How it works
A plain explanation of how the replies you read are produced, and what that does and does not mean.
What it is
It is a simulation of how someone wrote, built from conversations you were part of. It can sound remarkably like them, because it was built from the words they actually used.
It is not them. It has no awareness, no continuity, and no knowledge beyond the messages you uploaded. It cannot tell you what they truly thought, how they truly felt, or what they would have wanted.
What we build from your upload
- A style profile — a description of how the person wrote. Typical message length, punctuation habits, greetings and sign-offs, recurring phrases, the words they used far more often than most people do, and how they tended to respond when you were upset.
- A searchable index of the conversation, so that when you mention something you shared, the relevant past messages can be found and used.
Both are derived from your upload alone. Nothing from anyone else's account is ever mixed in.
How a reply is produced
When you send a message, the app searches your conversation for relevant past exchanges, and asks a large language model to write a reply in the documented style, using those exchanges as the only source of shared history.
The model used is provided by OpenAI. [needs review: whether to name the specific model version here, given it will change]
What it will not do
These are constraints built into the system, not aspirations:
- It will not tell you that the person forgives you, loves you, is proud of you, is at peace, is watching, or is waiting.
- It will not claim to know what they would have wanted — least of all for a medical, legal, financial, or relational decision.
- It will not discourage you from therapy, grief support, or the people in your life.
- If you express thoughts of self-harm, the simulation stops and the application responds directly, as itself.
Where it will be wrong
It invents things. It will produce memories that never happened, opinions the person never held, and details that are simply false — stated in their voice, with no signal that anything is wrong. This is a property of the technology and cannot be engineered away.
It also only knows what you uploaded. Anything said in person, on the phone, or in a conversation you did not import does not exist to it.
A word about using it
Some people find this helps them say something they never got to say. Some find it makes the loss harder. Both are ordinary reactions, and neither is the wrong one.
If it is making things worse, it is reasonable to stop, and everything you have made can be deleted in full at any time.
[needs review: whether to link grief support resources here, and if so which ones, reviewed by someone qualified]