Google began rolling out Guided Vision in Gemini Live to compatible Android devices on 1 October, a feature that turns the phone camera into a source of spoken description for people who are blind or have low vision.
The mechanism is a shared camera. During a Gemini Live session with Guided Vision enabled, Gemini describes what the camera is pointed at, answers follow-up questions, and, when the framing is wrong, gives spoken cues to correct it, asking the user to pan slowly right, tilt down or step back so it has the visual context to answer. Google’s announcement groups the uses in three ways: reading fine print such as nutrition labels, appliance dials, washing machine cycles or printed menus in dim restaurants; finding and localising objects in the immediate area, such as a dropped earbud or a jar in a crowded spice cupboard; and describing details such as colour, pattern and shape.
Google says the feature was built with the accessibility community rather than for it. It names Aira, an established remote visual interpretation service, as a partner: tens of thousands of hours of visual interpretation data went into training Gemini Live for conversational real-world context, and more than 1,000 members of Aira’s Trusted Tester network stress-tested the model while Aira specialists worked as subject-matter experts in establishing and evaluating the safety guardrails. That detail about how an accessibility feature was validated is unusual, and it is the part of the announcement most coverage omits.
The Verge adds the distribution detail. Guided Vision is available in the Gemini app and in Google TalkBack, and can be given a dedicated accessibility shortcut in Settings on devices running Android 9 or later. It arrived alongside a recent set of Pixel updates, and it parallels VoiceOver Live Recognition, which Apple has added to the iPhone and the Vision Pro. Google previewed the feature before today’s launch, describing it at the time as one of several tools for expanding digital and physical accessibility across Android; the preview was reported in early September.
The caveat matters more than the feature list. Google cautions users not to rely on Guided Vision for navigation, safe-travel guidance or obstacle detection, and not as a replacement for a cane or other mobility aid. That is an explicit limit on what an AI description of a scene can be trusted to do while a person is moving through the world, and it places the feature in the category of assistance rather than wayfinding.
Three things the announcement does not answer: whether an iOS version is planned, how the feature holds up in poor light or with a moving subject, and what happens to the camera frames the model processes. Google’s post discusses training data for the model but says nothing about retention or training use for the images captured in a session. Anyone weighing it for daily use should test those limits against their own conditions rather than the examples in the announcement.