Google I/O: Gemini gets
a mind of its own

The world’s most advanced AI is no longer waiting for instructions.

That was an unspoken message running through the Google I/O 2025 developer conference, held in Mountain View near San Francisco, California, this week. In the previous week, Google had announced a string of new features for its Android mobile platform. On Tuesday, the spotlight of the event shifted firmly to the AI platform Gemini, as a system preparing to operate on its own terms.

Demis Hassabis, CEO of Google DeepMind and winner of last year’s Nobel Prize for Chemistry for his AI breakthroughs, described Gemini as a “universal AI agent” during the opening keynote session. He didn’t announce another breakthrough, but rather a transformation already in motion.

“We’re giving Gemini memory, context, and the ability to take action,” he said. “This is the foundation for intelligent agency.”

Gemini can now carry out tasks across apps, respond to real-world context using a phone camera, and hold fluid, ongoing interactions. Hassabis showed it planning steps, using tools, and navigating digital environments with minimal intervention.

_{Aarush Selvan, head of Deep Research in Google’s Gemini division. Photo ARTHUR GOLDSTUCK.}

“We’re building agents that can reason, that can use other systems, and that can help people get more done in complex, open-ended settings.”

Among the features redefining what that means is one called Deep Research: a capability that allows Gemini to step away from the interaction, browse the web in real time, and return with a structured, fully referenced report. Behind that system is Aarush Selvan, a product manager at Google who has spent the past year building out Gemini’s role as a research partner.

On the sidelines of I/O, Selvan told Gadget in an exclusive interview that Deep Research broke the rhythm of “prompt and reply”. As a result, it gives the model time to do something far more complex.

“We’ve left the synchronous chat paradigm. Deep Research allows Gemini to pause, explore, and come back with something it has worked through.”

The model takes up to 15 minutes to browse the web, collect information, verify it against multiple sources, and then synthesise that information into a coherent, structured report. Every few sentences, Gemini cites its sources. At the end of the report, it offers a full list of the web pages it consulted.

As opposed to the sprint of most AI chat results, this is more like a slow-burn investigation.

“That time isn’t just about pulling in more content. It’s about giving the model the space to reason across what it finds. To filter, to weigh sources, to build a useful structure. We’ve found that the extra time results in a different quality of output.”

He described Deep Research as a turning point in how users relate to AI.

“When people come to Gemini, it’s often because they’re thinking through something: a topic, an idea, a project. The question becomes: what can the model do when it’s not just reacting, but working through the problem itself?”

Selvan said users awee already approaching Gemini for everything from writing assistance to historical analysis to technical planning.

“People use it to build business models, to brainstorm, to explore policy implications. Some just want to understand a topic more deeply. Deep Research gives them something to work with – not a final answer, but a meaningful starting point.”

His own experience includes using the tool while visiting South Africa.

“I’d read a lot about the country’s history as a teenager,” he said. “When I visited last year, I used Gemini to refresh that understanding.”

He also saw the tool as useful for thinking through hard or sensitive topics.

“There are times when you want to bounce ideas off something, but it’s not the kind of conversation you’d have casually. Gemini can be a companion in that process.”

To support that kind of depth, Selvan said, transparency was essential.

 “We make the sources obvious. It’s part of the experience. Every few sentences are linked, and you get the full trail of pages the model used. We want people to check, explore, go deeper.”

Different tasks require different approaches. For code, the model shows working examples. For maths, it walks through logical steps. For everything else, there’s a feature called “thoughts”: a view of the internal reasoning Gemini used to reach its conclusion.

“It’s about trust. People want to know how the model got there.”

Selvan acknowledged that Gemini’s growing capabilities don’t come with one-size-fits-all use cases.

“There will never be one app that uses everything the models can do,” he said. “What we’re building is a constellation: NotebookLM for long documents, Labs for image generation, Gemini app for personal tasks. Each one explores what this kind of model can unlock.

 “We’re just beginning to understand how people want to use this. Features like Deep Research are shaped by what we see in real usage. That’s where the insights come from.”

Asked what happens when people begin to depend on Gemini for structuring how they think, Selvan was reflective. “It gives people a framework. They still need to bring their own context, their own values. But it helps them work through the shape of the problem.”

Hassabis, too, avoided grand conclusions during his keynote.

“This is a beginning,” he said. “And beginnings are where everything gets interesting.”

* Arthur Goldstuck is CEO of World Wide Worx, editor-in-chief of Gadget.co.za, and author of “The Hitchhiker’s Guide to AI “.