Wave 9: Risks

The following report summarizes responses from 194 experts, in addition to 53 superforecasters and 612 members of the public collected between May 19, 2026 and Jun 10, 2026. Within expert respondents, 44 computer scientists, 37 industry professionals, 45 economists, and 68 research staff at policy think tanks participated.

Our wider website contains more information about LEAP, our Panel, and our Methodology, as well as reports from other waves.

Scenarios

Throughout this wave's survey, we ask respondents to consider three scenarios for AI progress by 2030 that we call "slow", "moderate", and "rapid".

In Wave 1 of LEAP, we asked respondents to forecast the percentage of LEAP panelists who would select each of these scenarios as best representing reality in 2030. In that wave, the mean aggregate forecasts by group (and the standard deviation of responses) are displayed in the table below (source).

Insights

1. Experts and superforecasters expect that a major AI-driven harm event occurring by 2050 is more likely than not, and forecast such an event happening by the mid-2030s.

   When asked about the probability that an event primarily driven by at least one AI system causes at least 50 deaths or $100 billion (in 2025 USD) in damages by the end of 2050, the median expert gave this a 62% probability and the median superforecaster 70%, which are far above the public's 35% forecast. Conditional on such an event occurring, both experts and superforecasters gave 50% odds that the first event arrives by 2035 (the public is slightly later, at 2037). Most respondents anchored on how low the threshold is, describing a qualifying event as “a single airplane” or “one major hurricane,” and reasoned that as AI is built into critical infrastructure, transport, and military systems, a qualifying incident becomes nearly inevitable. The most commonly cited events are AI-introduced errors in critical infrastructure systems and cyberattacks, followed by autonomous or AI-assisted military systems. Below these, transportation accidents and deliberate human weaponization of AI form a secondary tier, followed by a long tail of lower-frequency causes including healthcare failures, industrial accidents, and biological threats.

2. If AI capabilities progress rapidly, experts forecast roughly a 5x increase in the probability of a global AI-related catastrophe over their forecasts for a world where AI capabilities progress slowly.

   In the survey, we defined a global AI-related catastrophe as “a catastrophic event caused primarily by artificial intelligence systems in which more than 10% of the people alive at the start of a five-year period die by the end of that period.” If AI capabilities progress slowly, the median expert assigns this a 0.08% probability by 2030, 1% by 2050, and 2% by 2100; under a scenario where AI progress by 2030 is rapid, those forecasts rise to 1%, 5%, and 10%. The same pattern holds across other groups: by 2100, the median public forecast roughly triples from 5% to 15% and the median superforecaster more than doubles from 2% to 5%. For experts, this implies that roughly 80% of the catastrophe probability in a rapid-progress world would be absent if AI reached only the slow scenario's level of capacity by 2030 (see above for progress scenario details). In their written rationales, both optimistic and pessimistic forecasters agree that faster progress raises risk because it compresses the time available for policy and society to adapt. However, disagreement exists regarding how likely reaching this 10% threshold is: optimists treat ~800 million deaths as a near-disqualifying bar (noting that all of WWII killed roughly 3% of the world's population) and expect warning events to trigger protective mandates before the threshold is approached, while pessimists emphasize competitive AI race dynamics, the democratization of bioweapon capability, and the failure of institutional safeguards as plausible pathways to catastrophe.

3. If AI capabilities advance rapidly, experts attribute roughly two-thirds of all global catastrophic risk this century to AI.

   We also asked respondents for their total probability of a global catastrophe from any cause which is similarly defined as the death of more than 10% of the global population over a five-year period. We define an AI-related catastrophe to cover cases where AI is the direct or proximate cause of the deaths, or events that would not have occurred or would have counterfactually been extremely unlikely to occur “but for” the substantial involvement of AI within one year prior to the event (see the resolution criteria below for more detail). Under the condition AI capabilities progress rapidly, the median expert forecasts the probability of an AI-caused catastrophe by 2100 at 10%, compared to 15% for catastrophe from any cause. Computing each respondent's individual ratio of AI catastrophic risk to total catastrophic risk, the median expert attributes about two-thirds (67%) of total global catastrophic risk to AI in a world where AI capabilities progress rapidly. The share of total catastrophic risk that experts attribute to AI appears to rise with the level of AI capabilities: by 2100, it is roughly 30% under slow progress, 53% under moderate progress, and 67% under rapid progress. In their written rationales, some respondents explained why they perceived AI-related catastrophic risk to comprise a higher or lower percentage total catastrophic risk. Those who perceived a higher percentage argued that AI is becoming so entangled with every catastrophic pathway that a genuinely AI-free catastrophe is hard to imagine. For example, one respondent noted “there just aren’t that many risk factors that don’t at least indirectly flow through AI,” citing pandemic prevention, asteroid detection, and nuclear command and control. A few in this camp added that AI may actively reduce non-AI risk: “AI I think can actually make us safer with time, for example, by monitoring threats from out[er] space, or by helping analyze new viruses and craft vaccines.” Those who saw AI catastrophic risk as a smaller percentage rejected the premise that AI is entangled with everything, treating the dominant non-AI pathways as substantial standalone risks: “Pandemic risks and world war top the risks, followed by risks from climate change… I consider AI to be a small component of overall threats to the world,” wrote one, and another observed that “unlike the AI question, baseline risk here is already non-negligible without AI—pandemics, nuclear war, [and] climate collapse are all live on a 75-year horizon.”

4. Experts and superforecasters give a 50% chance that a major government restricts an AI release by 2030.

   The median expert and superforecaster each assign a 50% chance that the U.S., U.K., or E.U. will issue a binding directive, regulation, court order, or injunction to delay, seriously restrict, or condition the initial public release of an AI system on safety grounds by 2030, with the median expert giving 25% odds by 2028 and 75% by 2034. Forecasters divided over the path to resolution. On June 12, shortly after the survey closed, the U.S. federal government issued an export control directive suspending all access to Anthropic’s Mythos 5 and Fable 5 models by any foreign national. However, given these models had already been made publicly available, this directive does not meet our resolution criteria that such a direction must restrict the system’s release prior to its availability to the general public because of safety or national security risks. We will monitor ongoing news surrounding GPT-5.6 and related models to see if near-term government actions meet the criteria. Early-resolution respondents read Anthropic's voluntary restriction of its Mythos model as evidence the dangerous-capability threshold has arrived, and pointed to the EU AI Act's Article 93 powers which take effect in August 2026 as the earliest plausible trigger. Late-resolution respondents argued that voluntary industry restraint substitutes for state action and that governments typically regulate deployment and use rather than pre-release publication. Most across both early- and late-resolution respondents agreed that the likeliest catalyst is a galvanizing safety incident.

5. If AI capabilities progress rapidly, experts forecast a roughly 25% chance that two more of the current eight largest democracies backslide into authoritarianism by 2040.

   Two of the world's ten largest economies, China and Russia, currently score below 4 (“authoritarian”) on the EIU Democracy Index, and the median expert expects that count to hold at roughly two through 2040, essentially flat across every date and scenario. However, if AI progress is rapid, experts give a roughly 25% chance that two of today's democracies backslide, bringing the total to four or more authoritarian states by 2040 (out of the top ten economies), and a roughly 5% chance that the total reaches five or more. Those assigning higher counts argued that rapid AI progress amplifies the tools of surveillance, censorship, and repression and widens the range of political outcomes, disproportionately adding probability to less democratic results. India was a frequently cited candidate to cross the line, with the U.S. and U.K. raised to a lesser extent.

6. Experts predict that losses to cybercrime by 2031 would be roughly 52% higher in a world where AI capabilities progress rapidly than in a world where AI progresses slowly.

   If AI capabilities progress slowly, the median expert predicts that cybercrime losses reported to the FBI will increase from a baseline of $21 billion in 2025 to $25 billion in 2026, $35 billion in 2028, and $46 billion by 2031. If AI progress is rapid by 2030, they predict reported losses of $31 billion, $50 billion, and $70 billion, respectively. In a rapid progress world, experts give a one-in-four chance that losses are $93 billion or higher by 2031. Many respondents reported that AI-enabled deception would be a primary driver of future losses, citing deepfakes, automated targeting, and personalized messages as factors that would increase the rate of fraud. Forecasters who predicted high cybercrime losses argued that AI advantages attackers, with defenders structurally disadvantaged, with legacy systems and human behavior remaining highly vulnerable to attacks.

Questions

• AI Model Governance: By what calendar year will the United States (federal government), United Kingdom (central government), or European Union (supranational, rather than individual member-states) have issued a binding directive, regulation, court order or injunction to delay, seriously restrict, or condition the public release of an AI system(s), based on safety risks? ⬇️

• Financial Losses from Cybercrime: What will be the total financial losses (in billion 2025 USD) reported in the FBI Internet Crime Complaint Center (IC3) annual report for incidents reported during the following resolution years? ⬇️

• AI Catastrophic Risk: What is the probability that a global AI-related catastrophe will occur by each of the following resolution years? ⬇️

• Total Catastrophic Risk: What is the probability that any global catastrophe will occur by each of the following resolution years? ⬇️

• First Major AI Global Harm Event: What is the probability that an event, primarily driven by at least one AI system, will cause at least 50 deaths or $100 billion (in 2025 USD) in damages by the end of 2050? If such an event occurs by the end of 2050, in what calendar year will the first qualifying event occur? ⬇️

• Democratic Backsliding and Power Concentration: How many of the ten largest economies by GDP (as of 2025) will have an EIU Democracy Index score of below 4 (“authoritarian”) at the end of the following resolution years? ⬇️

For full question details and resolution criteria, see below.

Results

In this section, we present each question, and summarize the forecasts made and the reasoning underlying those forecasts. More concretely, we present background material, historical baselines, and resolution criteria; graphs, results summaries, results tables; as well as rationale analyses and rationale examples. We analyse these rationales alongside predictions to provide significantly more context on why experts believe what they believe, and the drivers of disagreement, than the forecasts alone.

AI Model Governance

Question. By what calendar year will the United States (federal government), United Kingdom (central government), or European Union (supranational, rather than individual member-states) have issued a binding directive, regulation, court order or injunction to delay, seriously restrict, or condition the public release of an AI system(s), based on safety risks?

Results. Experts and superforecasters give closely similar forecasts for when a qualifying government intervention will first occur. The median expert and superforecaster each assign a 50% chance of resolution by 2030, while the public is a little later at 2032. The median expert gives a 25% chance of resolution by 2028 and a 75% chance by 2034, indicating a meaningful probability that such an intervention occurs within the next decade. Experts and superforecasters do not differ significantly across any elicited percentile. The public, by contrast, is later than experts: expert-public differences are statistically significant at the 5th, 25th, 50th, and 75th percentiles, while the groups are much closer at the 95th percentile.

On June 12, shortly after the survey closed, the U.S. federal government issued an export control directive suspending all access to Anthropic’s Mythos 5 and Fable 5 models by any foreign national. However, given these models had already been made publicly available, this directive does not meet our resolution criteria that such a direction must “delay, seriously restrict or condition the release of an AI system prior to its availability to the general public, citing safety or national security risks.”

Rationale analysis

• The Mythos precedent: Both early- and late-resolution respondents cite Anthropic’s voluntary restriction of Claude Mythos Preview but often draw opposite conclusions from the incident. Early-resolution respondents tend to read it as evidence the dangerous-capability threshold has arrived and that involuntary restrictions will soon follow: “Mythos just gave national governments a reason to take action this year.” Whereas many late-resolution respondents read the same episode as proof that voluntary restraint is an adequate substitute for state action, with one writing that “AI companies are likely to cooperate in order to avoid the precedent of binding restrictions,” and another that “there won't be any need as AI companies will decide not to deploy extremely powerful models just for mere economic reasons.”
• The EU AI Act and its August 2026 enforcement powers: A frequent dividing line is whether the EU’s existing machinery is a ready vehicle for a qualifying action. Early-resolution respondents often treat it as such: “The EU AI Office gains explicit legal authority under Article 93 of the AI Act to restrict market availability of models posing systemic risks starting August 2, 2026—the earliest plausible trigger.” Some late-resolution respondents, however, argue the AI Act isn’t a good vehicle for imposing a pre-release restriction: “Even the AI Act doesn't actually give this power yet.”; “The question is whether [the AI Act] ever materializes as a halt, rather than a compliance demand.”
• Inciting incident: Many respondents, representing both poles, agree a notable safety incident is a likely path to resolution, with several forecasting that an incident involving a released model could result in restrictions on future models of the same class. But differences emerge on whether such an incident is imminent: “Once one major AI incident or near-miss occurs, at least one of these three jurisdictions will act and I expect this to happen within roughly 2 years,” writes an early-resolution respondent, while a late-resolution respondent writes, “I don't expect this to happen until there is an AI-caused catastrophe,” and "without that trigger institutional inertia wins.”
• The US political environment: Early-resolution respondents often treat the reported Trump executive order [it was signed on the last day the survey was open] as revealed intent: “Given that we see the Trump administration—which is highly anti-regulatory—considering this option, I think we will see more rational actors considering and implementing it relatively soon.” Another adds that it will be “good politics (voters are genuinely worried about AI.)” Late-resolution respondents tend to emphasize that the order was voluntary and [a stronger version] was ultimately shelved, and that meaningful U.S. action on this front will likely be deferred to a future administration: “Trump just backed away from AI regulation. So if the U.S. does it, I think it will happen after Trump leaves office”; “Unlikely in U.S. until after Trump.”
• Capability trajectory: Many early-resolution respondents argue that frontier capabilities are escalating rapidly, increasing both the risk posed by these models and the likelihood of restrictions being imposed: “I expect a qualifying government intervention to be more likely than not…mainly because frontier model cyber/bio/autonomy risk are becoming legible to governments…”; “Pandora’s box is about to be opened. The latest models by several companies already are far too dangerous.” Late-resolution respondents tend to downplay the risk: “AI has not yet done anything that is truly alarming or existentially disastrous.”
• Resolution-criteria threshold: Many early-resolution respondents read the bar as low: “The resolution criteria doesn't require the regulation to be particularly stringent, e.g. an order requiring the voluntary 2-3 month delay that Anthropic imposed on itself would count.” Others point to similar government restrictions: “Emergency action against a specific product can be near-instant once a galvanizing event occurs: the FAA grounded the 737 MAX three days after the second crash.” Whereas some late-resolution respondents read the same text as demanding a narrow, historically unprecedented form of action: “The resolution criteria are very narrow. General AI regulation, EU AI act compliance, transparency rules, or safety reporting would not count,” writes one, with another noting that “governments usually regulate deployment and use, rather than pre-release model publication.”
• Geopolitical and economic pressures: Several late-resolution respondents argue that the race between the U.S. and China for AI supremacy, paired with influence the AI industry exerts over politics, will stymie action: “There's a lot of money (and thus lobbying) involved, plus there's a lot of international competition where regulation would probably backfire or be a leaky bucket problem.”; “Too many rich and powerful men expect to make a lot of money off of AI, and they tend to wield power over elected officials.” Several early-resolution respondents point to the potential for the national-security element to cut the other way: “The most likely trigger is not broad safety regulation, but a U.S. national-security-framed restriction on frontier model weights—something closer to an export-control measure designed to prevent adversaries from accessing, copying, or distilling cutting-edge models.”

Financial Losses from Cybercrime

Question. What will be the total financial losses (in billion 2025 USD) reported in the FBI Internet Crime Complaint Center (IC3) annual report for incidents reported during the following resolution years?

Results. All participant groups forecast substantial increases in reported cybercrime losses between 2026 and 2031, and higher assumed AI progress is associated with higher losses. Unconditionally, the median expert forecasts IC3-reported losses of $27 billion in 2026, $40 billion in 2028, and $55 billion in 2031. Scenario assumptions matter increasingly over time: under Slow Progress, the corresponding expert medians are $25 billion, $35 billion, and $46 billion, while under Rapid Progress they rise to $31 billion, $50 billion, and $70 billion. The public is consistently lower, forecasting unconditional medians of $24 billion, $30 billion, and $40 billion, and expert-public differences are statistically significant across all reported dates, scenarios, and percentiles. Experts and superforecasters are much closer, with no statistically significant differences in 2028 and 2031, though superforecasters are somewhat higher by 2031, with an unconditional median $60 billion. Uncertainty also widens substantially over time: by 2031, the median expert’s 25th and 75th percentile forecasts are $40 billion and $70 billion unconditionally, and $52 billion and $93 billion under Rapid Progress.

Rationale analysis

• Offense–defense balance: A primary consideration for forecasters is whether AI will disproportionately advantage attackers or defenders. Many high-loss respondents argue the advantage lies with the attackers: “Defenders have to cover a widening attack surface, while attackers only need to find one vulnerability, one moment of misplaced trust, or one pathway into a financial relationship.”; “Defenders remain structurally disadvantaged…institutions move slowly, legacy systems remain everywhere, and human beings continue to be remarkably exploitable firmware.” A few low-loss respondents, however, argue the opposite is likely to be true: “Probably contrary to the typical rationale, I think rapid advancement in AI would actually produce counter scam technology. AI could analyze voices, speaking patterns, IP addresses, and other information and then prevent people from falling for scams.” Still others argue that, while attackers are likely to be disproportionately advantaged in the near term, defenders may win the day in the long term: “At least initially AI will accelerate losses. Maybe by 2031 AI will make systems secure enough to reverse the trend.”
• Trendline extrapolation: High-loss respondents often anchor on the recent growth in reported cybercrime losses, focusing in particular on the recent spike: “Internet Crime Complaint Center losses grew from $1 billion in 2015 to nearly $21 billion in 2025, and the growth continues apace. Extrapolation is a poor guide for long horizons, but for five years out it seems like a natural starting point.” Low-loss respondents tend to think the curve is likely to flatten before then. One writes, “There is a negative feedback loop on these numbers: the larger they get, the more efforts will be made to guard against these crimes,” while another argues that "the latest trend has us going to infinity too quickly…this curve will soon become a sigmoid.”
• AI's contribution to the trend: Many respondents treat AI-enabled deception as the primary driver of future losses: “AI can make phishing, impersonation, romance scams, investment fraud, and business-email-compromise schemes more scalable and more convincing. Deepfakes, automated targeting, and personalized messages may also increase the success rate of fraud.”; “AI directly scales the attack side with automated phishing, deepfakes, voice cloning, and large-scale social engineering.” Others express skepticism, with one writing, “I'm not really sure I see much connection between the AI progress and the level of cybercrime. I guess there are ways that advancing AI could increase or decrease cybercrime. But it seems pretty orthogonal to me.” Another argues that “an important factor for the growing cybercrime losses is due to the simple fact that more and more business and money stuff is handled through cyberspace.”
• Speed of societal adaptation: Many high-loss respondents expect victims and institutions will be slow to adapt, with one pointing to the “aging US population” and noting that “aging, asset-rich people are the most vulnerable cohort (ie, mental decline makes people vulnerable)” and another that the future is likely to resemble “a giant whack-a-mole ecosystem of criminal creativity that mutates faster than institutions can realistically adapt.” Low-loss respondents typically express more optimism: “I expect systems to be introduced to counter AI-enabled fraud at pace—things like multi-factor authentication, passkeys and hardware-backed authentication. Behavioral biometrics like typing rhythm, device handling, mouse patterns, and navigation habits add a signal that is harder for synthetic identities to imitate.”; “Trust may become less of a default social condition online and more of an operational achievement, something produced through verification, platform design, bank controls, and new habits of suspicion.”
• What rapid progress implies for agent use: Echoing points made when considering whether AI will disproportionately advantage attackers or defenders, several high-loss respondents predict rapid progress will result in, as one puts it, the emergence of “highly capable agents that can conduct large-scale fraud, social engineering, and cyber-enabled financial crime with very little human labor.” Whereas a low-loss respondent “suspect[s] that persistent personalized AI assistant agents will be deployed at scale in the next two years, and that one of their functions will be to screen inbound communications for scams.”
• Reported losses versus actual losses: Many respondents point to the potential for the gap between reported losses and actual losses to distort future reports. High-loss respondents often see reporting growth leading to higher numbers: “There is significant underreporting which could be mitigated if AI makes reporting easier, so I'm modeling a substantial upwards tail.” One low-loss respondent offers a counterpoint: “If US enforcement and reporting capacity remains constrained by resource pressures, and if broader conflicts of interest around crypto or financial regulation weaken incentives for aggressive fraud enforcement, reported IC3 losses may grow more slowly than underlying cybercrime.”

AI Catastrophic Risk

Question. What is the probability that a global AI-related catastrophe will occur by each of the following resolution years?

Results. All participant groups forecast that AI catastrophic risk rises with the assumed pace of AI progress. Unconditionally, the median expert assigns a probability of 0.3% by 2030, 2% by 2050, and 5% by 2100. Under Slow Progress, the corresponding expert medians are 0.08%, 1%, and 2%; under Rapid Progress, they rise to 1%, 5%, and 10%. The public consistently assigns higher probabilities than both experts and superforecasters, with unconditional medians of 1%, 3%, and 7%, and a median of 15% by 2100 under Rapid Progress. Superforecasters are more optimistic than experts, assigning unconditional medians of 0.1%, 0.88%, and 2.4%. Differences between experts and the public are statistically significant across all reported dates and scenarios. Expert-superforecaster differences are also statistically significant in the unconditional, moderate-progress, and rapid-progress forecasts at all three dates, while slow-progress forecasts are much closer. Dispersion remains wide even among experts: under Rapid Progress, the expert interquartile range is 2%-10% by 2050 and 3%-22% by 2100.

Rationale analysis

• Pace of progress as a risk factor: A clear majority of both optimists and pessimists view faster progress as raising risk because it compresses the time available for adaptation: “Faster progress both pulls forward the expected date of a potential catastrophe and increases its likelihood at any time, because we have less time for policy and society to respond.” Pessimists tend to focus exclusively on that factor, whereas optimists often either acknowledge the compression issue will raise overall risk but point to mitigating factors—“progress could either support more safety (a more advanced system should have better defenses against harmful uses) while also enabling more risk”—or, in a handful of cases, argue that rapid progress is a less risky path relative to the other scenarios: “It may be slow progress we most need to worry about, when models are already susceptible to adversarial attacks by people but aren't sufficiently developed to prevent them.”
• Safety mandates: Many optimists argue that governments and institutions, catalyzed by smaller warning events, are likely to intervene before the 10%-mortality threshold is in danger of being reached: “Prior to such a large event, there would very likely be smaller ones, perhaps in which the fatalities are in the thousands or tens of thousands. Such events would most likely bring on a backlash in which remedies were put into place to avoid future events. (See Chernobyl, Three Mile Island, and Fukashima.)” Pessimists typically are far more skeptical: “The political willingness to effectively regulate AI at the national level or cooperate internationally is sorely lacking”; “[AI]-race dynamics within and across countries, unethical or stupid people in charge, the push to put AI in weapons…preclude both serious thinking on AI alignment and worldwide regulations on AI.”
• The 10% threshold: Optimists treat ~800 million deaths as nearly disqualifying: “All of WWII, including civilian deaths, was 3%…So the question is how many nines of unlikely?”; “I just see this scenario as the stuff of Hollywood, but not realistic, especially when we consider we’d be talking about almost one billion deaths.” Pessimists often argue the high bar suggested by history is deceptive given that “our growing power to do damage is likely to outweigh” our resilience. Some also argue that the base rate, however low, is still meaningful: “The only clear qualifying event for the global population is the Black Death, but there are a few cases that seem like they’d qualify if you considered only ‘the known world to the population affected’ (the Plague of Justinian, Cocoliztli, New World smallpox). I get a rough base rate of 0.05% to 0.3% per year…It’s notable that several of these are cases of populations being exposed to a new pathogen they weren't adapted to.”
• Threat-vector pathways: Optimists and pessimists largely agree that engineered pathogens, AI-misalignment, autonomous weapons, and nuclear escalation are the most likely pathways. But pessimists tend to find at least one channel plausible: “Too many possibilities for my taste. And a few years from now, ANY major catastrophe will be AI-related.” While optimists are more skeptical: “After Covid-19, the world is better equipped to deal with a pandemic. Other than a pandemic, a nuclear war could potentially wipe out that many people, but human control prevails and it would need a very serious chain of misalignment problems…(see the Swiss cheese model—a lot of independent security holes need to be present at the same time).”
• Time-horizons: Most respondents deem the odds of a catastrophe occurring by 2030 to be exceptionally low, with many pointing to the fact that the slow/moderate/rapid AI-progress scenarios only speak to the state of progress by 2030. As one writes, “The 2030 resolution effectively corresponds to a single 5-year window (~2026–2030), during which AI capability is still developing.” A frequent corollary—used to justify convergence across scenarios at later dates—is that the state of AI progress by 2030 “[doesn’t] indicate much about the world or state of AI in 2100.”

Total Catastrophic Risk

Question. What is the probability that any global catastrophe will occur by each of the following resolution years?

Results. All participant groups forecast that total catastrophic risk rises with both the time horizon and the assumed pace of AI progress. Unconditionally, the median expert assigns a probability of 1% by 2030, 4.2% by 2050, and 10% by 2100. Under Slow Progress, the corresponding expert medians are 0.75%, 3%, and 8%; under Rapid Progress, they rise to 2%, 7%, and 15%. The public is generally more pessimistic, especially under scenario-conditioned forecasts: by 2100, the median public respondent assigns probabilities of 10% under Slow Progress, 12% under Moderate Progress, and 20% under Rapid Progress. Superforecasters are slightly more optimistic than both groups, assigning unconditional medians of 0.6%, 3%, and 9%. Differences between experts and the public are statistically significant across all scenario-conditioned forecasts, though the groups converge in the unconditional 2100 forecast, where both assign a median of 10%. Dispersion remains substantial: under Rapid Progress by 2100, the expert interquartile range is 7%-25%.

AI Share of Total Catastrophic Risk

The figure and table below are computed per respondent rather than from group medians: for each participant who answered both questions (at the 50th percentile), we take their personal AI share (AI Catastrophic Risk ÷ Total Catastrophic Risk), then report the weighted median and interquartile range across respondents. This is a median of individual ratios, not a ratio of medians. Only logically consistent respondents are included (i.e., those with Total > 0 and AI ≤ Total). The table shows the median (25th–75th percentile, n).

Rationale analysis

Because total catastrophic risk (this question) necessarily includes AI-caused catastrophic risk (the prior question), the analysis below focuses solely on why some respondents perceive a modest increase in risk between questions and others a major one.

• Modest increase: These respondents tend to reason that AI’s footprint is becoming so broad that a genuinely AI-free catastrophe is unlikely: “AI will be heavily interwoven in all aspects of life such that almost any major decision/activity would be limited by the ‘if not for’ test.”; “AI could become instrumental in everything that happens henceforth, and there would then be no meaningful difference between Q3 and Q4.”; “The bulk of my probability mass falls on AI-caused disasters, so I am only making small adjustments here…there just aren't that many risk factors that don’t at least indirectly flow through AI.”; “Little additional probability that wouldn't be causally linked to AI, as AI takes an increasing share of catastrophic risk-relevant decision-making (pandemic prevention, asteroid detection, nuclear command & control…)” Others argue that AI will help mitigate the risk of a non-AI catastrophe occurring: “While the risk of global catastrophe are slight in almost any scenario, AI I think can actually make us safer with time, for example, by monitoring threats from out space, or by helping analyze new viruses and craft vaccines.”
• Major increase: Many of these respondents reject the premise that AI will be entangled with nearly everything and treat the dominant non-AI pathways to catastrophe as carrying large standalone risks: “Pandemic risks and world war top the risks, followed by risks from climate change causing widespread crop failures or oceanic biosphere collapse. I consider AI to be a small component of overall threats to the world.”; “I consider catastrophes resulting from climate change to be the most likely, followed by wars between highly populated nuclear-armed nations.”; “Unlike the AI question, baseline risk here is already non-negligible without AI—pandemics, nuclear war, [and] climate collapse are all live on a 75-year horizon.”

First Major AI Global Harm Event

Question. What is the probability that an event, primarily driven by at least one AI system, will cause at least 50 deaths or $100 billion (in 2025 USD) in damages by the end of 2050? If such an event occurs by the end of 2050, in what calendar year will the first qualifying event occur?

Results. Experts and superforecasters both give more than a 50% chance that a qualifying AI-driven harm event occurs by the end of 2050. The median expert assigns a probability of 62%, while the median superforecaster is slightly higher at 70%; however, these forecasts are statistically indistinguishable. The public is substantially less pessimistic, with a median forecast of 35%, and both expert-public and public-superforecaster differences are statistically significant. There is also substantial disagreement within groups. Half of experts’ forecasts lie between 40% and 90%, and expert subgroup medians range from 50% among computer scientists to 75% among economists and industry respondents.

Conditional on such an event occurring by the end of 2050, experts and superforecasters expect it to arrive in the mid-2030s. The median expert and superforecaster each assign a 50% chance that the first qualifying event occurs by 2035, while the public is somewhat later at 2037. Expert-public differences are statistically significant across the elicited distribution, from the 5th through 95th percentiles. Experts and superforecasters, by contrast, give closely similar forecasts and do not differ significantly at any elicited percentile. Uncertainty remains substantial: the median expert gives a 25% chance that the first event occurs by 2031 and a 75% chance by 2040, with a 5% chance by 2028 and a 95% chance by 2048.

Rationale analysis (I)

• The 50-deaths/$100-billion threshold: The vast majority of respondents express profound pessimism that a global harm event of this magnitude can be avoided. Most anchor on the low bar set for a qualifying event and emphasize that there are multiple pathways for getting there: “It’s a single greyhound bus. It’s a single airplane.”; “$100B in damages is roughly one major hurricane, and 50 deaths is a single plane crash.”; “A 50-fatality threshold is trivially low for almost any military system. A single drone-swarm engagement in a medium-intensity conflict could exceed it.” Many also contrast the resolution criteria here with the survey’s catastrophe questions, with one calling it “a small bar by global-risk standards.” Of the minority of forecasters who express optimism that a qualifying event can be avoided, a few single out the financial arm as being demanding: “$100 billion is a high bar for accidental manmade harm. Only a relative handful of events clear it.”
• Autonomy and oversight expectations: Pessimists generally expect real decision authority to migrate to AI-controlled systems well prior to the 2050 resolution date: “AI will be built into all critical infrastructure by then.”; “The mere effect of explosive development and proliferation of these systems, making them increasingly pervasive in a wide range of domains, almost guarantees…that at some point one will meaningfully contribute to a major incident.”; “If AI is common (as it certainly will be by 2050), this seems almost inevitable.” A handful of optimists, however, expect institutions to withhold decision authority in key domains, with one writing, “It is difficult to envisage that AI will autonomously manage nuclear plants or power grids without oversight.”
• The ‘primarily driven’ threshold: The resolution criteria specify that for an incident to count it must be “primarily driven by at least one AI system” and the optimist minority often points to this requirement as likely to be decisive—i.e., they are not predicting that an event of this magnitude will not occur so much as they are contesting the notion that it is likely to be attributed to AI: “Because the question requires the event to be primarily driven by AI, I think the probability remains very low, though not zero.”; “The technology itself is not going to be the cause of the event, it will certainly involve human operators.” Several pessimists agree that this argument has merit and cap otherwise higher forecasts: “I am not going higher than 80 percent because the proximate cause criterion is strict. Many incidents that look AI-driven will have human decisions in the chain that disqualify them,” argues one, while another writes, “I hold roughly 8% back mainly for attribution risk.”
• Historical baseline: Pessimists frequently treat existing incidents and near-misses as evidence the remaining gap is small, with several citing the 737 MAX incidents: “Consider that Boeing 737 MAX planes already killed >300 people due to an autonomous control system. I guess that doesn’t count as AI under MIT AI Risk Initiative’s definitions (fine), but it points to this being quite likely as AI becomes increasingly integrated.” Another noted the ‘Autonomous Kargu-2 Drone Allegedly Remotely Used to Hunt down Libyan Soldiers’ incident reported by the resolution source and still others pointed to rising counts: “The severity and number of damages by AI jumped up considerably in 2025 and early 2026.” Optimists tend to read the same record and come away reassured: “Although AI incidents are rising, no confirmed case has yet met the 50-death or $100B threshold, which keeps the near-term base rate limited.”
• AI’s net effect on risk: Some argue AI is likely to displace error-prone humans rather than add new danger, with one writing, “I model this not as AI creating new accidents but as the transfer of causal responsibility for existing large accidents from humans to AI…Waymo's reported 90 percent reduction in serious-injury crashes relative to human drivers is my anchor.” Pessimists more often treat ubiquity itself as the hazard, even if AI results in fewer overall safety incidents: “AI will be everywhere. AI makes mistakes. Even with perfectly aligned AI this will definitely happen.”

Rationale analysis (II)

• Early-resolution respondents often argue that this could resolve soon given that AI is already sufficiently capable and integrated into critical systems: “The military capability to do this already exists, it’s just a question of when a state decides it is acceptable to use completely AI powered targeting.”; “This question will almost certainly resolve by the end of 2027 as a result of development of autonomous drones in their war with Russia.” Others reason incidents of this nature are likely to occur when the technology is still nascent, i.e. “while there is a considerable gap in our understanding of the system” and “before protective practices have a chance to develop.” As one writes, “An analogy would be a high rate of airplane incidents and accidents as a percentage of total, in the early years of rapid growth of the industry.”
• Late-resolution respondents typically point to the need for greater autonomy and deployment to materialize before an incident becomes likely: “Widespread full autonomy and agentic deployment will ramp up gradually through the 2030s rather than arriving suddenly.”; “Deployment into safety-critical systems is still relatively limited.” Several expect humans to remain in the loop where stakes are high—“AI systems will continue to operate with substantial human oversight, especially in domains where failures could cause large-scale harm”—with one adding that this dynamic means near-term incidents will likely be “human-using-AI rather than AI-as-primary-agent.”

Question (I). What is the probability that an event, primarily driven by at least one AI system, will cause at least 50 deaths or $100 billion (in 2025 USD) in damages by the end of 2050?

Question (II). If such an event occurs by the end of 2050, in what calendar year will the first qualifying event occur?

Democratic Backsliding and Power Concentration

Question. How many of the ten largest economies by GDP (as of 2025) will have an EIU Democracy Index score of below 4 (“authoritarian”) at the end of the following resolution years?

Results. Experts and superforecasters forecast relatively little near-term change in the number of top-10 economies classified as authoritarian. Unconditionally, the median expert forecasts 2 such economies in 2027 and 2030, rising only slightly to 2.1 by 2040, while the median superforecaster remains at 2 across all three dates. The public is consistently more pessimistic, with median forecasts of 2, 3, and 3 respectively. Under Rapid Progress, the gap widens further: the median expert forecasts 2 authoritarian top-10 economies in 2027 and 2030 and 3 in 2040, compared with 3, 4, and 4.6 for the public. Experts and superforecasters are broadly similar across most dates and scenarios, while the public assigns more probability to higher-count outcomes, especially from 2030 onward, and many of these differences are statistically significant. Even under Rapid Progress, however, the central forecast in every group remains below five of the ten largest economies.

Rationale analysis

AI’s effect on democracy: The primary disagreement was over whether AI will lead to any democratic backsliding at all; many low/no-backsliding respondents question that it will, regarding it as a “neutral tool that people [can] use to push society toward or away from authoritarianism”: “Technology produces good and bad effects, and in aggregate I think this is just going to cancel out relative to other societal factors impacting politics.”; “There are plenty of other techno-positive visions where AI actually enables democracy, holds power to better account, and it makes it harder for authoritarians to propagandize and subject citizens.” High-backsliding respondents tend to focus on the potential for AI to strengthen mass surveillance, censorship and repression, with one writing that “with AI the ability both to monitor and influence will become more pronounced. AI will make it easier for a government to control the population…[and] control or influence the many.” Most high-backsliding respondents also view rapid AI progress as amplifying the risk: “These technologies are likely to be quite effective even under the Slow scenario, though under the Rapid scenario they will be frighteningly effective.”

Institutional resilience: Low/no-backsliding respondents often stress that democracies are resilient: “Liberal democracy is a remarkably enduring form of government, many of the countries on that list have remained democratic despite existential wars of survival, invasion and occupation and centuries of technological and social change.”; “Institutional inertia, courts, regional governments, elections, bureaucracies, markets, and civil society create drag against full consolidation of power.” Others argue that even if backsliding were to occur, it would take “years if not decades,” therefore they see “little reason for any changes to happen by 2027” and predict “little change by 2030” as well. One notes “the case of Orban shows that voters can push back against authoritarians when corruption and incompetence becomes too much to stomach.” Several high-backsliding respondents, however, note that “a single violent coup potentially flips a country rapidly” and that the rapid AI progress scenario in particular “increases the variance... [and] disproportionally adds probability mass to less democratic outcomes,” so the near term is not as inevitable as base rates imply.

The <4 threshold: Low/no-backsliding respondents often stress that a country can suffer “substantial democratic backsliding, institutional capture, expanded surveillance, media manipulation, and attacks on civil society while remaining above that line,” and that recent erosion has “at most” shifted democracies to “flawed” status; many note that falling below 4 “requires a full collapse,” with countries passing “through the ‘flawed’ and ‘hybrid’ bands first.” As one forecaster writes, the Economist Intelligence Unit “never rated Hungary below a 6, even at the height of Orban's power.” High-backsliding respondents, however, tend to treat the <4 threshold as particularly plausible for the later resolution dates: “Over fifteen years, shocks such as war, severe economic crisis, contested elections, constitutional breakdown, or AI-enabled power concentration could push one or more democracies below 4.”

• The swing states: Much of the discussion centers around which countries might dip below the <4 threshold. High-backsliding respondents often point to India, and to a lesser degree the U.S., as potential candidates: “The only realistic marginal candidate is India (6.96), which has declined roughly 0.1 points/year under Modi over the last decade”; “The next three to possibly go over the line—in order—are the U.S., India, and the UK.” Low/no-backsliding respondents typically express more optimism, albeit guarded; one predicts that India “reaches the hybrid-regime zone (~5–6) by 2040 but [does not fall] below 4”— and several note that a post-Putin democratizing Russia could instead move the count below 2: “It's possible that the index may increase in certain countries, in particular in Russia. It's not inconceivable to think that by 2040 Putin will have been dead long enough for a successor regime to have moved the country in a more open and democratic direction.”

Footnotes

1. In some cases, the "aggregate" refers to the mean; in others, the median is used, depending on which is more appropriate for the distribution of responses. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

2. We occasionally elicit participants' quantile forecasts (estimates of specific percentiles of a continuous outcome) to illustrate the range and uncertainty of their predictions. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷