Ten projects share an expert-first theme

Google Research used its 14 October Research@ event to present ten AI initiatives designed to extend scientific and professional work. The portfolio ranges from planetary forecasting and clinical research to document verification, numerical models and collaborative interface generation. Rather than one foundation-model launch, it is a view of how Google is packaging specialised models, agents and datasets for experts.

The common claim is amplification rather than replacement. Doctors, scientists, public-health officials, teachers and developers remain responsible for goals and decisions, while AI handles search, simulation, forecasting or verification at a scale that would otherwise be difficult. The credibility of that model depends on access to evidence, known limitations and evaluation in the setting where a tool will be used.

Earth AI becomes more programmable

Google is expanding early access to experimental Earth AI capabilities. Users can conduct deeper geospatial analysis in Google Earth, generate risk assessments through a planetary prediction engine and integrate Geospatial Reasoning agent skills through an experimental agent-to-agent API. Dataset Search adds access to more than 100 million public datasets, while TimesFM supports nowcasting and forecasting.

The geospatial stack includes more than 70 tools and foundation models for remote sensing and population dynamics. Google.org is also providing US$10 million for organisations working on crisis resilience and public health. Programmatic access can let teams embed satellite interpretation, routing and demographic embeddings into their own workflows, but the experimental label makes local validation essential before operational decisions.

Agents assist science and technical review

A multi-agent orchestrator can reproduce published methods and apply them in new settings. Google cites simulated catheter designs intended to reduce bacterial infection as one example. Related systems include Empirical Research Assistance, which can test many code variations, and Co-Scientist, which generates hypotheses. These tools can widen the search space, but simulated or generated findings still require independent scientific confirmation.

Deep Review provides a different form of assistance: a four-stage agentic process that checks logical consistency and grounds facts across complex documents. Google says an early version has reviewed more than 16,000 submissions to major computer-science conferences and is now being opened to the public for technical material including legal agreements and financial reports. Reviewers should treat it as another critical reader, not a substitute for accountable professional advice.

Numerical and health models move closer to workflows

TimesFM and TabFM are designed for time-series and tabular distributions rather than general text. Both are available to BigQuery users, and Google plans to integrate them into Gemini for business so people can forecast sales or hospital capacity over large operational tables. Keeping sensitive data within an organisation’s perimeter is a stated design goal, though administrators must still configure access and retention correctly.

MedGemma has passed ten million downloads from more than 65,000 users, according to Google. New research adds tool use and explores mammography, ultrasound and electrocardiograms. AMIE research has expanded towards multimodal and long-term clinical reasoning; in one study involving 98 patients, clinicians said AI summaries helped preparation in 75 per cent of cases. Google stresses that these systems remain part of a research journey.

Disaster prediction, teaching and prototyping

Groundsource used Gemini to convert public crisis reports into a dataset containing more than 2.6 million flood events. Google says the data helped train a model that predicts urban flash-flood likelihood up to 24 hours ahead and has been released for research. For education, a generative-UI experiment lets teachers create guided STEM simulations, with more than 30 teacher-vetted examples and a school pilot.

Play with Putty applies generative interfaces to collaborative prototyping, allowing teams to edit and interact with generated software together in real time. A waitlist is open through Google Labs. Across all ten projects, the key question is not whether AI can produce an impressive artefact, but whether experts can inspect assumptions, correct errors and connect results to real evidence. Google’s portfolio is strongest where those human checkpoints remain explicit.

Access and evidence differ across the portfolio

Some of the announced systems are public, others are experimental or available through a pilot or waitlist. Readers should not treat the ten highlights as one uniform product release. A researcher considering an agent API, a BigQuery model or a school simulation needs to confirm eligibility, region, data terms and whether the interface is supported for production use.

The evidence also varies. Google cites peer-reviewed clinical studies, technical reports, public benchmarks, usage counts and demonstrations. Each supports a different conclusion. A benchmark can compare predictive performance, while a field study can reveal workflow effects and a download count shows interest rather than accuracy. Responsible adoption starts by matching the claim to the evidence behind it.

The breadth of Research@ illustrates a shift from general conversation towards systems that combine models with tools, retrieval and domain data. That increases practical value and the number of failure points. Evaluators should inspect the full pipeline, including datasets, tool permissions and human escalation, rather than attributing every result to the model named in the interface.