Gemini Live's new Guided Vision describes your surroundings. It is not a cane replacement.

What Google just shipped

On 1 October 2026, Google began rolling out Guided Vision inside Gemini Live, a feature built for people who are blind or have low vision. A user shares their phone camera during a Gemini Live session, and Gemini narrates what it sees in real time. It works on Android 9 and newer, in any market where Gemini Live is already supported, and needs an updated Gemini app and current Android system software. Google names no price or subscription requirement anywhere in the launch post, so it appears to ride on whatever Gemini Live access a user already has.

Google's own framing for why it built this is blunt: "technology is most powerful when it opens up the world for everyone." That line sits inside a wider push the company describes as expanding accessibility "across Android," and Guided Vision is the most concrete piece of that push shipped so far this year.

Four jobs it is actually built for

Google's launch post is specific about what the camera gets pointed at, which matters for a reality check: a vague "it helps blind people see" claim is marketing, a list of named tasks is a product. The post describes four categories. Reading fine print and complex text: nutrition labels, appliance dials, washing machine cycle settings, printed menus. Finding and localizing objects: a dropped earbud, something buried in a cabinet. Describing and matching objects: checking whether two clothing items pair together, identifying a jacket's color. Exploring an immediate environment: a room's layout, or what is sitting on a table. None of those four is navigation, and that gap is deliberate, not an oversight.

How the guiding actually works

The core trick is that Gemini does not just describe, it coaches. If the camera is pointed too high, too close, or angled off to one side, Gemini gives spoken prompts to reframe: pan right, tilt down, step back. Google's own announcement puts it directly: "Gemini acts as a conversational partner" that offers "natural verbal prompts to help you reframe" the shot. That back-and-forth is the difference between a one-shot image caption and something closer to a conversational AI agent walking a person through a task.

Google says the underlying model was trained on "tens of thousands of hours" of visual interpretation data developed with Aira, a company that has run human-staffed visual interpreting services for blind and low-vision customers for years. Aira's own specialists acted as subject-matter experts while Google and Aira worked out safety guidelines for what the model should and should not attempt to say.

Where Google draws the line

The most telling part of the launch post is not a capability claim, it is a limitation. Google states plainly: "As with any generative AI technology, Guided Vision is an assistive utility and can make mistakes. It is not a medical device, mobility aid, or white cane replacement, and it's not intended for navigation, safe-travel guidance, or obstacle detection."

That is a meaningful line to draw, because those exact tasks (crossing a street safely, detecting a curb or a step, deciding whether a path is clear) are the highest-stakes problems a blind or low-vision person actually faces outdoors. Guided Vision is scoped to indoor, low-stakes, high-frequency tasks: reading mail, matching colors, finding an object on a cluttered counter. Google naming what the tool will not do, in its own launch copy, is the kind of disclosure that is easy to skip and that most AI launch posts skip.

Who actually tested it before launch

Google says more than 1,000 members of Aira's Trusted Tester network, people who are themselves blind or have low vision, stress-tested Guided Vision across their daily routines before it shipped. The company also says it gathered feedback from blind and low-vision communities in India, Brazil, Singapore, Indonesia and Japan, specifically to check that descriptions, reframing cues and answers felt natural across languages and cultural context, not just for English-speaking US testers.

That is a larger and more geographically spread testing base than most consumer AI features get before a public launch, and it matters more here than usual. A feature that tells someone what is in front of them has to be reliable across accents, lighting, home layouts and language, or it simply does not work for the people it is meant to serve.

How to turn it on

Guided Vision is reached three ways: through the Gemini app's profile settings, through the Android Accessibility Shortcut, or from inside the TalkBack menu for people already using that screen reader. None of it requires a separate app or subscription beyond having a supported Android device and an updated Gemini app. Coverage depends on Gemini Live being available in a user's market, so it is not yet universal.

Our take

Guided Vision is a real, tested accessibility feature, not a demo dressed up as a product: more than a thousand blind and low-vision testers across five countries used it before Google shipped, and the model learned from genuine visual-interpretation expertise rather than generic image captioning data. What makes it credible is what Google admits it is not. A company that says its own tool "can make mistakes" and rules out navigation and obstacle detection, the two hardest and most dangerous use cases, is drawing an honest boundary rather than overselling a research preview as a white cane replacement. Judge Guided Vision on reading labels and finding lost keys, where the evidence says it works, not on tasks Google itself says it was never built for.