Google's annual I/O conference tends to get covered as a product show. This year the more important story is how far Google has pushed AI into work that used to require specialized human experts: scientific research, medical diagnosis, weather forecasting, and software engineering.
The centerpiece of the research announcements is Gemini for Science, a suite of tools built on years of internal research, two papers of which were published in the journal Nature on the same week as I/O. One tool, called Computational Discovery, generates and scores thousands of code variations in parallel, letting scientists test hypotheses that would take months to explore by hand. Another, called Hypothesis Generation, runs a kind of automated debate between AI agents to propose and challenge new research ideas, with citations attached to every claim so scientists can verify the sources.
For context: with millions of scientific papers published every year, no individual researcher can read everything relevant to their field. These tools do not replace scientific judgment. They compress the reading and hypothesis-sorting work so that human experts spend more time on the decisions that actually require human thinking.
The health announcements are where the numbers get striking. In a study of nearly 14,000 participants using an experimental symptom-checking tool on Fitbit, independent clinicians reviewing the same conversations preferred the AI's assessments roughly twice as often as those from other clinicians. In a separate study of 1,779 participants preparing for doctor visits, 15 percent more users felt better prepared after using the AI tool, and 13 percent more felt confident they could make the most of their appointment. These are early research numbers, not clinical approvals, but they point toward where AI is headed in the doctor-patient relationship.
MedGemma, Google's open medical AI model that any developer can download and run on their own servers, has now passed 5 million downloads. Hospitals and health tech companies are already building on it: Taiwan's National Health Insurance Administration used it to analyze over 30,000 pathology reports. A tuberculosis screening tool built on MedGemma runs entirely on a phone with no internet connection. For healthcare operators in regions with limited infrastructure, that last point is significant.
On disaster response, Google's WeatherNext model predicted Hurricane Melissa's rapid intensification and Jamaican landfall five days in advance in October 2025, giving authorities enough time to warn the public. The flood forecasting system now covers 2 billion people across 150 countries. These are not experimental features; they are live and in use by national weather agencies.
For business operators, the most practically relevant announcement may be Gemma 4, Google's open AI model that any organization can download, run on its own servers, and customize. It hit 100 million downloads in its first month. It runs under a fully open Apache 2.0 license, meaning no usage restrictions, no fees, and no requirement to send data to Google's servers. For organizations with strict data privacy requirements, or those who want to build AI tools without becoming dependent on a single cloud provider, this is the most important development of the entire conference.
Google also launched Antigravity 2.0, a coding platform that can run multiple AI agents simultaneously to write, test, and debug software. On stage, it built a working operating system from scratch. The practical implication for non-technical businesses is that the cost and time to build custom software tools is falling sharply. Teams that could not afford custom software development are increasingly within range of it.
Google is spending roughly 180 to 190 billion dollars on infrastructure this year, about six times what it spent in 2022. That scale of investment does not go into products that stay experimental. The tools announced at I/O are the visible surface of a much larger infrastructure build that will make AI capabilities cheaper and more accessible across every industry over the next two to three years.