Download PDF Spring Bridge on AI: Promises and Risks April 15, 2025 Volume 55 Issue 1 This issue of The Bridge features fresh perspectives on artificial intelligence’s promises and risks from thought leaders across industry and academia. AI's Capabilities Make It a Powerful Tool for Driving Societal Impact Monday, April 14, 2025 Author: Yossi Matias, Avinatan Hassidim, and Philip Nelson AI innovations, developed and deployed responsibly, can help preserve our climate, improve health outcomes, and create a more accessible world for everyone. Many of us entered the sciences driven by a profound curiosity and by a desire to solve the challenges around us and make new discoveries about our world. With artificial intelligence (AI), we are now living in a golden age of research, where the pace of innovation rivals the early days of the internet. The incredible capabilities of AI, including generative, predictive, and diagnostic, make it a transformative tool for addressing some of humanity’s most complex challenges. We are already seeing how AI can help preserve our climate, improve health outcomes, propel scientific understanding, democratize education, and create a more accessible world for everyone. Along with the significant benefits to humanity, there are also legitimate concerns about this nascent technology, from hallucinations and perpetuating biases to privacy and security risks. The development and deployment of AI technologies in a responsible and collaborative manner can help mitigate the risks and maximize the benefits, with alignment across stakeholders and the public and private sectors. Below, we demonstrate the critical role that AI can play in society, drawing on examples of AI innovations from teams at Google Research and collaborative efforts across the research ecosystem to ensure that AI, used responsibly, can truly benefit everyone. Developing AI Solutions to Address Climate Crises The climate crisis is one of today’s most pressing global challenges; it demands urgent and decisive action. At forums such as the annual COP conference, scientists and governments have aligned on the importance of mitigating the extent of climate change while simultaneously helping communities adapt to its catastrophic impact. AI is playing a crucial role in accelerating climate action. It is used to enhance our holistic understanding of the climate system, help mitigate the climate impact of various industries, improve our ability to predict and adapt to climate-related events, and optimize climate action strategies. Healthcare is another domain where AI has already started to deliver on its potential to improve lives. Mitigating Climate Change Take transportation, which is responsible for a significant amount of global and urban greenhouse gas emissions. Road transportation is especially problematic at city intersections, where pollution can be 29 times higher than on open roads (Kumar and Goel 2016). To help address this, Google Research’s Project Green Light uses AI-powered traffic flow modeling and Google Maps’ comprehensive understanding of global road networks to create smart recommendations for traffic light optimization. Results from the pilots with cities indicate a potential for up to a 30% reduction in vehicle stops and a 10% decrease in greenhouse gas emissions. Green Light is now being scaled and deployed around the world from Seattle to Bangalore to Jakarta. In aviation, about 35% of the global warming impact comes from the clouds created by contrails, the condensation trails sometimes seen behind airplanes (Jaramillo 2022), which form in humid conditions when water vapor condenses around the tiny particles of pollutants emitted by airplane engines. By combining weather data, satellite data, and flight data, AI can predict when and where contrails are likely to form, enabling pilots to adjust the altitudes of their flights accordingly. American Airlines and Google Research demonstrated how airlines can reduce their climate impact: A program involving 70 flights over six months resulted in a 54% reduction in contrails compared to flights without AI-guided route adjustments (Elkin and Sanekommu 2023). At scale, such initiatives could have a meaningful, cumulative environmental impact. While these AI-based solutions do not require high amounts of energy, it is important to note that AI technologies, particularly generative AI, often consume significant amounts of energy. This is a distinct challenge that merits its own mitigation solutions, such as more energy efficient models and hardware. Helping Communities Adapt and Build Resilience AI is being utilized to help communities build resilience and better prepare for and respond to natural disasters. It can analyze vast datasets, including historical records and real-time weather and satellite data, to predict and track extreme weather events with greater accuracy than previously possible. Forecasting floods is one such example. Every year, floods disrupt the lives of millions worldwide, causing thousands of fatalities and significant financial damages. Google Research’s AI-powered forecasting system uses a state-of-the-art hydrological model to forecast the amount of water flowing in a river and is able to leverage global, publicly available data to generate predictions even in regions with limited historical data. This improves forecasts across regions in Africa and Asia to be similar to those that are currently available in Europe and provides coverage in 100 countries for 700 million people (Cohen 2024; Nearing et al. 2024), with flood forecasts distributed on Google Maps and Android devices. To maximize the value of the research, the data has been made available to researchers and partners through an application programming interface and through the Google Runoff Reanalysis & Reforecast dataset which includes forecasts dating back to 1981. Giving the ecosystem access to more and higher-quality data should enable the development of more effective mitigation strategies. As global temperatures rise, wildfires are also devastating communities around the world, as seen recently in Los Angeles. Since 2020, Google Research has employed AI models to track wildfire boundaries and has surfaced alerts providing helpful information to people who are caught near a wildfire (Royz and Tendler 2024). This is now available in over 20 countries. To further advance wildfire research, Google Research released FireBench, an open-source machine learning (ML) benchmark dataset. Plus, Google Research collaborated with the Earth Fire Alliance and the Moore Foundation to contribute to the creation of FireSat (Van Arsdale 2024), a purpose-built constellation of satellites designed specifically to detect and track wildfires as small as a classroom (roughly 5x5 meters; figure 1). With FireSat, authorities will have near real-time information about the location, size, and intensity of early-stage wildfires updated every 20 minutes so firefighters and emergency responders can respond effectively. FireSat’s data will also be used to create a global historical record of fire spread, helping scientists to better model and understand wildfire behavior. Advancing AI to Help More People Worldwide Live Healthier Lives Healthcare is another domain where AI has already started to deliver on its potential to improve lives. It is being used to expand access to quality care worldwide, provide customized information to people on their health journeys, generate insights for public health authorities, and empower clinicians to provide timely, accurate diagnoses to patients. Going forward, AI-powered technologies will allow experts to reimagine patient care and will increasingly be used by both patients and clinicians, helping people everywhere live longer and healthier lives. Improving Healthcare for Everyone Through AI Diagnostic Tools AI accelerates medical imaging and diagnostics by rapidly analyzing vast amounts of complex data and providing objective assessments, which can lead to more timely and effective interventions, and thus better health outcomes. Breast cancer, one of the most common cancers globally, is an example where early detection is possible through screenings and can lead to better chances of survival. Together with partners including iCAD (Corrado 2022), NHS, Imperial College London, and Northwestern Medicine, Google Research has developed and integrated AI models for mammography into breast cancer screening workflows to help radiologists identify breast cancer earlier and more consistently. Published research shows that this technology can identify signs of breast cancer as effectively as trained radiologists (McKinney et al. 2020). Google Research also developed ML tools for lung cancer screenings: The system analyzes CT scans and provides a cancer suspicion rating, helping radiologists identify potential malignancies more accurately (Kiraly and Pilgrim 2024). Study results indicate the potential for one person out of every 15–20 patients screened to be able to avoid unnecessary follow-up procedures, reducing their anxiety and the burden on the healthcare system (Kiraly et al. 2024). DeepHealth and Apollo Radiology International are exploring how to integrate this technology into clinical practice. In a world with a shortage of teachers and overcrowded classrooms, there is a huge opportunity for AI to make a difference in formal education settings, helping both educators and students. It is important to ensure that AI is also deployed in less economically developed countries, taking into account practical challenges such as unreliable internet connections, the availability of trained staff, and the need to operate within established frameworks for follow-up care. AI-powered diagnostic tools can play a crucial role in boosting access to care in low-resourced regions across the Global South. One example is screening for diabetic retinopathy, the leading preventable cause of blindness. Over 500 million adults worldwide have diabetes, but many do not have access to eye specialists and the regular screenings required for the timely detection and treatment of the disease. Back in 2015, Google Research sought to address this together with the Aravind Eye Hospital in India and Rajavithi Hospital in Thailand, along with the founder of Thailand’s national diabetic retinopathy screening program, Dr. Paisan Ruamviboonsuk. They pioneered Automated Retinal Disease Assessment, an AI-based diagnostic tool for high-quality diabetic retinopathy screening. Ophthalmologists helped researchers train the model, which can be operated by nurses and has screened over 600,000 patients to date. Forus Health, AuroLab, and Perceptra aim to reach six million people in India and Thailand over the next decade using this technology (Sawhney 2024). It is an example of how years of collaborative research can be piloted and then applied at scale in clinical settings to make quality care accessible to underserved populations. When developing AI applications for healthcare, it’s important to prioritize patient safety and privacy, and to deliver equitable results regardless of the socioeconomic status, race, gender, or other demographic factors of patients. That is why piloting research in real-world settings, under medical guidance from partners, and with people of diverse backgrounds matters. Frameworks can also help prevent biases in AI datasets and tools. Last year, Google Research introduced the Health Equity Assessment of ML performance (Schaekermann et al. 2024), a four-step process for quantitatively estimating the performance of ML tools for groups with, on average, worse health outcomes. This could inform improvements in product development and real-world testing to make health AI technologies more effective for more people. The framework was tested on the Skin Condition Image Network dataset, developed by Google Research and Stanford University. It contains over 10,000 user-contributed images of skin, nail, or hair conditions, and aims to provide a more representative dermatology dataset across diverse skin tones and body parts. The intention is that this dataset will be a useful resource for all those working to advance inclusive dermatology research, education, and AI tool development. It is important for such datasets to be diverse and accurately reflect the complexities and variations within the population, so that when AI algorithms are trained on the data, they are free from biases and can promote better health outcomes for all. Delivering Personalized Healthcare with Generative AI Arguably the most profound shift to date in AI research has been the recent and rapid rise of generative AI and its mainstream adoption across society. Now being applied to the medical domain, this technology is poised to enable highly personalized insights which could revolutionize the care patients receive and pave the way for more preventative healthcare solutions. Google Research is exploring the value of generative AI in healthcare to assist clinicians, researchers, and patients. Notably, Med-Gemini (Corrado and Barral 2024) is a family of AI models fine-tuned for medical applications. On the MedQA benchmark, which uses US Medical Licensing Exam-style questions, Med-Gemini achieved a state-of-the-art accuracy of 91.1%, surpassing previous models (Saab et al. 2024). It can interpret complex 3D scans, answer clinical questions, and generate state-of-the-art radiology reports (figure 2). These results are a testament to the advanced multimodal and reasoning capabilities of today’s generative AI and lay the foundation for more accurate and individualized medical care. For all the advances in AI technology, the physician-patient conversation is, and will remain, a cornerstone of medicine. Google Research is conducting research to inform how generative AI could be used to augment diagnostic medical reasoning and conversations between clinicians and patients. AMIE (Articulate Medical Intelligence Explorer; Karthikesalingam and Natarajan 2024) is a large language model (LLM)-based research system trained and evaluated along many dimensions that reflect quality in real-world clinical consultations from the perspectives of both clinicians and patients (Tu et al. 2024). In simulated consultations, AMIE performed at least as well as primary care physicians when evaluated on clinically meaningful aspects of consultation quality, and clinicians assisted by AMIE arrived at more comprehensive differential lists than those without AMIE assistance. Following these promising results, AMIE is being tested with healthcare organizations to see how it could support clinical conversations, with oversight from medical professionals. As a powerful tool for delivering insights, AI could also help personalize and improve overall wellness. Google Research is exploring this with projects such as the Personal Health Large Language Model (PH-LLM) and LLM agents that analyze wearable data from Fitbits and other devices (Merrill et al. 2024). The PH-LLM, a fine-tuned version of Google’s Gemini model, is designed to give personalized advice on sleep and fitness. It is trained to understand both written text and data from wearable sensors. In tests, PH-LLM’s fitness recommendations were as good as those from human experts. The model demonstrated a strong understanding of health topics, scoring 79% on sleep-related and 88% on fitness multiple-choice tests, outperforming average human scores (Cosentino et al. 2024). This research shows the significant progress and potential of AI models to provide unique health insights to users. Given the sensitivities around personalized healthcare and insights, it is important to undertake this research with strong attention to user privacy and safety. Empowering the Next Generation of Learners with AI Tools Grounded in Learning Science Learning is limitless. People learn as students, on the job, for a new career, and throughout their lives, pursuing hobbies and becoming curious about new things as their lives change and interests evolve. Generative AI is unlocking the potential within learners of all ages by democratizing access to information, personalizing experiences, and encouraging creativity. It now allows for engaging conversational experiences that foster understanding and knowledge acquisition. In collaboration with pedagogy experts, Google Research developed LearnLM, a family of AI models fine-tuned for education. LearnLM is grounded in educational research and tailored to how students learn, with the aim of creating more engaging, personalized, and effective learning experiences (Gomes 2024). In a recent technical report, LearnLM outperformed other leading AI models when it comes to adhering to the principles of learning science (Wiltberger 2024), such as explaining concepts at appropriate levels, providing effective guidance, interactively guiding learners, and encouraging active engagement. The technology is already integrated into existing platforms, including Google Search and YouTube, to enhance understanding rather than simply providing answers. For example, in Google Search, AI Overviews can be adjusted to simplify language or break down complex topics. Nothing can replace the magic of the teacher-student relationship. But in a world with a shortage of teachers and overcrowded classrooms, there is a huge opportunity for AI to make a difference in formal education settings, helping both educators and students. LearnLM is being used to create new educational tools, such as Illuminate, an experiment that breaks down research papers into audio conversations, and Learn About, an experiential tool that guides learners through various topics using different media and activities (figure 3). AI could also serve as a personal assistant to teachers, supporting them with tasks such as lesson planning, testing, and grading, enabling them to spend more time with their students. Google is piloting programs in Google Classroom to simplify lesson planning, for example, suggesting multimedia resources about a topic or allowing educators to adjust their content for different class levels. Given the access young students will have to these tools, it is vital to emphasize safety, data privacy, and age-appropriateness. Tech companies have a responsibility to consult with experts on child safety and development, partner with pedagogical experts, conduct rigorous testing, and provide resources for responsible AI use. Employing AI to Create a More Accessible World for Everyone Groups who may be at risk of being overlooked are sometimes the ones who stand to benefit the most from AI. AI-powered technologies can help create more inclusive, accessible, and personalized experiences for people with disabilities in their everyday lives. By better understanding user interactions and adapting interfaces in real time, AI can give users with disabilities access to information and communication tools so they can interact more easily with other people and with the digital world. For people who are d/Deaf or hard of hearing, AI can provide real-time transcriptions of speech and other sounds, making daily conversations and video and audio content more accessible. This technology is seen in products like Google’s Live Transcribe, which provides real-time speech-to-text transcriptions, and Live Caption, which automatically generates captions on videos and audio across Android devices and Chrome browsers. For people who are blind or low-vision, AI can read online content out loud, providing the most relevant information first and offering descriptions of images, making it more helpful than traditional screen readers. The newly released NotebookLM can convert written documents or long videos into an engaging conversation about the topic, which can benefit many users. AI can also help the hundreds of millions of people globally who have non-standard speech to be heard and understood. Project Relate, for example, is an Android app developed by Google Research that uses machine learning to understand an individual’s non-standard speech patterns. It enables users to communicate more easily through features such as real-time transcription, synthesized voice repetition, and direct interaction with Google Assistant. AI can also help people with conditions like amyotrophic lateral sclerosis (ALS) and similar motor impairments to communicate through eye movements. These users often rely on speech-generating technology, but it can be extremely slow and therefore limiting. Google Research partnered with Team Gleason, a non-profit organization that serves individuals with ALS, to create SpeakFaster (figure 4), a prototype that uses large language models to accelerate eye-gaze-based typing by predicting full phrases from abbreviated text and conversational context with high accuracy. Initial studies demonstrate that the system requires 57% fewer motor actions than traditional predictive keyboards in offline simulations, resulting in text-entry rates 29-60% faster than traditional baselines (Cai et. al. 2024). In addition, AI speech synthesis can be used to recreate voices for people with ALS, allowing them to communicate with friends and loved ones in a more authentic manner than standard, robotic-sounding voices of traditional solutions. Such technologies can help remove barriers to the world and meaningfully change lives. Removing language barriers can also help make the world more accessible. Much of the investment in language models to date has been in English. However, multilingual AI can help break down language barriers and help make online information and other AI tools accessible to users from all backgrounds. Google Assistant can now read web articles aloud in 42 languages and can translate the text when needed, and Google Translate recently expanded to include 110 new languages using AI, increasing access to information for over 614 million more people (Caswell 2024). It is also valuable to invest efforts in multilinguality so that language models are accurate and accessible in different languages and are relevant to people across different countries and cultures. AI Must Be Deployed Responsibly For all the recent progress, the use of AI to advance societal causes is still in the early stages. As the technology matures, society will benefit from a focus on safety for children, patients, and less advantaged populations, from an emphasis on both quality and equality, and from putting checks and regulations in place. If the risks and challenges can be properly addressed, there will be a plethora of opportunities to create useful, real-world applications that optimize classrooms, clinical settings, urban spaces, online access, and more. Google’s approach is grounded in AI principles that guide its efforts to develop the technology responsibly and collaboratively so that it works for everyone. Google advocates for bold innovations that can advance the frontier of AI and scientific research to benefit people everywhere, and the organization develops and deploys AI innovations responsibly, in a way that addresses user needs and rights such as safety, security, and privacy. That includes implementing appropriate due diligence and feedback mechanisms, aligning with international legal principles, and employing rigorous design, testing, and monitoring to mitigate any harmful outcomes and avoid unfair bias. This work is best done in collaboration with a wide range of partners, including researchers across industry and academia and with governments and civil society, to maximize the benefits of AI for all. AI, developed and deployed boldly, responsibly, and collaboratively, can help provide solutions to previously unsolvable problems. It is already having a profound impact, and as we look to the future, the potential of this technology to address societal challenges and improve people’s lives has never been clearer. Stakeholders from the public and private sectors have good reason to come together to realize the vast AI opportunity and establish partnerships to scale the technology safely in domains like healthcare and education. The authors of this article look forward to joining forces with others dedicated to leveraging AI to benefit humanity as we ask the biggest of questions and pioneer groundbreaking answers. We are committed to pursuing AI responsibly, and with the conviction and determination that it will be a powerful force for good. References Cai S, Venugopalan S, Seaver K, Xiao X, Tomanek K, Jalasutram S, Ringel Morris M, Kane S, Narayanan A, MacDonald RL, and 6 others. 2024. Using large language models to accelerate communication for eye gaze typing users with ALS. Nature Communications 15:9449. Caswell I. 2024. 110 new languages are coming to Google Translate. Google, The Keyword, June 27. Cohen D. 2024. An improved flood forecasting AI model, trained and evaluated globally. Google Research, Nov 11. Corrado G. 2022. Partnering with iCAD to improve breast cancer screening. Google, The Keyword, Nov 28. Corrado G, Barral J. 2024. Advancing medical AI with Med-Gemini. Google Research, May 15. Cosentino J, Belyaeva A, Liu X, Furlotte N, Yang Z, Lee C, Schenck E, Patel Y, Cui J, Douglas Schneider L, and 24 others. 2024. Towards a Personal Health Large Language Model. arXiv:2406.06474. Elkin C, Sanekommu D. 2023. How AI is helping airlines mitigate the climate impact of contrails. Google, The Keyword, Aug 8. Gomes B. 2024. How generative AI expands curiosity and understanding with LearnLM. Google, The Keyword, May 14. Jaramillo P, Ribeiro SK, Newman P, Dhar S, Diemuodeke OE, Kajino T, Lee DS, Nugroho SB, Ou X, Hammer A, and 2 others. 2022. Transport. In: IPCC, 2022: Climate Change 2022: Mitigation of Climate Change. Contribution of Working Group III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. Shukla PR, Skea J, Slade R, Al Khourdajie A, van Diemen R, McCollum D, Pathak M, Some S, Vyas P, Fradera R, Belkacemi M, Hasija A, Lisboa G, Luz S, Malley J, eds. Cambridge University Press. Karthikesalingam A, Natarajan V. 2024. AMIE: A research AI system for diagnostic medical reasoning and conversations. Google Research, Jan 12. Kiraly AP, Cunningham CA, Najafi R, Nabulsi Z, Yang J, Lau C, Ledsam JR, Ye W, Adrila D, McKinney SM, and 15 others. 2024. Assistive AI in lung cancer screening: A retrospective multinational study in the United States and Japan. Radiology: Artificial Intelligence 6(3):e230079. Kiraly A, Pilgrim R. 2024. Computer-aided diagnosis for lung cancer screening. Google Research, March 20. Kumar P, Goel A. 2016. Concentration dynamics of coarse and fine particulate matter at and around signalised traffic intersections. Environmental Science: Processes & Impacts 18:1220–35. McKinney SM, Sieniek M, Godbole V, Godwin J, Antropova N, Ashrafian H, Back T, Chesus M, Corrado GS, Darzi A, and 21 others. 2020. International evaluation of an AI system for breast cancer screening. Nature 577:89–94. Merrill MA, Paruchuri A, Rezaei N, Kovacs G, Perez J, Liu Y, Schenck E, Hammerquist N, Sunshine J, Tailor S, and 10 others. 2024. Transforming wearable data into health insights using large language model agents. arXiv:2406.06464. Nearing G, Cohen D, Dube V, Gauch M, Gilon O, Harrigan S, Hassidim A, Klotz D, Kratzert F, Metzger A, and 8 others. 2024. Global prediction of extreme floods in ungauged watersheds. Nature 627:559–63. Royz M, Tendler A. 2024. Wildfire boundary maps expand to new countries in Europe and Africa. Google, The Keyword, July 29. Saab K, Tu T, Weng WH, Tanno R, Stutz D, Wulczyn E, Zhang F, Strother T, Park C, Vedadi E, and 57 others. 2024. Capabilities of gemini models in medicine. arXiv:2404.18416. Sawhney R. 2024. How AI is making eyesight-saving care more accessible in resource-constrained settings. Google Research, Oct 17. Schaekermann M, Spitz T, Pyles M, Cole-Lewis H, Wulczyn E, Pfohl S, Martin Jr D, Jaroensri R, Keeling G, Liu Y, and 17 others. 2024. Health equity assessment of machine learning performance (HEAL): a framework and dermatology AI model case study. The Lancet eClinicalMedicine 70:102479. Tu T, Palepu A, Schaekermann M, Saab K, Freyberg J, Tanno R, Wang A, Li B, Amin M, Tomasev N, and 15 others., 2024. Towards conversational diagnostic AI. arXiv:2401.05654. Van Arsdale C. 2024. A breakthrough in wildfire detection: How a new constellation of satellites can detect smaller wildfires earlier. Google, The Keyword, Sept 16. Wiltberger S. 2024. LearnLM outperformed other AI models in a recent technical study. Google, The Keyword, Dec 19. About the Author:Yossi Matias is vice president at Google and the head of Google Research. Avinatan Hassidim is a vice president at Google Research. Philip Nelson is a director of engineering at Google Research.