Cities and researchers are increasingly using artificial intelligence (AI) to map and predict extreme-heat risk, but a new review published in Climate Risk Management finds that existing research overwhelmingly emphasizes prediction over operational response. The review’s authors describe a substantial “research-to-implementation gap”: studies focus on where people are at risk, with far less attention to how forecasts can be turned into government action.
Why this matters
As climate change intensifies heat extremes, better forecasts alone will not protect vulnerable communities unless governments can act on them. A 2025 Yale study found that U.S. deaths associated with high temperatures rose by 53%, from an annual average of 2,670 in 2000–2009 to more than 4,000 in 2010–2020.
Findings of the review
The Climate Risk Management review examined how AI and machine learning are being used to manage extreme heat. Key statistics from the papers analyzed include:
- 98.4% of the studies focused on prediction.
- 59.2% addressed prevention.
- By contrast, only 10.4% covered response, 3.2% recovery, and 1.6% preparedness.
The authors label this imbalance a research-to-implementation gap: while many projects map hazards and vulnerable populations, far fewer explore how those outputs help officials decide what to do next.
Expert perspectives
Karishma V. Patel, a RAND policy researcher and co-author of the review, said AI’s value is greatest where prediction is central to decision-making—either by scaling insights or by providing higher spatial or temporal granularity.
Jennah Gosciak, a Cornell researcher and PhD student in Information Science who has worked as a Public Interest Technology Fellow with New York City Emergency Management, told Axios that turning sophisticated forecasts into usable government tools requires “robust data systems” that can ingest diverse sources and make data actionable for practitioners. She also noted that simpler heat-vulnerability indexes can be advantageous for longer-term planning because they allow officials to explicitly choose which vulnerable populations to prioritize.
Gosciak contrasted “slow versus fast data”: predictive algorithms may support time-sensitive, trigger-tied decisions, while simpler indexes can better serve strategic, long-range planning.
Practical example: Google’s Heat Resilience data
In June, Google announced it had expanded its AI-powered Heat Resilience data from an initial 14-city pilot to more than 50 cities worldwide. The dataset combines AI and high-resolution imagery to map roof reflectivity down to individual buildings, providing building-level inputs that urban planners can use to target heat interventions. Google said pilot data informed decisions in multiple cities, including cool-roof ordinances and adaptation plans.
Reality check
More precise information does not automatically translate into better outcomes. Experts emphasize that models must be coupled with operational systems and decision frameworks to produce impact. Patel suggested that the next frontier is less about incremental improvements in forecasting accuracy and more about connecting predictions to interventions: integrating changing signals, offering hyperlocal forecasts, and enabling near real-time decision-making.
Such integration could help officials respond to heat risks as they unfold and better account for compound threats involving heat, air pollution, wildfire smoke, and drought.
Bottom line
AI and machine learning can play a substantial role in identifying and forecasting heat risk, but their practical value depends on whether they solve concrete problems or remove bottlenecks in decision-making and operations. As Karishma V. Patel put it: building AI-enabled technology is insufficient unless it actually solves a problem or eliminates an operational constraint.
