A real problem that used to happen
In earlier versions, reading the same chat screenshot two or three times would occasionally produce titles with different phrasing — this time "Zhang San · Product Review," next time maybe "Product Review Meeting." The information wasn't wrong, just unstable. The root cause: the title field used to be handed entirely over to the AI model to generate freely, and the model's phrasing habits aren't perfectly consistent from one run to the next, even with the exact same input.
There was also a more specific kind of error: in a group chat, if someone says "let me introduce you two — Zhang San... Li Si..." — this "go-between making an introduction" scenario — the model would sometimes mistake the go-between's name for one of the actual parties to the event and put it in the title. But the go-between themselves isn't attending the meeting; the title should highlight the two people actually meeting, not the person who introduced them.
The fix: the model only produces the parts, and deterministic code does the assembly
The current approach splits "generating the title" into two halves. The AI model no longer produces a finished title sentence directly — it only outputs a handful of structured pieces of information, like "what name shows at the top of this message," "is this a one-on-one or group chat," "which names are mentioned in the body," and "roughly what is this message about."
Those pieces get handed to a fixed set of backend assembly rules — for one-on-one chats, the name shown at the top of the conversation takes priority; for group chats, the group name or email addresses are never assembled into the title; and a go-between whose only role in the text is "introducing two people" is excluded from being a title subject by the rules as well. No matter how many times the same image is read, the pieces fed into the assembly rules stay stable, and the rules themselves are fixed code, not something generated "depending on the model's mood" — so the resulting title naturally comes out stable too.
What this means for you
You don't need to care how any of this works under the hood — the only thing you'll notice is the result: for the same type of screenshot, the shape of the title is predictable. A one-on-one message title will include the other person's name; a group chat title will summarize the topic instead of randomly grabbing some name that appeared in the group. If you rely on titles to scan your calendar and recognize what an event is at a glance, that kind of consistency is more useful than "clever one time, different the next."
That said, this isn't a cure-all
The title assembly rules solve a "phrasing consistency" problem, not a "recognition accuracy" problem. If the screenshot itself is incomplete or the text is blurry enough that a name gets misread, the assembly rules won't magically correct it — which is also why chat screenshots always go to the results page instead of a zero-confirmation write (see the complete guide for details), leaving the final check in your hands.
