In plain English
- What happened
- Most chatbot vendors let a general product name like a model alias point to a newer version behind the scenes, so the same question can get a different answer without any notice in the product itself. OpenAI's model documentation explains the difference between a rolling alias and a fixed, dated snapshot, and its deprecations page lists retirement dates for specific model versions.
- Why it matters to your business
- A small business that built prompts or canned replies around a particular tone or format can find them quietly broken with no error message, because the model changed rather than anything the owner did. Whether the fix is a settings change or a rewritten prompt depends on which part moved, and that is checkable.
- Do this
- Open ChatGPT's settings, check whether Memory is turned on, and try the same question again with it off to see whether saved personalization, not a model update, is behind the change.
If an AI assistant is giving different answers than it did when a business first started using it, the tool has probably changed, not the business. Vendors regularly swap the model sitting behind a general name for a newer version, and separately, features that remember details about a user or their business can shift wording and emphasis in later replies. Both of these are documented by the vendors themselves, and both can be checked.
Why the same question gets a different answer
OpenAI's model documentation describes two ways to refer to a model: a rolling alias, like a general product name, that OpenAI can point to a newer version at any time, and a dated snapshot, which stays fixed until it is retired on a published date. A business asking questions through the ChatGPT interface is almost always on the rolling alias, which is why answers can shift with no announcement.
Separately, ChatGPT's memory feature lets the assistant carry details from earlier conversations into new ones, which can change the tone or specificity of a reply over months of use even if the underlying model has not changed at all. The two causes look identical from the outside but call for different checks.
OpenAI's deprecations page lists specific retirement dates for individual model snapshots, meaning a version a business relied on can stop responding entirely on a date set by the vendor, not by the business.
What the vendor pages actually show
| Reference type | What happens over time |
|---|---|
| Rolling alias (e.g. a general product name) | Points to whichever snapshot OpenAI currently serves; can change without notice |
| Dated snapshot (e.g. a specific version string) | Stays fixed; retired only on a date published on OpenAI's deprecations page |
Anthropic follows a similar pattern for Claude: its model deprecation policy states that specific model versions are retired on announced dates, and recommends that developers who need stability request a named version rather than a general one. Most small businesses interact through a consumer chat window rather than the API, so they inherit whichever version the vendor is currently serving without being asked.
What to check
- 1
Find out which model actually answered you
In ChatGPT, the model name is shown near the top of the chat window; on an API integration, check whether requests specify a named alias or a dated snapshot, per OpenAI's model documentation.
- 2
Turn off memory and re-ask the same question
In ChatGPT's settings under Personalization, memory can be switched off. Asking an identical question again with it off shows whether saved details, not a model change, shifted the answer.
- 3
Check the release notes against the date things changed
OpenAI publishes a running list of updates to ChatGPT. Comparing that list to when a prompt stopped behaving as expected narrows down whether a model update lines up with the change.
- 4
Pin a dated version where consistency matters more than novelty
For anything built on the API rather than the chat window, requesting a specific dated snapshot means updates only happen when chosen, not automatically.
- 5
Re-test saved prompts on a schedule
A prompt written for an older model version can behave differently once the version changes underneath it. Reviewing frequently used prompts every few months catches this before a customer sees the difference.
Reality check
A different wording in a reply is not proof the model got worse, and there is no dashboard that tells a business owner exactly which recent update touched their particular question. Chasing every change note is not a good use of time; checking memory settings and the model reference when an answer genuinely surprises usually is enough.
The underlying model, not the person typing the question, is usually what moved the second time an old prompt stops working the way it used to. The vendor's own model and memory pages are the fastest way to confirm which one changed, and this article reflects those pages as they read in September 2026.
Sources
