Google Earth AI Wants to Turn Place Into Public-Health Infrastructure
Google is making a bigger claim for geospatial AI than faster map production. Its latest research suggests that reusable mathematical representations of places could become part of the basic analytical infrastructure used to anticipate disease, fill gaps in health surveillance and direct limited resources.
The idea is being tested through Google Earth AI’s Population Dynamics Foundation Model, or PDFM. Instead of representing a location only through satellite pixels or census variables, PDFM combines aggregated search trends, patterns of mobility and place activity, weather, air quality and characteristics of the built environment. It compresses those signals into location embeddings, effectively numerical fingerprints of places that can be added to existing statistical and machine-learning workflows.
That distinction matters. Public-health teams frequently work with datasets that are several years old, spatially incomplete or trapped inside administrative boundaries. Google is not proposing that PDFM replace epidemiological records. It is proposing a reusable context layer that might help models work when conventional inputs are late or sparse.
Five tests, not one benchmark
The accompanying 41-page research paper, currently a preprint, brings together five partner-led evaluations across four countries and very different health problems.
Along the United States-Canada border, adding Canadian context increased the variation explained in county-level measles, mumps and rubella vaccination coverage from 16% to 22%. For cardiovascular mortality across roughly 3,100 US counties, PDFM-based models performed comparably to models using older census variables, while reducing large outlier errors.
In Mexico, combining PDFM with Google’s TimesFM model improved one-month dengue forecasts in up to 72% of municipalities experiencing active transmission. In the Democratic Republic of Congo, a model developed with WHO’s Regional Office for Africa improved the quality of shortlists for cholera emergence four to eight weeks ahead. At eight weeks, the average number of correct locations among the five highest-risk health zones rose from 1.78 to 2.10.
The fifth evaluation added geographic context to postpartum-depression risk screening. The improvement in headline predictive accuracy was small, but the signal transferred into states excluded from training. In simulations, that could help a constrained health system reach more rural mothers or reduce unnecessary follow-ups. The researchers are careful to say that area-level embeddings do not replace individual income, insurance or clinical information.
These are encouraging results, but not proof of a universal public-health model. Different experiments use different targets, baselines and geographies. The paper has not completed peer review, and several gains are statistically meaningful without being operationally transformative on their own.
From prediction to an agentic workflow
The more consequential development may be how Google packages the models. According to Google’s account, WHO AFRO used a prototype Geospatial Reasoning agent during the ongoing Ebola outbreak in the DRC to examine remote mining corridors, mobility and exposure. Google says the workflow identified 48 exposed settlements and more than 45,500 people at risk in minutes, supporting mobile-laboratory deployment and border surveillance.
That is a company-reported operational example, not an independently audited outcome. It does not establish that the model reduced transmission or saved lives. It does show the direction of travel: a public-health analyst asks a question in ordinary language, while an agent discovers datasets, performs spatial processing and presents candidate areas for action.
A crowded geospatial intelligence layer
Google is not alone in building reusable spatial foundations. WorldPop produces open, high-resolution demographic datasets used in health systems and humanitarian response, with an emphasis on transparent methods and government collaboration. NASA and IBM’s open-source Prithvi family learns primarily from Landsat and Sentinel-2 imagery and can be adapted for floods, fires, crops and other Earth-observation tasks.
PDFM’s differentiation is its combination of environmental signals with aggregated human behavior and place activity. That may capture conditions that imagery alone cannot see. It also creates a dependence on signals and infrastructure that few organizations other than Google can reproduce.
The commercial path is already visible. The embeddings are available in preview as Population Dynamics Insights through Google Maps Platform, while selected academic and public-health users can request no-cost access. What begins as research infrastructure may therefore become a proprietary layer underneath government and humanitarian decisions.
The governance problem is spatial too
Google describes the underlying signals as aggregated and privacy-preserving. That does not resolve every governance question. Search activity and mobility are uneven proxies for human behavior, particularly in places with limited connectivity. A model can appear globally transferable while performing best where Google’s data exhaust is richest.
Public-health agencies will need more than an accuracy score. They will need geographic error analysis, uncertainty at the decision scale, documentation of changing source coverage, audit trails for agent-produced maps and a clear process for local experts to reject misleading outputs. Communities should also have a voice when behavioral data helps determine whether their area is labeled high risk.
The strategic shift is nevertheless important. Geospatial foundation models are moving beyond extracting objects from imagery. They are becoming representations of how places function. If those representations are independently validated and governed well, they could shorten the distance between a weak signal and a public-health intervention. If they are treated as an automated substitute for local knowledge, they could make old data inequalities harder to see.








