Google used its annual I/O developer conference this week to lay out one of its broadest platform updates in years, spanning a redesigned Android, new privacy controls, a conversational artificial-intelligence system and a new machine-learning model intended to make Search better at answering complicated questions.

The company said Android now runs on more than 3 billion active devices, giving even incremental changes to the operating system enormous reach. But the first beta of Android 12 is not incremental in appearance. Google described it as the biggest design change in Android’s history, with a new visual system called Material You that changes colors and interface elements based on a user’s wallpaper and preferences.

Behind the design overhaul is a larger strategic message: Google wants its most important products to feel more personal while also convincing users that personalization need not come at the expense of privacy. That tension ran through much of I/O, alongside a second theme — putting more advanced artificial intelligence directly into consumer products.

Android 12 puts privacy in the interface

Google’s developer preview of Android 12 introduces a new privacy dashboard that shows which apps accessed location, camera and microphone data over time. New status indicators will appear when the microphone or camera is active, and quick-setting toggles will let users disable those sensors across the device.

The operating system will also allow users to give apps an approximate location rather than a precise one. That is a meaningful shift because location permission has often been an all-or-nothing decision. Google is trying to turn privacy into a visible, routine control rather than a setting buried several menus deep.

The company paired those changes with broader privacy measures across its services. In a separate security announcement, Google introduced a quick-delete option for erasing the last 15 minutes of Search history and said it would continue expanding privacy protections built into its products by default.

LaMDA points toward more natural computing

The most technically ambitious demonstration involved LaMDA, short for Language Model for Dialogue Applications. Google presented it as a research system designed to carry on open-ended conversations without being restricted to a narrow set of preprogrammed topics.

Traditional voice assistants are generally strong when a request can be mapped to a familiar command — set a timer, play a song, check the weather — but much weaker when a conversation moves unpredictably. Google says LaMDA is being trained on dialogue so that responses remain relevant and specific even when the subject changes.

The company showed illustrative conversations in which the model spoke as if it were Pluto or a paper airplane, emphasizing the system’s ability to sustain a topic while producing varied answers. Google also stressed that conversational fluency is not sufficient by itself. The company said it is researching measures for factuality, safety and quality before such systems are integrated more broadly into products.

That caution is important. Large language models can generate plausible-sounding responses that are incomplete or wrong. Google’s challenge is to turn research demonstrations into services reliable enough for everyday use without creating a new channel for confident misinformation.

MUM aims at questions that currently require many searches

Google also unveiled the Multitask Unified Model, or MUM, as a new approach to difficult search queries. The company says MUM is built on a transformer architecture, is trained across 75 languages and can understand information across different formats.

The example Google offered was a user who has hiked Mount Adams and wants to prepare for Mount Fuji in the fall. Today, answering that question can require separate searches for elevation, weather, trail difficulty, equipment and seasonal conditions. Google wants MUM eventually to understand the larger task and connect those pieces of information more directly.

A broader Search roadmap presented at I/O showed how the company sees artificial intelligence moving beyond ranking web pages toward interpreting complex intent. That could make Search more useful, but it also increases the importance of transparency about how answers are assembled and which sources are represented.

Workspace becomes less like a collection of files

The conference also reflected the changes in work produced by the pandemic. Google introduced Smart Canvas, a series of updates intended to blend Docs, Sheets, Slides, Meet and other collaboration tools into a more connected workspace.

New “smart chips” can display information about people, files and meetings inside documents. Users will be able to start meetings directly from Docs, Sheets and Slides, create checklists, use new table templates and work with pageless document formats designed for screens rather than printed paper.

The changes are aimed at a workforce that has spent more than a year moving among video meetings, shared documents, chat windows and project tools. Google’s response is to make the document itself more of a workspace — a place where teams can bring together people, data, tasks and conversation.

A platform company broadens its definition of the platform

I/O traditionally gives developers a roadmap for Google’s operating systems and application programming interfaces. This year’s event suggested a wider definition. Android remains essential, but the company increasingly treats artificial intelligence, identity, privacy controls and cloud-based collaboration as common layers across all of its products.

For developers, Android 12 brings immediate work: new design conventions, permission behavior and performance requirements must be tested before the final release. For consumers, the more significant changes may arrive gradually as technologies such as LaMDA and MUM move from demonstrations into products used by billions of people.

Google is betting that the next stage of computing will be simultaneously more personalized and more intelligent. The central question after I/O is whether it can make those systems more capable while also making users feel that they retain meaningful control over the data and decisions beneath them.