What happens to your messages

You type something you would not say to most people, and it goes to a server. What happens next is determined entirely by documents and decisions on the other end, and it is the one part of these products where you can get real answers — because unlike the model or the persona, this part is written down somewhere by legal necessity.

Most people never read those documents. Which is fair, because they are long and dull. This post is about the six things in them that actually matter here, and the wording that tells you the answer.

Your messages are stored, essentially always

Start here, because it removes a common assumption. A companion app has to store your conversation to function at all — the history is what gets reassembled into every reply. There is no version of this product that keeps nothing.

So the question is never whether messages are stored. It is: for how long, who can see them, what they are used for, and what happens to them when things end.

The six questions, and where the answers live

How long are messages retained? Look in the privacy policy for a section headed retention, data retention, or how long we keep your information. What you want is a period — a number of months, or “for as long as your account is active, and X after deletion”. Vague retention language (“as long as necessary for our legitimate business purposes”) is an answer too: it means the period is whatever they decide.

Can a human read them? Look for human review, manual review, moderation, or trust and safety. Almost every service reserves the ability to have staff review content, usually for moderation or abuse investigation, and often for quality improvement. This is normal and it is also the item that surprises people most, so it is worth knowing rather than assuming.

Are they used to train models? This is the one to read carefully. The wording to look for is improve our services, train, model development, or machine learning. Note two things: “we do not sell your data” is a different claim from “we do not train on your data”, and a policy that is entirely silent on training use has not promised you anything. Where an opt-out exists it is usually in settings under something like data controls or privacy, and it is often on by default.

Are they shared with anyone else? Look for third parties, processors, service providers, subprocessors. Most apps do not run their own model, which means your messages are transmitted to whoever supplies it — a genuine third party with its own retention and its own policy. That is unremarkable architecture and it does mean there is more than one company holding the text.

What does deleting actually delete? Look for deletion, your rights, account closure. Distinguish deleting a conversation in the interface from deleting your account, and both from what is retained afterwards — backups, moderation records, and legally required records commonly persist for a stated period after a deletion request. A policy that gives a number here is being straight with you.

Can you export? Look for data portability or export. Where a real export exists you can keep your own copy of a conversation, which matters more than it seems, because everything else in this list is outside your control and that is the one thing inside it.

The product

THE PRODUCT — your messages

  · Stored at all
                    → yes, necessarily. The history is what
                      gets reassembled into every reply.

  · "We don't sell your data"
                    → a narrower claim than "we don't train
                      on it". Both sentences can be true and
                      only one is the question.

  · A policy silent on training use
                    → has promised nothing. Silence is an
                      answer.

  · Deleting a conversation
                    → not the same as deleting the account,
                      and neither is the same as it being
                      gone from backups.

  · Retention, review, training use, deletion
    windows
                    → THE OPERATOR DECIDES, and revises. The
                      policy states today's version.

  · The exact answers for any given app
                    → CHECK THE POLICY. Search it for
                      "retention", "train", "human review",
                      "third parties", "delete", "export".

The two ends nobody plans for

Shutdown. Companion apps close. When one does, the conversations go with it, usually with short notice and sometimes with none. There is rarely an export offered on the way out, because a company winding down is not building features. If a history matters to you, an export taken while the service is healthy is the only version of that you control.

Acquisition. Privacy policies almost universally include a clause permitting transfer of personal data in a merger, acquisition or sale of assets. Worth reading once, because it is the mechanism by which the company holding your messages becomes a different company with a different policy, without you doing anything. Look for change of control or business transfer.

Neither of these is sinister and both are routine. They are just the parts of the arrangement that the day-to-day experience gives you no reason to think about.

What you can actually do

Four things, all modest.

Read the retention and training sections once. Ten minutes, two searches in a document. It is the highest-information thing available on this entire subject.

Check settings for data controls. Where a training opt-out or a chat-history control exists, that is where it is, and it is usually not surfaced during onboarding.

Take an export if one exists. Before you want it, not after — the moment you most want it is usually the moment the service is going away.

Assume a human could read any given message. Not because one will, but because that assumption is accurate and it is the only privacy model that survives contact with how these services are actually built.

What this doesn’t tell you

It does not tell you what any specific app does with your messages. That is in their documents, it differs across the category, it changes, and any claim made here would be both unverifiable and out of date.

It does not tell you whether a policy is honoured, which is a question about enforcement rather than wording.

And it does not tell you what to do about a conversation you regret having sent, because there is often nothing to do — which is the actual reason to know all of this in advance rather than afterwards. Same principle as the pricing: the arrangement is much easier to evaluate before you are invested in it.