Insights · June 4th, 2026
In June 2026, a research team led by MIT FutureTech and the University of Queensland published one of the most systematic attempts yet to rank the dangers of artificial intelligence. Using the Delphi method—three rounds of anonymous, iterative consultation designed to surface genuine consensus and expose where it breaks down—they asked 272 AI experts from 37 countries to judge 24 distinct AI risks across severity, who is most exposed, and who is most responsible for fixing them.
For leaders trying to separate signal from noise in the AI conversation, the results are a rare piece of structured guidance.
What are the risks?
The study drew its risk list from the MIT AI Risk Repository, which distilled more than 1,700 documented risks from 74 frameworks into 24 subdomains across seven families.
The breadth is the point: AI risk is not one thing, and treating it as a single issue obscures where the real exposure lies. The seven families, with their component risks, are:
- Discrimination & toxicity: unfair discrimination and misrepresentation; exposure to toxic content; unequal performance across groups.
- Privacy & security: loss of privacy through leaked or inferred personal data; AI security vulnerabilities and attacks.
- Misinformation: false or misleading information; pollution of the information ecosystem and loss of a shared sense of reality.
- Malicious actors & misuse: disinformation, surveillance, and influence at scale; AI-enabled weapons and cyberattacks (including CBRNE); fraud, scams, and targeted manipulation.
- Human–computer interaction: overreliance and unsafe use; loss of human agency and autonomy.
- Socioeconomic & environmental harm: power centralization; inequality and declining employment quality; devaluation of human creativity; competitive “race” dynamics; governance failure; environmental harm.
- AI system safety, failures & limitations: AI misalignment; dangerous capabilities; lack of capability or robustness; lack of transparency or interpretability; AI welfare; multi-agent risks.
This map is useful on its own: it lets a leadership team locate which handful of these 24 actually intersect their business rather than reacting to whichever risk is dominating headlines.
Summary of the Key Findings and Discussion Points
Experts rated each risk under two scenarios: “business as usual,” where organizations and governments carry on without new AI-specific mitigations, and “pragmatic mitigations,” where they make cost-effective, realistic efforts to manage the risks.
The headline number is sobering. Under business as usual, experts judged that 18 of the 24 risks carry at least a 10% probability of a catastrophic outcome over the next five years (2025–2030)—where catastrophic was defined as more than one million deaths, more than USD $100 billion in financial loss, or civilization-scale intangible harm such as democratic collapse.
The five highest-severity risks were dangerous AI capabilities, competitive dynamics (the race to ship), AI-enabled weapons and cyberattacks, power centralization, and false information. Encouragingly, experts believed pragmatic mitigations would meaningfully reduce severity across all 24 risks. Less encouragingly, five risks stubbornly remained above the 10% catastrophic threshold even with sensible mitigations in place: dangerous capabilities, weapons and cyberattacks, environmental harm, inequality and unemployment, and power centralization. And every one of the 24 risks stayed above a 5% probability of catastrophic harm.
Two further findings deserve executive attention. First, the experts identified a structural mismatch between vulnerability and responsibility.
The actors judged most vulnerable to AI harms were ordinary AI users and affected stakeholders—the public, consumers, and employees subject to AI decisions. The actors judged most responsible for addressing those harms were general-purpose AI developers and governance bodies (governments, regulators, standards organizations). Those who bear the risk are not those positioned to mitigate it—a textbook moral hazard.
Second, when experts named their top concerns, the most-cited were weapons and cyberattacks (27%), power centralization (24%), disinformation and influence (22%), loss of a shared sense of reality (22%), and dangerous capabilities (22%).
The discussion frames the persistent risks as collective-action problems. Any single developer who slows down to invest in safety bears the competitive cost alone, while the benefits accrue to society—so market forces push against precaution. The authors also warn of an “accountability sink,” where responsibility spread across many actors becomes responsibility held by none. (One caveat from the study: AI welfare was measured but kept out of the main severity charts, because the framework centered on human and institutional harm; its low score shouldn’t be read as a verdict that the risk is trivial.)
Why It Matters
It is tempting to file expert risk surveys under “interesting but abstract.” This one resists framing for three reasons.
First, the magnitudes cross governance thresholds that businesses already recognize. The authors note that under most established risk-governance frameworks—the kind used in dam safety, aviation, and pharmaceuticals—a 10% probability of catastrophic outcome over five years would be classified as intolerable and would automatically trigger mandatory mitigation. AI is being deployed at scale while sitting well above thresholds that no other safety-critical industry would accept.
Second, the study pinpoints where harm lands. Across nearly all risks, three sectors were judged most vulnerable: information, finance and insurance, and national security. These are not peripheral industries—they are the connective tissue of the modern economy, and they are precisely where AI is being embedded into high-stakes, fast-moving decisions. If your organization touches data, money, or critical infrastructure, you are operating in the blast radius the experts identified.
Third, the vulnerability-responsibility gap is a strategic warning, not just an ethical one. In aviation, pharmaceuticals, and nuclear power, the gap between who is exposed and who is accountable is bridged by mandatory standards, liability regimes, and insurance. For AI, the authors observe, those mechanisms are “nascent or absent.” That vacuum will not last. Liability rules, transparency requirements, and mandatory insurance are the predictable policy responses—and they will reshape the cost structure of building and deploying AI.
What This Means for CEOs
For executives, the study converts a sprawling debate into a manageable set of priorities. A few practical takeaways follow.
Triage your AI risk exposure deliberately. The lesson of a 24-item risk list is not that everything is dangerous—it is that resources are finite and severity varies. Use the seven-family map above to identify which risks your business actually touches. A financial-services firm should weight fraud, security vulnerabilities, and systemic model failures heavily; an information or media business should focus on disinformation and erosion of consensus reality.
Treat AI safety as risk management, not public relations. The “intolerable” framing matters internally. Boards already understand catastrophic-tail risk in financial and operational terms. Positioning AI exposure in that familiar language—probability-weighted impact, mitigation in proportion to expected harm—will get more traction than abstract appeals to responsible AI.
Build defense in depth rather than relying on your vendor. Experts placed primary responsibility on foundation-model developers, but they also assigned meaningful responsibility to deployers, infrastructure providers, and users. Distributed responsibility can be a strength when each layer adds a real safeguard—or a weakness when each assumes someone else is handling it. Don’t let your organization become part of an accountability sink. Assume the upstream model is imperfect and add your own monitoring, controls, and human oversight.
Anticipate the regulatory turn. The mechanisms the experts expect—liability, mandatory insurance, transparency and disclosure rules—are coming, and the most exposed sectors will likely see them first. Firms that build documentation, audit trails, and incident-response capability now will absorb new requirements far more cheaply than those scrambling later.
Resist the race when it compromises safety. The competitive-dynamics finding is a direct message to leadership: the pressure to ship fast is itself one of the highest-rated risks. The strategic edge over the next five years may belong less to whoever moves fastest than to whoever earns durable trust in sectors where AI failure is catastrophic.
The experts’ closing note is worth repeating to any leadership team: the window for avoiding the worst outcomes remains open, but it is narrowing.
Source: Prioritization of Risks from Artificial Intelligence A Delphi Study of 272 International Experts – read here.
Read other articles in the series:
The CEO’s guide to AI: The Pope Has an AI Strategy Memo for Your Boardroom
The CEO’s guide to AI: State Media Control Influences Large Language Models
The CEO’s guide to AI: Social engineering is turbocharged by AI
The CEO’s guide to AI: Black-hat LLMs and Cyber-threats
The CEO’s guide to AI: We’re nowhere near AGI
The CEO’s guide to AI: Young workers as the canaries in the coalmine
The CEO’s guide to AI: The risk of vibe coding complacency
The CEO’s guide to AI: The Case for Superhuman Adaptable Intelligence (SAI)
About Nikolas Badminton
Nikolas Badminton is the Chief Futurist & Hope Engineer at futurist.com. He’s a world-renowned futurist keynote speaker, consultant, author, media producer, and executive advisor that has spoken to, and worked with, over 500 of the world’s most impactful organizations and governments.
Nikolas is an artificial intelligence expert and his 2026 keynote ‘The AI Leader: Create Incredible Productivity, Profit & Growth’ is the level up for the modern CEO and executive leader.
Please contact futurist speaker and consultant Nikolas Badminton to discuss your engagement.
