Gemini moves from chat to controller

Google is trying to turn Gemini from a chatbot you talk to into a software layer that presses buttons for you on the phone screen. On Pixel phones in the March 2026 Pixel Drop, Gemini can use visual reasoning to work through certain third-party apps instead of stopping at a text answer. (androidpolice.com) That changes the job description of the assistant. A normal chatbot gives instructions like a recipe card, but this version can follow the recipe inside apps by placing an order or booking a ride when the right permissions are enabled. (androidpolice.com) Google is also adding a filing cabinet for these conversations. “Notebooks” in the Gemini app let users create a topic-specific workspace, attach files from Google Drive or a computer, and keep the chat tied to that subject instead of starting from scratch every time. (blog.google, lifehacker.com) Those notebooks are linked to NotebookLM, which is Google’s research tool for reading your documents and answering questions from them. Google said the sync works both ways, so a notebook started in Gemini can continue in NotebookLM and then come back into Gemini with the same project context. (blog.google, 9to5google.com) Put those two moves together and Gemini starts to look less like a search box and more like a dispatcher. One part keeps the memory of the task, and the other part reaches into apps to carry the task out. (androidpolice.com, blog.google) Google made a second change this week that shows how seriously it now treats Gemini as a front door for real-world action. On April 7, 2026, the company said Gemini would show one-tap access to crisis hotlines and steer users toward human support when prompts suggest severe mental health distress or suicidality. (blog.google, 9to5google.com) Google paired that product change with money. The company said it will provide $30 million through Google.org to expand crisis centers and train hotline workers, including support for organizations in the United States and abroad. (blog.google, economictimes.indiatimes.com) The timing is awkward because a separate report this week described the risk of giving app-level Gemini features deeper reach. Security researchers at CloudSEK said they found 32 live Google application programming interface keys hardcoded in 22 Android apps with more than 500 million installs combined, and they argued those keys could expose Gemini endpoints when artificial intelligence features were enabled on a project. (securityweek.com, digit.in, theoutpost.ai) That is not the same thing as Gemini reading every phone by default. The reported problem is closer to leaving a master key taped under the doormat: if developers embed the wrong Google key in an Android app, attackers may be able to call Gemini-related services they were never meant to reach, which can create data exposure and surprise costs. (securityweek.com, firstpost.com) So the story this week is not just that Gemini got smarter. Google is wiring memory, app control, and crisis routing into the same assistant, which makes Gemini more useful when it works and more consequential when permissions, safety rules, or developer integrations fail. (androidpolice.com, blog.google, blog.google, securityweek.com)