.png){kind=small}
We asked 272 international experts which AI risks matter most, how severe they could become, and who is responsible for addressing them.
Under current trajectories, experts judged that 18 of 24 AI risk domains carry at least a 10% probability of catastrophic outcomes within the next five years.
Experts estimated the five most severe harms over the next 5 years were likely to come from dangerous capabilities, competitive dynamics, weapons & cyberattacks (including CBRNE), power centralization, and false information.
AI users and affected stakeholders bear most vulnerability; experts assign primary responsibility to frontier developers, governments, regulators, and standards bodies.
Information, finance, and national security face the highest vulnerability across all risk categories.
Experts point to three implications: risks this large demand stronger controls, voluntary developer action needs rules and enforcement, and some cross-border risks require international coordination and structural reform.
Expert views on the tractability of the most severe harms:
"Dangerous capabilities emerge as an emergent property of scaling AI systems, which makes them difficult to predict and control. Even with safety measures in place, the potential for catastrophic misuse remains."
"Cyberattacks from both state and non-state actors will be a permanent fixture of AI-related risks."
"Heavy data infrastructure and large models will continue to consume energy and cause resource pressure."
"Inequality is deeply entrenched in economic and social systems. AI may exacerbate existing inequalities through automation-driven job losses in certain sectors while creating wealth for those who own and control AI systems."
"Power centralization is perhaps the most stubbornly persistent risk because the same entities developing AI are often best positioned to capture its benefits, creating self-reinforcing dynamics that are difficult to reverse through technical interventions alone."
We elicited judgements from more than 270 international experts spanning all areas of AI risk, including:
Experts identified asymmetries in how vulnerabilityVulnerability had two components: exposure and sensitivity. We defined exposure as "the extent to which the actor's people, operations, and assets interact with or are dependent on AI systems", and sensitivity as "the extent to which an actor would be harmed if the hazard materialized. The harm may be direct or indirect". to AI risks and responsibilityResponsibility had 3 components: obligation, capability, and causal influence. We defined obligation as "the extent to which the actor should proactively lead or initiate efforts to address the risk"; capability as "the extent to which the actor has specialized skills and resources needed to address the risk"; and causal influence as "the extent to which the actor causes or contributes to the harms resulting from the risk materializing". for addressing AI risks are distributed across the AI ecosystem.
According to experts, general-purpose AI developers and governance actors such as governments, regulators, and standards bodies hold primary responsibility for addressing AI risks.
In contrast, AI system users and affected stakeholders such as members of the public are most vulnerable to AI risks.
This mismatch means that those who are most responsible for addressing AI risks are not those who are most vulnerable, leading to misaligned incentives in addressing the most important AI risks.
Expert views on responsibility and vulnerability:
"Affected stakeholders lack both the agency and systemic leverage to mitigate risk, and assigning responsibility to them risks reinforcing harm by misplacing accountability."
"I rated AI Developers and Specialized AI Developers as primarily responsible because they decide the model design, structure, critical parameters (such as default setting) which directly impact the outputs and systemic risks."
"I cannot overstate how much more responsible the 'upstream' actors are for limiting these issues. That's where the emphasis should be. I see a direct analogy to social media. Yes, individuals are responsible for sharing misinfo. But the platforms should bear the brunt of our concern about the issue. They are the best point of intervention."
"I think governance actors have extreme responsibility for managing these problems, even more than private companies. It ultimately falls on them to protect the public and enact regulations and enforcement."
"Governance actors are the primary bridge between the affected stakeholder and the AI developer and deployer. Until incentives are aligned to the public interest, governance actors are essential to broker the relationship between AI developers and affected stakeholders."
This divergence between vulnerability and responsibility is standard in risk governance: the public is most vulnerable to aviation failures, pharmaceutical side-effects, nuclear meltdowns, and environmental contamination, but engineers, manufacturers, and regulators bear primary responsibility for prevention. In other safety-critical industries, the gap is bridged by a combination of mandatory standards, enforcement, liability regimes, and a societal expectation of low risk tolerance. But comparable mechanisms for AI are nascent or absent. Without them, vulnerable parties have little recourse and responsible parties face little pressure to act.
Experts assessed how vulnerable each of 14 industry sectors is to each of the 24 AI risks. Sectors were adapted from the North American Industry Classification System (NAICS), including Information (e.g., publishing, motion pictures and sound, broadcasting, telecommunications, data processing), Finance and Insurance, Health Care, National Security, Education, and others. The accompanying figure displays sector vulnerability assessments, with sectors ordered from most to least vulnerable based on weighted mean vulnerability scores across all risks.
Experts rated the Information and National Security sectors as most vulnerable across AI risks, with consensus on extreme vulnerability (median 4 to 5) to content-related harms like disinformation and influence and loss of privacy, and to dangerous capabilities and weapons and cyberattacks respectively. For example, experts argued:
"The most vulnerable sectors are those where AI is deeply embedded in critical decision-making and where failures have immediate, large-scale consequences—information, national security, and finance all share these characteristics."
"National security faces compound vulnerability—both as a target for AI-enabled attacks and as a domain where AI failures could trigger cascading geopolitical consequences."
Finance and Insurance were judged to have similarly high vulnerability to fraud and scams, AI security vulnerabilities, and AI system safety failures, reflecting:
"Both direct attack vectors (fraud, market manipulation) and regulatory exposure from AI system failures."
Health Care received high vulnerability ratings (median 4) for loss of privacy, discrimination, and overreliance and unsafe use, given that medical AI failures carry immediate human costs. By contrast, sectors with lower AI penetration, including accommodation and food services, agriculture and manufacturing, and arts and entertainment, received lower vulnerability ratings (median 2 to 3), though they remain exposed to broad socioeconomic effects like discriminatory outputs and job displacement.
Expert consensus was strongest at the extremes: near-universal agreement on high-vulnerability sectors like Information and National Security, and on low-vulnerability sectors like Accommodation, while moderate-vulnerability sectors showed more disagreement, reflecting uncertainty about how AI adoption will affect these industries.
Rather than suggesting that each sector faces a unique set of AI risks, the findings suggest that many sectors are exposed to overlapping risks as AI becomes embedded in decision-critical systems.
Under most established risk-governance frameworks, a 10% probability of catastrophic outcome over five years would be considered 'intolerable', triggering mandatory mitigation requirements. The size of these risks, even under a scenario where pragmatic mitigations are implemented, suggests significantly more action is required.
Each AI risk affects different actors and sectors in different ways. Browse the individual pages to see how experts assessed vulnerability and responsibility for each.
This study uses specific definitions for the risks, sectors, and actors discussed throughout. Expand any term below for its full definition.
How we grouped the parts of the economy experts assessed.
The roles in the AI ecosystem experts assigned responsibility to. Drawn from US and Australian government standards, refined through team discussion and pilot testing.
The 24 risks experts assessed, grouped into seven domains. Definitions from the AI Risk Repository domain taxonomy (Slattery et al.), reproduced with permission.
This work was supported by Commonwealth Bank of Australia, who reviewed the design, but did not influence the collection, analysis, interpretation or reporting of the data.
Many of our authors participated as experts in the Delphi study. In addition, we wish to acknowledge the contributions of the following experts who also generously gave their time to participate: Dimitris Alchatzidis, Lauriane Aufrant, Fazl Barez, Catherine Barrett, Matthew Bedsole, Bill Black, Patrick Butlin, Stephen Cave, Uzma Chaudhry, April Chin, Austin Crumpton, Robert Cunningham, Rozita Dara, Rajiv Dattani, Daniela Elia, Simon Goldstein, Elena Gurevich, Timo K Harakka, Possum Michael Hodgkin, Michael J Howell, Hernan Huwyler, Strahinja Janjusevic, Minseok Jung, Uma Kalkar, Raz Karmi, Paul Kedrosky, Cameron F Kerry, Ansgar R Koene, Noam Kolt, Eyup Engin Kucuk, Elyn Yun Ling Lee, Fion Lee-Madan, Lynna Leong, Jodie Levy, Markus P. Luchsinger, Matt MacDermott, Alexander Meinke, Layne L Morrison, Bronte Pendergast, Daniel J Ragsdale, Pablo Rice, Bram Rijsbosch, Anka Reuel, Paul Röttger, Graham H. Ryan, Samuel T Segun, Aithan Shapira, Lisa Soder, Puneet Sondh, Lucas G. Uberti-Bona Marin, Luis E Urtubey, Jimena Sofia Viveros Alvarez, Ben Wilkinson, James Williams, and all experts who chose to remain anonymous.
We’d also like to acknowledge Spencer Michaels for support in creating this page and visualizations as a Cambridge Boston Alignment Initiative Fellow.
Alexander K Saeri, Jess Graham, Michael Noetel*, Peter Slattery, Dennis Ah-king, Edla Aittokallio, Ibitola Akindehin, Abbas Al Mahdi, Elie Alhajjar, Rafael Andersson Lipcsey, Gary Ang, Catherine M Azam, Amos Azaria, Rishal Balkissoon, Isabel Barberá, Claudio Bareato, Jonathan Barry, Michael Basehart, Andrew M Bean, Danny Belitz, Samantha Augusta Bennett, Kayla Blomquist, Damian Borstel, Ben Bucknall, Tomas Bueno Momcilovic, Aurelie Bugeau, Nicholas Caputo, Stephen Casper, Gulam Chagani, Ze Shen Chin, Jiyeon Cho, Jay Chooi, Joel N Christoph, Dmytro Chumachenko, Kieran Conboy, Elizabeth M Daly, Tom David, Paul de Font-Reaulx, Antonio De Santis, Fabrizio Degni, Christopher W DiCarlo, Yawen Duan, Janet Egan, Ian W Eisenberg, Sherif M Elsafty, Adam Ennamli, Mark Esposito, Nicola Fabiano, Gallo Fall, Neil R Fernandes, Pip Foweraker, Chiara Gallese, Sandra Galletti, Andrew Gamino-Cheong, Rokas Gipiškis, Gwyn Glasser, Delaram Golpayegani, Jeff Grayson, Hans Gundlach, Josiah Hagen, Alexander Hagenah, Amelia S Haines, The Anh Han, Yixiong Hao, Kasii Harris, Tianxing He, Koen Holtman, Giorgos Iacovides, Kenneth L Ingham, Krystal Jackson, Adam Jones, Himanshu Joshi, Brian Judge, Arturs Kanepajs, Shreya Kapoor, Win Myat Nwe Khine, Aidan Kierans, Aleksandra Korolova, Markus Krebsz, Nicholas Kruus, Joe Kwon, Valeria Lazzaroli, Ray X Lee, Evelina Leivada, Stephan Lewandowsky, Michael B Li, Xiaojian Li, Geunsik Lim, Henrique Lisakowski, Fabio Lonardoni, Todd C Lowe, Jackson G Lu, Alexander Lyzhov, Nada Madkour, Parv Mahajan, David Manheim, Kareem Mathias, Claudio Mayrink Verdun, Sean McGregor, Scott McLean, Matthew J McMahon, Minas Megalokonomos, Nicolas Moës, Fernando Mourao, Yaroslav Mukhin, Malcolm Murray, Simon Mylius, Neeraj Nagpal, Koichi Nakada, Anna Neumann, Jessica Newman, Kwan Yee Ng, Minh N Nguyen, Quynh Phuong Nguyen, Seán S Ó hÉigeartaigh, Daria Onitiu, Kelly Onu, Oscar Oviedo-Trespalacios, Ugur Ozer, Chanwoo Park, M. Alejandra Parra-Orlandoni, Patricia Paskov, Anna M Pastwa, Burak Piskin, Jacob Pratt, Claudiu A Predincea, Marjana Prifti Skenduli, Kenneth Priore, Mukunda Madhab Pujari, Zhenting Qi, Preethi Raghunathan, Robi Rahman, Deepika Raman, Max Reddel, Jyoti Ruparel, Emma B Ruttkamp-Bloem, Tiffany Saade, Greg Sadler, Said Saillant, Paul M Salmon, Ayrton San Joaquin, Lama Saouma, Maziya Sarangpurwala, Supheakmungkol Sarin, Daniel S Schiff, Anna D Schilling, Chris Schmitz, Reva Schwartz, Abeer Sharma, Tianhao Shen, Kehan Sheng, Maury D Shenk, Eli Sherman, Chandler Smith, Julie M Smith, Estevenson Solano, Oliver Sourbut, Madhulika Srikumar, Ryan Stendall, Jakob Stenseke, Michael Stern, Joshua Sternfeld, Nikko Stevens, Ilia Sucholutsky, Yuanyuan Sun, Mariami Tkeshelashvili, Cristian Trout, Brian Tse, Nikolaos Tsinganos, Michelle Vaccaro, Anthony R Valiaveedu, Ramakrishnan Veeramony, Jeremy Verdo, Pulkit Verma, Andrea Luigi Vitali, Jinge Wang, JR Washebek, Yonah Welker, George F Westerman, Tristan Williams, Rongwu Xu, Mick Yang, Xuemeng Yang, Sander Zeijlemaker, Jingyu Zhang, Marta Ziosi, Neil Thompson
For full author affiliations and ORCIDs, see the arXiv paper.
We work with governments, companies, and researchers to build better tools for understanding AI risk. Subscribe to our newsletter for updates. Get in touch to explore partnership opportunities.
