Microsoft|AI for (Public) Goods

Makes it practical to apply deep learning to satellite and remote sensing imagery, powering work from precision agriculture and disaster monitoring to climate change research.

Expands the set of druggable proteins by rapidly pinpointing hidden pockets on structures once considered untargetable, opening new avenues for drug discovery.

Maps building footprints in rapidly urbanizing cities from only a few hundred local labels, giving planners current data on growth and post-conflict recovery.

Enables large-scale acoustic monitoring of tropical biodiversity by automatically identifying bird and frog species from continuous soundscape recordings.

Maps where and when the world builds solar and wind power so policymakers and researchers can measure and accelerate progress toward clean energy and sustainable development goals.

Gives disaster first responders accurate building damage maps within two hours of a satellite pass so relief efforts can be targeted where they are most needed.

Gives agricultural monitoring a large, openly available benchmark for automatically mapping crop field boundaries from satellite imagery across 24 countries, enabling scalable food security and land use assessments.

Enables automated acoustic monitoring of birds and amphibians of conservation concern across Puerto Rico's forests by accurately identifying their calls in continuous field recordings.

Enables automated identification of 100 bird species from audio recordings, supporting large scale acoustic monitoring of avian biodiversity.

Automatically flagging true beluga calls in passive acoustic recordings speeds up monitoring of an endangered Cook Inlet population so conservation managers can track its seasonal occurrence and respond to threats.

Lets research groups with limited computing resources adapt large protein language models to predict protein interactions and structural symmetry using far less memory and orders of magnitude fewer parameters.

Delivers a decade-long, cloud-free global map of flood extent that pinpoints historically flood-prone communities and supports real-time disaster response.

Speeds post-disaster damage mapping so humanitarian responders can direct shelter, medical aid, and food to the hardest-hit communities within hours of a disaster.

Maps 1,363 utility-scale solar farms across India and reveals that over 74 percent were built on ecologically or agriculturally valuable land, giving planners the evidence to site future solar capacity without displacing natural ecosystems.

Lets frontline health workers screen young children for stunting and malnutrition from a single smartphone depth scan, without specialized anthropometric equipment in low-resource settings.

Enables classifying satellite imagery in regions where labeled training data is scarce or unavailable, supporting applications like disaster response and land-use monitoring without any task-specific training.

Delivers quarterly global maps of building density and height so planners and researchers can track where settlements are growing and inform climate resilience and adaptation efforts.

Exposes where aerial road-mapping models fail on tree-occluded roads so geospatial maps stay accurate even where canopy hides the ground.

Lets data holders release GAN-generated synthetic images that resist membership inference attacks, keeping individual training records private while preserving the data's usefulness for downstream tasks.

Predicts how proteins assemble into symmetric complexes from a single sequence at roughly 80,000 proteins per hour, enabling proteome-scale insight into protein function that would otherwise require slow all-atom structure modeling.

Advances a satellite and AI platform that lets agencies and scientists locate endangered whales across remote oceans, filling gaps in marine mammal distribution to support conservation of at-risk species.

Lets researchers train Earth observation models by streaming imagery straight from cloud storage at the same accuracy and speed as local disk, removing the cost and friction of copying petabyte-scale datasets.

Lets conservationists monitor the activity and abundance of regionally rare Himalayan bird species from passive field audio recordings to guide habitat protection in Nepal.

Lets researchers rapidly discover predictive geospatial features and delineate similar regions, such as maize yield response zones in Rwanda, without heavy computing or coding.

Protects fragile remote ecosystems by enabling continuous, autonomous wildlife monitoring where power and connectivity are scarce, streaming real-time conservation insights to the people who safeguard them

Makes COVID-19 diagnosis from chest x-rays more trustworthy by stopping models from latching onto dataset artifacts, so predictions hold up on patients and scanners the model has never seen.

Helps clinicians distinguish COVID-19 pneumonia from other lung conditions and predict patient mortality by combining chest CT scans with clinical data.

Enables fully automated detection and segmentation of metastatic prostate cancer lesions in whole-body PSMA PET/CT scans to support personalized radiopharmaceutical therapy and treatment response monitoring.

Accelerates ecological monitoring of climate-threatened Himalayan glaciers by letting experts map and correct clean-ice and debris-covered glacier extents from satellite imagery.

Enables assistive apps to recognize banknotes across 17 currencies and 112 denominations so people who are blind or have low vision can identify money independently.

Automatically detects, classifies, and counts calls from four acoustic blue whale populations in ocean recordings so each population's conservation status can be better assessed.

Speeds leprosy diagnosis in underserved communities by letting health workers screen skin lesions at over 90% accuracy, helping prevent irreversible nerve damage and curb ongoing transmission.

Locates industrial poultry barns across the United States from high-resolution aerial imagery, giving regulators and researchers a nationwide view of concentrated animal feeding operations and their environmental footprint.

Produces the first nationwide open map of poultry factory farm locations, giving regulators and the public data to monitor air, water, and health risks that were previously undocumented.

Lets open-data users quickly discover and vet whether a dataset fits their needs and responsible AI policies before they build on it.

Shows that AI otoscopy tools which detect middle ear disease lose substantial accuracy on patients from new clinics, flagging a gap that must be closed before such tools can be trusted for real world ear health screening.

Flags satellite inputs that fall outside a model's training distribution so globally deployed Earth observation systems stay trustworthy in low-data regions.

Pinpoints Renewables Acceleration Areas across Europe where wind and solar can scale up to meet REPowerEU targets while sparing the lands most important for biodiversity and rural communities.

Helps radiologists catch breast cancers earlier by flagging suspicious screening MRI scans and highlighting the tissue regions that drive each prediction.

Forecasts household food insecurity a month ahead so humanitarian agencies can target aid to the most vulnerable families in southern Malawi before shocks hit.

It reveals when buildings such as poultry barns and solar farms were constructed or demolished from satellite image time series, enabling longitudinal monitoring even when footprint labels exist for only a single date.

Clarifies which tumor characteristics deep learning can reliably measure from cancer PET scans, guiding safer use of AI-derived imaging biomarkers in clinical oncology.

Reveals which long-term COVID-19 symptoms burden medically underserved patients so safety-net providers and policymakers can better target their care.

Aligning image resizing and normalization with pre-training conventions makes standard ImageNet-pretrained models a competitive, low-cost baseline for satellite and aerial image analysis, lifting downstream accuracy by up to 32 points.

Maps how giraffe social ties and movements differ by sex and age, informing conservation planning for endangered Masai giraffes in Tanzania.

Gives journalists and human rights defenders a rigorous, continuously updated benchmark to measure whether AI deepfake detectors actually hold up against emerging generators and adversarial attacks in high-stakes real-world cases.

Gives fire management agencies reliable spatio-temporal forecasts of where wildfires are likely to occur so they can act early to prevent, detect, and suppress them.

Quantifies for every US state how many COVID-19 deaths could have been averted with higher vaccination coverage, giving public health leaders evidence to target vaccination efforts.

Establishes maternal obesity as a modifiable risk factor for sudden unexpected infant death, informing prevention guidance that could lower infant mortality.

Shows that audio-language models can recognize wildlife sounds without any labeled training examples, opening a path to cheaper annotation-free acoustic monitoring for conservation.

Lets conservationists re-identify individual giraffes and other wildlife across repeat surveys far faster and cheaper, supporting population monitoring of endangered species.

Enables reproducible, cross-herd evaluation of automated caribou counting from aerial survey imagery so wildlife managers can monitor Arctic herds without retraining detectors for every new survey.

Enables faster, automated mapping of Himalayan glacial lakes to support outburst flood risk assessments that protect downstream communities and infrastructure.

Enables reproducible, retraining-free counting of caribou from aerial survey imagery so wildlife managers can monitor Arctic herd populations across different years and regions.

Lets conservation teams count wildlife from aerial and drone imagery using cheap point labels instead of costly bounding boxes, cutting the time and expense of large-scale animal surveys

Enables early detection of retinopathy of prematurity from smartphone-recorded retinal videos so that preventable childhood blindness can be caught and treated where pediatric eye specialists and imaging cameras are scarce.

Gives clinicians a validated risk tool that flags hospitalized COVID-19 patients at high risk of death and organ failure within the first 24 hours of admission, supporting faster triage and goals-of-care decisions.

Enables accurate detection of the rare retinal disease macular telangiectasia type 2 from OCT scans even when labeled patient data is scarce.

Puts automated wildlife detection and species classification within reach of conservationists with little or no coding background, accelerating large-scale biodiversity monitoring.

Opens biodiversity monitoring to practitioners without programming expertise by letting multimodal language models classify species and interpret animal behavior at scale, accelerating conservation.

Identifies the modifiable maternal factors that drive the largest share of late stillbirths in the US, pointing to where prevention efforts can save the most fetal lives.

Shows that low-cost banner ads on a news platform can recruit tens of thousands of people worldwide to self-report COVID-19 symptoms, enabling rapid population-level disease surveillance when testing is scarce.

Reconstructs the low and medium voltage electricity network across Kenya's Kakuma and Kalobeyei refugee camps so utilities and humanitarian agencies can plan reliable power access for more than 200,000 displaced residents.

Reveals that gaps in how US households actually use their devices persist far beyond broadband access, pointing policy toward digital skills and literacy rather than infrastructure alone.

Measures how the risk of hospitalization and death after a breakthrough COVID-19 infection differs by vaccine brand, age, and prior infection so public health officials can better target vaccination and booster efforts.

Eases the slow, labor-intensive work of writing ophthalmology board and residency exams by showing GPT-4 can generate training questions rated on par with those from expert human committees.

Shows New York City policy-makers which neighbourhood social conditions most drive diabetes disparities so they can target local interventions where they matter most.

Helps clinicians decide whether a patient with a pancreatic cyst should be discharged, monitored, or sent to surgery, cutting unnecessary operations while catching cancers earlier.

Provides land use recommendations that help India deploy 500 GW of renewable energy by 2030 while reducing conflicts over crowded lands and safeguarding socio-ecological values

Enables accurate land cover mapping from remote sensing imagery while cutting the training data and compute needed by up to 75 percent.

Measures how much web search actually exposes users to unreliable information sources, showing that most such engagement comes from users deliberately seeking those sites rather than from the search algorithm.

Helps Kenyan health decision-makers anticipate where acute child malnutrition will rise across sub-counties so interventions can be targeted before crises escalate.

Reveals which baseline patient factors predict long-term quality-of-life outcomes for people with acute myeloid leukemia, helping clinicians anticipate decline and guide treatment decisions.

Gives state health agencies a way to anticipate weekly spikes in drug overdose deaths using only publicly available data, so interventions can be aimed before fatalities rise.

Shows that routine contrast enhanced CT scans can approximate the gold-standard MRI measure of liver fat, opening a path to detect fatty liver disease in patients who lack access to specialized MRI.

Enables accurate, interpretable automated screening for macular telangiectasia type 2, a rare sight-threatening retinal disease, from routine OCT scans at expert-grader level.

Reveals which long-term conditions COVID-19 survivors develop and how social factors like race, income, and education shape that risk, informing more equitable post-COVID care.

Shows that sudden unexpected infant death rates rose after the onset of the COVID-19 pandemic and hit infants of Black, younger, and lower-income mothers hardest, pointing prevention efforts toward the families most at risk.

It reveals how strongly a baby's growth predicts late stillbirth risk across the full birthweight range, so fetuses at the extreme centiles can be identified and monitored earlier.

Shows that historical shifts in how left-handedness was reported fully account for the long-cited claim that left-handed people die nine years younger, correcting a widespread public health misconception.

Lets conservationists identify wildlife sounds like birds, frogs, and whales without the costly manual annotation that limits acoustic monitoring, extending detection to species and habitats never seen during training.

Enables broader and more reliable acoustic detection of endangered Cook Inlet beluga whales across noisy, underrepresented habitats to strengthen their conservation.

Measures whether automated CT pancreas segmentation reaches human-level reliability so that CT based early cancer detection and quantitative biomarker tools can be trusted in clinical deployment.

Measures how reliably general-purpose language models answer ophthalmology board questions, showing reasoning models can match or exceed residents and could serve as trustworthy study aids for eye-care trainees.

Provides an expertly annotated acoustic reference dataset that enables automated detection and identification of Neotropical bird species to support biodiversity monitoring in Colombia.

Enables scalable, non-invasive monitoring of tropical bird biodiversity by giving machine learning practitioners an expertly labeled benchmark of Neotropical soundscapes for training and evaluating acoustic detection models.

Measures how vaccine type and a prior infection change the odds of severe COVID-19 among breakthrough cases, giving clinicians and health officials evidence to guide vaccine and booster decisions.

Enables scalable monitoring of thousands of remote archaeological sites from satellite imagery to protect cultural heritage from looting.

Turns dense state Medicaid eligibility rules into an accessible web tool so people facing redetermination can check whether they still qualify for health coverage.

Turns handwritten markings on seized elephant tusks into a low-cost forensic signal that helps investigators link ivory shipments and disrupt trafficking networks

Enables monitoring of nearly 2,000 archaeological sites across 16 countries to detect looting and disturbance from monthly satellite imagery, helping protect cultural heritage from destruction.

Predicts which SARS-CoV-2 epitopes trigger an individual's protective T-cell response from their HLA genotype, enabling personalized assessment of immunity and reinfection risk.

Helps under-resourced conflict-analysis nonprofits turn messy armed conflict reports into usable structured data with far less manual work and reduced reporting bias.

Gives humanitarian responders accurate maps of refugee camp buildings, solar panels, roof materials, and sanitation facilities so they can better plan infrastructure, distribute resources, and respond to disasters.

Identifies treatable maternal infections during pregnancy as risk markers for sudden unexpected infant death, pointing to prenatal screening and treatment as a lever to help protect newborns.

Estimates how long protective immunity lasts after MenAfriVac vaccination to help decide when booster doses are needed across Africa's meningitis belt.

Improves acoustic identification of both common and rare tropical bird species in remote rainforests, strengthening biodiversity monitoring where annotated call data is scarce.

Helps law enforcement identify trafficked children faster by automatically spotting school uniforms in intercepted images to narrow down a child's school of origin.

Shows that limited language support in AI systems is an independent barrier to adoption, with low-resource-language countries having about 20 percent fewer AI users, giving policymakers a measurable target for closing the global AI divide.

Shows that machine learning models built and tested on differentially private synthetic data can misjudge their real-world accuracy and bias, helping teams share sensitive data privately without unknowingly deploying unreliable or unfair models.

Measures how waterholes, homesteads, and big trees around Oshikango, Namibia shifted between 1943 and 1972, turning decades-old aerial photographs into a quantified record of postwar environmental and land-use change.

Pinpoints which areas of Malawi are most at risk of food insufficiency using only publicly available data, helping direct food assistance toward the communities repeatedly hit by food crises.

Shows how to share sensitive tabular data as privacy-preserving synthetic datasets without sacrificing model accuracy or fairness, so regulated fields like health care and humanitarian aid can build machine learning models on data that could not otherwise be released.

Quantifies how a medical specialty's average compensation tracks with the share of its trainees who are female, giving residency program directors evidence to argue for income and specialty-choice equity for women physicians.

Distills eleven practical takeaways that help future AI for good collaborations deliver real-world impact across sustainability, health, humanitarian aid, and social justice.

Helps clinicians catch diabetic eye disease earlier by accurately flagging which patients need referral from color fundus photographs.

Lets conservation teams map hard-to-find livestock enclosures across the Serengeti Mara from satellite imagery without any starting labels, making landscape monitoring feasible on minimal annotation budgets.

Detects small, easily missed metastatic lesions in men with recurrent prostate cancer from PSMA PET scans, supporting earlier and more targeted treatment decisions.

Speeds the development of longer-lasting next-generation solar cells by forecasting and explaining how they degrade from only a few hours of testing, cutting stability screening time 5 to 20 fold.

Lets ecologists count and locate cattle and elk across large landscapes directly from satellite imagery, enabling wildlife and rangeland monitoring without costly field surveys.

Maps dwelling types from satellite imagery so at-risk homes in India's low-income settlements can be scored and prioritized for protection before floods and other hazards strike.

Grounds datacenter capacity planning, efficiency investment, and energy policy in production-realistic per-query energy and water estimates for large-scale AI inference, replacing figures overstated by 4 to 20 times.

Delivers accurate local land-use and land-cover maps for African agricultural regions to support the resource management and planning that strengthens food security.

Enables a single satellite-imagery model to handle segmentation, classification, and zero-shot recognition well, so earth-observation teams no longer trade vision accuracy for language capability.

Defines a shared vocabulary of agency capabilities and a productivity benchmark so AI assistants can measurably speed up the everyday map-making work of GIS practitioners.

Flags abandoned fishing nets in sonar imagery so cleanup teams can locate and remove the gear that entangles and kills marine wildlife.

Quantifies how much multifactor authentication protects online accounts, showing it keeps over 99.99 percent of enabled accounts secure and cuts compromise risk by more than 99 percent, strengthening the case for turning it on by default.

Measures how reliably people can tell real photographs from AI-generated ones, quantifying the misinformation risk that makes watermarking and detection tools necessary.

Exposes hidden dataset biases that let melanoma-detection models cheat, guarding against unreliable AI in early skin cancer screening.

Helps materials scientists and chemists trust machine learning predictions and surface genuine scientific insights rather than spurious correlations by making models interpretable and explainable.

More accurate local land-cover maps help target agricultural monitoring and food security efforts in African regions that global maps map poorly.

Gives institutions a rigorous way to measure how much a generative model leaks about the individuals in its training data before that model or its synthetic outputs are shared.

Maps where Punjab farmers adopt water-saving rice practices so groundwater conservation and climate adaptation policies can target the fields that need them.

Lets conservationists identify wildlife species in camera trap photos without expert-labeled training data, making large-scale population monitoring cheaper and faster to deploy.

Maps the vast landscape of open datasets hosted on GitHub so researchers and organizations can discover and reuse them to accelerate AI work on complex societal problems.

Pinpoints which chronic kidney disease measures flag outpatients at higher risk of dying from COVID-19, helping clinicians and policymakers prioritize protection and vaccination for the most vulnerable.

Sharpens regional poverty rate estimates by fusing satellite imagery with a short survey, helping target social programs to households below the poverty line.

Gives transplant programs and policymakers more accurate kidney offer acceptance predictions to support fairer risk-adjusted program evaluations and allocation policy.

Clarifies how the survey behind prompt validation was built so large language models can be assessed more reliably as ophthalmology education tools.

Lets sites with limited, biased medical imaging data collaboratively produce higher-quality, less-biased synthetic images without any of them exposing their patients' raw data.

Shows that blood vessel patterns in routine retinal photographs can serve as a reliable non-invasive biomarker for detecting diabetic retinopathy and, by extension, screening for cardiovascular disease risk.

Warns practitioners that unverified open health data from India can yield misleading pregnancy-outcome predictions, helping guard against flawed AI decisions in resource-poor public health settings.

Helps radiologists catch small, still-curable pancreatic tumors on CT scans earlier, improving the odds of timely treatment for a cancer usually found too late.

Guides agricultural survey planners to spend limited budgets on the label checks that most protect crop map accuracy, which underpins food security decisions.

Measures how well tumor segmentation models transfer between cancer types, pointing toward PET/CT lesion detection that generalizes without type-specific labeled data.

Exposes hidden dataset biases that undermine AI screening for childhood ear disease, so diagnostic models can generalize across populations and help prevent avoidable hearing loss.

It gives researchers and policymakers zip-code-level broadband coverage figures to target the digital divide while mathematically protecting the privacy of individual households.

Makes prostate cancer metastases stand out on PSMA PET scans without needing hand-labeled lesions, supporting easier detection of spreading disease.

Shows that site-level incentives, not just algorithmic recommenders, trap users in unreliable-content rabbit holes, pointing regulators and platforms toward the economic drivers of misinformation exposure.

Produces accurate cropland maps across Africa from almost no hand labeling, helping agricultural planning and food security efforts reach regions where high-resolution maps are scarce.