’s memory is frozen at a date in the past; capability moved this month while recency did not; so measure the snapshot and keep the record straight. Open each part for where it changes the work.
01 A model’s memory is frozen at a date in the past
The knowledge a model answers from, when it is not looking something up live, comes from its training, and training ends at a cutoff. Right now that cutoff is September 2021 — almost two years before this is being written. Everything a model "knows" from memory about your company is therefore a snapshot taken at or before that moment, and snapshots age. If you rebranded last year, launched a new flagship since, repositioned around a different problem, changed your pricing, or were acquired, none of that is in the frozen memory. Ask the model who you are and, answering from that memory, it will describe the company you were at the cutoff with the same fluency it would describe the company you are today. That fluency is the trap: there is no visible timestamp on a confident answer, nothing that says "as of September 2021", so a dated description reads exactly like a current one unless you already know better, and most prospects do not.
02 Capability moved this month; recency did not
It would be easy to assume that the newer models released this month know more recent things, and that assumption is exactly the one to drop. The update is real and useful — the models can now call functions and return structured output, handle far more text at once, and cost less — but the knowledge cutoff did not budge: the new versions know the world up to the same September 2021 as the ones before them. Capability and recency are separate dials, and only one of them turned. This matters because it kills a comforting intuition: that as the models visibly improve, their picture of your company is quietly refreshing too. It is not. A version bump can make a model a better reasoner and a better tool while leaving its memory of you exactly as stale as it was. The only things that actually advance the knowledge are a new training run with fresher data or a live lookup at query time — neither of which a capability release implies. The release notes will tell you the model is smarter; they will not tell you it has met the current version of you.
03 So measure the snapshot and keep the record straight
Because you cannot reach in and edit the model’s memory, the work is measurement and maintenance rather than a one-time fix. Start by reading the snapshot: ask the models, the way a customer would, who you are and what you offer, and record where the answer is current, where it is dated, and where it is wrong. Treat that as a baseline you re-check on a schedule, the same discipline you would apply to any metric that drifts. Then keep the record straight where you can reach it — your own clearly published, well-marked-up current information, and accurate reflections of the present you in the independent sources a model learns from — so that the next training run and any live lookup find the right material. And set the internal expectation honestly: this is a lag, not a light switch, so today’s corrections mostly land in the model’s next snapshot. That patience, paired with steady monitoring, is the AC Group’s habit applied to a new surface across 27 years: watch what is being said about you, and keep the underlying record true.
A rebrand the model never heard about
A company spent the past year on a serious repositioning: a new name for its flagship product, a sharper focus on a different buyer, updated pricing, and a fresh round of messaging across its site. Internally the change is complete and everyone has moved on. Then a prospect, doing early research, asks a chatbot what the company is known for — and gets the old story back: the previous product name, the previous positioning, a description that was accurate in 2021 and is now a year out of date. The prospect, with no reason to doubt a confident answer, forms an impression of a company that no longer exists, and carries that stale impression into the next conversation.
Nothing here is a malfunction. The model answered from a memory that predates the rebrand, faithfully, and the company’s own excellent new pages had no way to reach into that frozen snapshot. What the company can do is the work this note is about: notice the gap by asking the models what they say, make sure the current story is published clearly and corroborated by independent sources so the next training run and any live lookup find the new reality, and accept that the fix is a lag rather than a switch. The team that checks what the model says catches the stale story while it is just an inconvenience; the team that assumes the model keeps up learns about the gap from a prospect who already walked away with the wrong picture. Same rebrand, very different outcome, and the difference is whether anyone was measuring.
What to do with this
Make the snapshot visible, then keep the record true. Ask the models the questions a customer would ask about you — who you are, what you sell, how you are positioned, what it costs — and write down, plainly, where the answer is current, where it is dated, and where it is simply wrong. Re-check on a steady cadence, because the snapshot will shift as new training runs land and as retrieval improves, and you want to see those shifts rather than be surprised by them. This is just monitoring a metric that shifts, applied to the model’s memory of you.
Then work the two levers you actually have. Keep your own current information clearly published and well marked up, so a model that looks you up live finds the present you, and so the next training run ingests the right facts. Make sure independent sources reflect the company you are now, not the one you were, because a model trusts corroboration and corroboration is where stale pictures get corrected at the root. And hold the patient expectation internally: what you fix today mostly improves the model’s next snapshot, not the one answering right now, so the discipline is to keep the record accurate continuously rather than to chase an instant correction. That steady, unglamorous habit — measure what is said, keep the truth reachable, wait for the snapshot to catch up — is exactly how the AC Group has managed reputation through every change in how it travels for ' + years + ' years.
Which changes the gap hurts most
Not every difference between the old you and the current you carries the same cost, so it helps to know which ones bite. The most damaging are changes of identity and category: a rebrand or a name change, because the model may not connect the old name to the new one at all and simply answers about a company the prospect cannot find; and a repositioning into a different category, because the model files you under the old one and never surfaces you for the queries that now matter. Close behind are a new flagship the model has never heard of, so it describes your line-up as if your best current product does not exist, and a pricing or packaging change, where the model quotes figures that are not just old but wrong in a way that erodes trust the moment a prospect checks.
Mergers, acquisitions, and ownership changes belong on the same list, because they alter the very entity the model is trying to describe, and a confident answer about the pre-merger company can mislead. The common thread is that the costliest gaps are the ones that change who you are or where you sit, not the ones that merely add detail. A model that is a year behind on a minor feature is a small problem; a model that is a year behind on your name, your category, your flagship, or your price is describing a different company to the people deciding whether to consider you. Knowing which of your own changes fall in that heavier bucket tells you where to look first when you read the snapshot.
When a stale snapshot doesn’t matter
It would be dishonest to imply this is a crisis for everyone, because for many companies the gap is harmless. If you have not changed in any consequential way since the cutoff — same name, same category, same core offer, steady positioning — then a memory frozen two years ago describes you about as well as a fresh one would, and the staleness is a non-issue. Stability is its own protection here: the less you have moved, the less there is for the model to be wrong about. Plenty of solid businesses are in exactly this position, and the right response for them is to confirm it with a quick check and then spend their attention elsewhere rather than manufacturing an alarm.
The point of this note is not that every model answer is dangerously old; it is that you cannot know whether yours is without looking, and that the act of looking is cheap relative to the cost of being surprised. So the honest framing is conditional: if you have changed in something that matters, the gap is working against you right now and quietly; if you have not, you are probably fine and can prove it in minutes. Either way the move is the same — check what the snapshot says — and the only companies this genuinely catches out are the ones that assumed, without checking, that the model had kept up with a change it had no way to know about.
The stale-snapshot problem: quick answers
Why would a model know an outdated version of my company?
Because the knowledge a model carries in its weights was frozen at a training cutoff, and that cutoff is often well in the past. The models OpenAI shipped this month are newer and more capable — they can call functions and read more text at once — but their world knowledge still stops in September 2021, nearly two years ago. Anything that happened to your company after that point simply is not in there: a rebrand, a pivot, a new flagship product, a change of positioning, a merger, even your current pricing. When someone asks the model about you and it answers from memory rather than from a live lookup, it answers as the company you were at the cutoff, not the company you are now. It is not lying and it is not broken; it is faithfully reporting a snapshot that has aged. The danger is that it sounds just as confident describing the old you as it would the current one.
Doesn’t a newer model mean newer information?
No, and this month is a clean illustration of why. The update that landed added real capabilities — function calling, a larger context window, lower prices — without moving the knowledge cutoff at all; the new versions know the world up to the same September 2021 as the ones they replaced. Capability and recency are two different dials. A model can get smarter at reasoning, better at following instructions, and more useful as a tool while still being frozen at the same moment in time for what it actually knows about the world. So "they released a new model" tells you nothing about whether it has caught up on your company. The only things that move the knowledge forward are a fresh training run that includes more recent data, or a live retrieval step that lets the model look something up at the moment you ask — and neither of those is implied by a version bump.
Can’t live web browsing fix the staleness?
It can help, but as of this writing it is not a reliable backstop. Retrieval — letting the model look something up live rather than answer from memory — is the right long-term answer to staleness, because it pulls from the current web instead of a frozen snapshot. But today the browsing features are early, uneven, and sometimes switched off entirely, so you cannot assume a given answer about your company came from a live lookup rather than from stale memory. When retrieval does fire, recency depends on what it finds and how fresh that is; when it does not, you are back to the cutoff. The practical stance is to treat live retrieval as a partial, improving mitigation rather than a solved problem, and to keep watching what the model says about you in both modes, because the memory-based answer is the one most likely to be quietly out of date.
What can I actually do about a stale snapshot?
Three things, in order. First, find out what the snapshot says: ask the models the questions a customer would, about who you are, what you offer, and how you are positioned, and write down where the answer is current, where it is dated, and where it is simply wrong. Second, fix the record where it is reachable — make sure your own current information is clearly published and well marked up, and that independent sources reflect the present you, so that the next training run and any live lookup have accurate material to draw on. Third, set the expectation internally that this is a lag, not a switch: the corroboration you build today mostly pays off in the model’s next snapshot, not this one, so the work is to keep the record accurate continuously rather than to expect an instant correction. You cannot edit the model’s memory directly, but you can shape what it learns next and what it retrieves now.
A note on sources and timing
This is written in June 2023, just after OpenAI released updated models with function calling and a larger context window while leaving the knowledge cutoff at September 2021 — capability forward, recency unchanged. We have described only what was public as of this writing, and we have been careful not to overclaim: live retrieval, where it works, does reduce staleness, and the cutoff moves forward with future training. The durable point does not depend on any specific date: at any moment, the memory a model answers from has an age, that age can be well in the past, and a capability update does not refresh it — so what a model "knows" about your company is something to monitor, not assume. Keeping the record true and watching what gets said is the standard the AC Group has held for 27 years, now pointed at a surface whose memory of you is older than it sounds.