How to Tell Users You Changed the AI Model

Swapping the model behind your product is usually invisible and occasionally load-bearing. Announcing every change trains users to ignore your changelog. Announcing none of them means the person whose workflow just broke finds out by having it break.

The test is not how big the change is technically. It is whether a user built a habit around behaviour that is now different.

When to say nothing

No announcement needed when the swap is genuinely behind the curtain:

• Output quality improved and nothing about the shape, length, or format changed

• You moved between snapshots of the same model family with no behaviour delta you can detect

• The change is a cost or latency optimisation invisible in the output

• You route across several models already and this adjusts the routing

The common thread: nobody has to do anything, and nobody will notice unless they are told. Announcing these has a real cost. Every non-event in a changelog lowers the odds anyone reads the entry that matters.

When to say something

Announce when a user could reasonably have built on the old behaviour.

• Output format or length changed. Anyone parsing, copying into a template, or having trained their expectations on the old shape is affected.

• Refusal boundaries moved. Requests that used to work now decline, or the reverse. This generates support tickets faster than anything else on the list.

• Tone or style shifted noticeably. Especially where users have shared the output with their own customers.

• Capability was removed. A modality, a language, a context length they were using.

• Price or limits changed for them. Always announce, always in advance.

• Regulated or contractual context. If a customer's compliance documentation names the model, the change is a contractual event, not a product note.

If you are unsure, the practical proxy is your own eval suite. A model swap that moves your eval scores meaningfully in any direction, including up, changed something users can see.

What to actually write

Users have three questions, and generic release notes answer none of them. What changed for me, do I need to do anything, and how do I get the old behaviour back.

## Improved responses in Draft Assistant

We upgraded the model behind Draft Assistant on 14 August.

What changed: drafts are noticeably more concise. If you relied on
the longer format, add "write in full paragraphs" to your template.

What to do: nothing, unless you have saved templates that expect the
old length. Existing drafts are unchanged.

Rolling back: the previous behaviour is available under
Settings > Assistant > Style > Detailed until 30 September.

Four things make this work. It leads with the user-visible effect rather than the model name. It names a date. It gives the person whose workflow broke a specific fix in one line. And it offers an exit with an end date, which is the part most teams skip and the part that turns an angry ticket into a shrug.

You do not have to name the underlying model, and many products deliberately do not, so they stay free to switch. If you have named it publicly before, keep naming it. Going quiet after being specific reads as concealment even when it is not.

Timing

Advance notice for anything that costs a user work. Same-day is fine for improvements. Never announce a degradation after the support tickets arrive, because at that point you are not informing anyone, you are responding.

A workable default: quality improvements go out the day of the change, format or behaviour changes get a week, capability removals and pricing get thirty days. If your change is forced by a provider retiring a model, you inherit their timeline, which is usually months. Use it, and see what to do when an AI model gets deprecated for the migration itself.

The disclosure floor

Separate from product communication, there is now a baseline obligation in some jurisdictions. Under the EU AI Act transparency rules, which the Commission started enforcing on 2 August 2026, interactive AI systems must tell users they are dealing with AI rather than a person, and certain generated content has to be labelled.

That is about disclosing that AI is involved at all, not about announcing which model. It does not require a changelog entry for a swap. But if your disclosure text names a specific provider or model, that text is now something you have to keep accurate when you change it, which is a good reason to keep it generic.

Where the entry lives

A changelog users can actually find, ideally with a feed. If your product does not have one, adding a changelog page is a small job that pays off well beyond model changes.

Match the channel to the impact. Changelog for everything worth announcing. In-app notice for changes that affect a workflow, shown to the users who actually use that feature rather than everyone. Email for pricing, limits, and removals only, because email spend on non-events is how you get unsubscribes before the message that matters.

Do the swap in an order that makes the note writable

The communication is easier when the migration was disciplined, because you know what changed rather than guessing. Run the new model against your evals before switching, keep the old one live behind a flag, and switch a small percentage of traffic first. Then your changelog entry describes observed differences instead of speculation, and your rollback line is true because you have tested it.

The mechanics are in how to migrate from one AI model to another, and the decision of whether to move at all is in how to know when to upgrade to a newer AI model.

Frequently asked questions

Do I have to tell users which model I use?

Generally no, and staying generic keeps you free to switch. Enterprise buyers in regulated sectors will ask, and some contracts require it, in which case the model name becomes a term you have to manage rather than a detail.

What if the new model is worse for some users?

Say so, and keep the old path available with a stated end date. A published limitation costs far less than the same limitation discovered by a customer who was told everything improved.

Should I announce that I use AI at all?

If users interact with it directly, yes, and in some jurisdictions you must. Being upfront also avoids the much worse conversation where a customer discovers it themselves.

How do I know whether a model change is user-visible?

Run your evals, then read twenty real outputs from before and after side by side. Twenty is usually enough for a difference in tone or length to become obvious, and it catches things a score does not.