Interviews
Aira on How Blind Testers Trained Google’s Gemini Guided Vision
Overview
Watch this interview
Captions, where YouTube has them, are on the video's CC button.
Overview
Google’s Guided Vision in Gemini Live was trained on real Aira visual interpreting sessions. Everette Bacon and Lynette Frison from Aira join Steven Scott and Shaun Preece to explain how the partnership with Google DeepMind came about and what blind testers taught the AI. They also discuss why AI image descriptions still treat blind adults like children.
Expanded Summary
Everette explains how Aira’s Build AI programme gave Explorers free minutes in exchange for sharing their sessions, which Google DeepMind then used to train its Astra model. Lynette ran the Trusted Tester programme. Hallucinations were the testers’ biggest complaint, and one early version told a tester to go and find a sighted person. A human visual interpreter still watches every Aira AI session and steps in if something goes wrong.
Steven and Shaun ask what blind users bring that mainstream data cannot. Blind people ask for different things: which way to turn a bottle so the label can be read, or where the buttons sit on a remote rather than what colour they are. The conversation then turns to sensitive requests. Finding dog poop is the number one reason Explorers call, and medical and personal questions come up too.
The panel also covers whether Aira will charge for its AI, how soon humans could drop out of the loop, and frustration with AI tools that will not describe legal images or read text with swear words in it. The episode ends with news that Build AI is now available in the UK.
Chapters
0:00 – How the Aira and Google DeepMind partnership began
4:01 – Running the Trusted Tester program
6:01 – Hallucinations and the human in the loop
11:02 – Guided Vision versus Aira AI
12:28 – What blind users’ prompts teach the AI
15:06 – Sensitive requests, from dog poop to medical questions
24:43 – How close is AI without a human backup
25:56 – Will Aira charge for its AI
30:48 – Image description limits and infantilization
37:39 – Build AI expands to the UK
Relevant Links
Aira: https://aira.io
Aira Build AI FAQ: https://aira.io/build-ai-faq/
Guided Vision in Gemini Live: https://blog.google/innovation-and-ai/products/gemini-app/guided-vision-gemini-live/
Google DeepMind: https://deepmind.google/