LIVE-12: AI confidence level feedback and control

Last updated: November 5, 2026

Part of the EchoVision Wish List, under Live AI.

• Status: Under review.
• What owners asked: Live AI sometimes gives wrong answers when it presumably has trouble deciphering what it is seeing. Would it be possible to have Live AI express its level of confidence in the accuracy of its response when that level is below a certain threshold?
• Related: SCENE-06: Read clocks, dials and screens more reliably and LIVE-11: Read signs, labels and displays aloud
• Discussed in: Live AI Hallucinations, on the pioneers list (members only)

The need

Live AI usually answers well, but now and then it answers wrongly with full confidence, most often when it can't see something clearly. Owners gave two examples: a bottle of cognac called a Kentucky bourbon, and a dishwasher where Live AI missed the highlighted setting and listed two options that model does not have. An owner who already knows the answer can catch the mistake. An owner who doesn't know may be misled, and for things like medication or appliance settings, that matters.

Today's workaround

• Ask again, or ask Live AI to look again, and compare the answers.
• Challenge the answer and say what seems wrong; Live AI will often correct itself.
• Treat every answer with some caution, and confirm anything important, such as medication, another way.
• Report a wrong answer by pressing the AI button five times, so the team can learn from it. See Support in the user guide.

What it would look like

• When Live AI is not sure, it says so, for example "I'm not certain, but this looks like a bourbon" or "I can't read the highlighted setting clearly."
• When it is sure, it answers as it does today, with no extra words.
• It may suggest how to get a better look, such as moving closer, adding light or turning the item.
• Possibly a setting for how readily it warns, from "only when very unsure" to "whenever there is some doubt", or off.
• Possibly the same behaviour in Scene Description and questions, if owners want it there.

Pros and cons

• Pro: owners know when to double-check, instead of trusting or doubting every answer equally.
• Pro: fewer misleading answers on labels, displays and controls.
• Pro: builds trust, because owners learn when Live AI is reliable.
• Con: AI models are often confident when wrong, so the warning itself may not always be reliable.
• Con: too many warnings could make answers longer and more tiring to listen to.
• Con: a confidence warning could give a false sense of safety for medication, where no AI reading should be relied on alone.

Open questions

• Can the AI services behind EchoVision give a dependable confidence estimate?
• Should this be in Live AI only, or also in Scene Description and questions?
• Should owners control the threshold with a setting, and if so, where?
• How should EchoVision handle prescription bottles and other medication, where a misread could cause harm?

History

• November 6, 2026: Under review. From a discussion on the pioneers list.