The prompt box is disappearing into everything

Meta’s reported prototype has a blunt promise: AI glasses that can capture your entire day so you can later ask what you saw or heard. That is either a brilliant personal memory layer or the fastest way to make everyone around you wonder whether they have become an entry in somebody else’s AI archive.

The obvious reading is that AI products are getting more multimodal. Voice here, video there, cameras on your face. True, but too shallow. The bigger shift is that the prompt box is dissolving. AI is moving from a place you visit into a layer that watches, listens, edits and remembers.

That changes the product problem completely.

The interface is now the world

OpenAI’s GPT‑Live release points in the gentler direction: live interaction inside an existing assistant. This is still recognisably a user-controlled exchange. You talk, it responds. The prompt box has become more conversational, but the boundary is visible.

Google Photos is further along the same path. Its new AI Video Remix tool puts generative editing inside a photo library, making AI feel less like separate creative software and more like a normal thing you do to existing clips.

Meta’s reported glasses prototype takes the leap. The Decoder reported on always-on AI glasses that capture your day. That is a product manager’s dream: zero-friction memory and contextual assistance. It is also the point where “Did the user consent?” becomes a weak question, because the product involves everyone near the user.

This is the part builders should sit with. Ambient AI does not merely need better models. It needs better social contracts.

Typing a prompt is an act of intent. Speaking to an assistant is still fairly legible. Letting software work across your camera roll, your voice conversations and your field of view turns intent into something inferred. The product no longer asks, “What do you want me to do?” It guesses from what it can see.

That is powerful. It is also how trust gets spent.

Provenance becomes product design

Google’s deepfake detector story is the counterweight. TechCrunch reported that Google’s system was used to debunk a McConnell hoax image. Good. A practical win for provenance at the exact moment synthetic media is becoming mundane.

But the lesson is not “detection will save us”. Provenance systems work best when platforms, model providers and distribution channels participate. That is a coordination problem, not a UX flourish.

There is an old retail parallel here. Supermarkets did not scale because shoppers personally trusted every producer. They scaled because packaging, labels, expiry dates and receipts made trust portable. The boring infrastructure made the abundance usable.

Ambient AI needs its version of that. If Google Photos can remix a clip, the edited media should carry intelligible provenance. If GPT‑Live makes AI interaction feel more natural, users need clear controls over what is heard, seen and retained. If Meta builds glasses that remember the day, consent cannot be buried in a settings menu. It has to be part of the interaction model, visible to the wearer and legible to bystanders.

The companies that treat this as compliance will ship creepy magic. The ones that treat it as core product design may earn the right to make AI ambient.

The prompt box is disappearing. The receipt for what AI saw, changed and remembered now has to appear somewhere else.

---