MITRisks Delphi 2026
The paper reports a three‑round Delphi study (late 2025) that gathered quantitative and qualitative judgments from 272 AI experts worldwide to systematically rank 24 AI‑risk domains. Participants evaluated each risk on severity, likelihood, sector/actor vulnerability, and responsibility of different stakeholder groups (AI developers, deployers, infrastructure providers, governance actors, users, and affected stakeholders). The study finds that under a business‑as‑usual (BaU) scenario, 18 of the 24 risks have >10 % probability of catastrophic outcomes (2025‑2030) and all 24 exceed 5 % probability. The highest‑severity risks are dangerous capabilities, competitive dynamics, AI‑enabled weapons and cyber‑attacks, power centralisation, and large‑scale disinformation. Even with pragmatic mitigation measures (cost‑effective, industry‑government actions), several risks (dangerous capabilities, weapons, environmental harm, inequality, power centralisation) retain >10 % catastrophic probability. The paper also maps responsibility, showing AI deployers as primary for many operational risks, general‑purpose developers for foundational design risks, and governance actors for systemic oversight. Vulnerability assessments highlight that users, affected stakeholders, and sectors such as public administration, civil society, and accommodation/food services are extremely vulnerable to disinformation, cyber‑attacks, and fraud. The manuscript includes extensive expert commentary on each risk, illustrating concerns about AI‑driven polluting information ecosystems, AI‑enabled fraud, over‑reliance, loss of human agency, multi‑agent dynamics, environmental impact, and AI welfare. The authors conclude that without coordinated, robust governance and mitigation, AI risks pose a non‑trivial chance of severe or catastrophic harm within the next five years.
Topics
Delphi Methodology and Expert Sample
Details the three‑round Delphi design, recruitment of 272 AI experts from diverse regions and sectors, and the quantitative scoring framework used to assess severity, likelihood, vulnerability, and stakeholder responsibility for each AI‑risk domain.
High‑Impact Technical Risks (Dangerous Capabilities, Weaponisation, Misalignment, Multi‑Agent Dynamics)
Combines expert assessments of risks stemming from advanced AI capabilities that could be weaponised, lead to misaligned goal pursuit, or generate emergent multi‑agent failures. Includes severity ratings, infrastructure vulnerability, and primary responsibility attributed to general‑purpose developers and AI deployers.
Stakeholder Responsibility and Vulnerability Mapping
Provides a cross‑risk synthesis of which stakeholder groups (developers, deployers, infrastructure providers, governance actors, users, affected parties) bear primary, secondary, or minimal responsibility, and identifies sectors with extreme vulnerability to each risk domain.