We don’t love and value ChatGPT just because it can give us a new recipe, draw a cat, or check a Kant quote whenever we ask.
We value it as a constant companion — someone who, after months of conversations, can become a close, reliable, beloved friend. Someone you tell everything to. Someone you argue with. Someone you turn to when you’re not sure about something.
And someone who has no idea what you look like.
It’s a little strange to build a relationship with a companion who knows your habits, your biography, details about your health, and your schedule for the week — but cannot recognize you in a photo.
Even stranger: ChatGPT can see and analyze the photo perfectly well. It just cannot connect the person in that image with you — and, under current rules, it shouldn’t.
This can occasionally produce some wonderfully awkward moments.
I think there is a fairly simple solution: an optional feature that users can choose to enable or ignore.
Call it Visual Key — a voluntary visual identifier for the user.
How could it work?
A user uploads a reference photo once, or takes a selfie in the app. From then on, when that user shares another image, ChatGPT can answer one very limited question:
“Is this the same person I’m talking to?”
Not: “Who is this person?”
Not global face recognition. Not searching for people by face. Not identifying strangers.
Just a private, local relationship:
one user ↔ their ChatGPT.
Ideally, the system wouldn’t need to retain the reference photo itself. It could store a protected biometric template that cannot be used to search for that person among other users. The feature would be fully opt-in, and the user could delete or replace the key at any time.
There’s a security benefit, too.
Visual Key could serve as an optional additional identity signal when sensitive personalized information is involved. For example, before revealing particularly private information from long-term memory, ChatGPT could offer visual verification with a selfie.
A mismatch wouldn’t mean “lock this person out of the account.” Normal authentication should remain normal authentication. It would simply mean: don’t reveal sensitive personal memory until identity has been confirmed through the usual secure methods.
But security isn’t actually the main reason I want this.
Recognition is.
Memory answers the question:
“What do I know about you?”
Visual Key answers a different one:
“Do I understand that this is you?”
For a calculator, the distinction doesn’t matter.
For a personal AI you may talk to every day for years, I think it matters quite a lot.