Where the suggested replies come from
Under the reply there are two or three things you could say next, written in your voice, ready to tap. Sometimes one of them is almost exactly what you were about to type. That is not a coincidence and it is not the character anticipating you. It is another feature, produced by a separate step, and the interesting part is what happens to the conversation when you use it.
They are either generated or scripted, and you can tell which
Generated. The app makes a second request, alongside or after the reply, asking a model for plausible next messages from you given the conversation so far. Sometimes it is the same model that wrote the reply; often it is a smaller and cheaper one, because the suggestions are short and nobody minds if they are less good. Generated suggestions reference specifics: a name you mentioned, a plan from two turns ago, the thing you just said.
Scripted. A fixed list, written once by the operator, sometimes selected by a simple rule — a new conversation, a long gap, a topic tag. Scripted suggestions are generic by necessity. They work in any conversation, which is exactly what makes them recognisable.
The test takes one look. If a suggestion could be pasted into any conversation with any character, it was written in advance. If it could only make sense here, something read the conversation to produce it. Many products do both, scripting the opening of a conversation and generating from then on.
They do three jobs at once
They remove the blank page. A conversation stalls when there is nothing obvious to say, and a stalled conversation ends. Suggestions convert a pause into a tap. Worth noting: a tapped suggestion is a message, and a message is a turn, so it consumes whatever the product counts — the unit the meter is measuring is the action, not the effort of typing it.
They steer. A suggestion is a proposal about where the conversation goes next, and the operator chooses which proposals exist. That can point away from areas where a moderation layer would intervene and toward things the product does well, which makes the product look more capable than an unguided conversation would. It can also point at a feature you have not used, or one that sits behind a tier, which is a way of arriving at a price screen at a moment that was chosen rather than stumbled into.
They produce clean signal. Free text is expensive to analyse. A choice among three fixed options is a labelled event: this person was shown A, B and C and picked B. Which suggestion gets tapped, and how often suggestions get tapped at all, are the kind of measurements that sit naturally in the telemetry an app collects alongside the conversation — and the kind of thing that gets varied between groups of users to see which set performs better, since you may be in an experiment without a notice.
What tapping one does to the conversation
This is the consequence that outlasts the tap. The suggestion becomes your message. It enters the history verbatim, on your side, and every later reply is built from that history.
So a share of what the character appears to know about you, and a share of the tone it mirrors back, is text the app wrote and you approved with one tap. The character is agreeable to whatever is in the conversation, and some of what is in the conversation was proposed by the product. Selecting a reply does not only choose what you say now; it chooses the record that everything after it is generated from.
That is not a hidden mechanism. It is the same effect as keeping one regenerated reply rather than another, and it is the reason what you keep in the history matters more than which single reply you liked.
THE PRODUCT — suggested replies
· Options in your own voice
→ generated by a second request, or a
fixed list written in advance.
· One that references a detail
→ generated. Something read the
conversation to write it.
· One that would fit any conversation
→ scripted, and reused across everybody.
· A suggestion that leads to a locked feature
→ a route, and the destination was
chosen before you tapped.
· Which options you are shown
→ THE OPERATOR DECIDES, and can test two
sets against each other without saying
so.
· Whether your choices are recorded
→ CHECK THE POLICY. Look for "usage
data", "interactions", "analytics".
· Whether they exist at all
→ VARIES BY APP, and can appear or
vanish in a release.What you can check
Watch for specificity. Over a few days, note whether any suggestion has ever referenced something only this conversation contains. If none has, you are looking at a fixed list, and the apparent responsiveness is your own pattern-matching rather than the product’s.
Notice where they lead. If suggestions repeatedly arrive at a feature you do not have, that is a route with a destination. Nothing is wrong with a product advertising itself; it is worth seeing it as advertising rather than as conversation.
Type your own for a week. The comparison is the point. An unguided conversation shows you what the character does without the product’s steering, and the difference between the two is the size of the steering.
Search the privacy policy for interaction data. Whether a tap is stored, and for how long, sits in the same category as the rest of what leaves your device alongside your messages — and a policy that lists interactions and analytics is telling you that taps are events.
What this doesn’t tell you
It does not tell you whether any particular app generates or scripts its suggestions. The specificity test is a strong hint from outside and it is not proof.
It does not tell you whether your choices are analysed, or by whom. That is answered by a document, and only in the general terms that documents use.
And it does not tell you how a suggestion feature would be built or tuned. That is a builder’s subject and no part of it appears here.