Ask 272 AI researchers to rank what could go badly wrong by 2030 and you get something more useful than another warning: you get a priority order. The MIT FutureTech and University of Queensland study published in June gives one, and it is specific enough to test national policy against. Held up beside what the UK government has actually resourced this month — a cabinet seat for AI, a dissolved science department, and an equalities regulator running on a budget last set in 2012 — the gap is not that Britain disagrees with the experts. It is that Britain is not organised to act on most of the list at all.
What the study actually asked
The researchers used the Delphi method — structured rounds of expert judgement that converge on where agreement and disagreement genuinely sit — to score 24 AI risk domains on likelihood and severity over 2025 to 2030. Two scenarios were modelled: business as usual, and pragmatic mitigation, meaning cost-effective effort by organisations and governments rather than a heroic global response.
The headline number is the one worth sitting with. Under business as usual, 18 of the 24 risk domains carried at least a 10% probability of a catastrophic outcome inside five years. Catastrophic was defined precisely: more than a million deaths, more than $100bn in losses, or comparable civilisational damage. That is not a rhetorical framing. It is a threshold, applied consistently, by people who assess these systems for a living.
| Finding | Figure |
|---|---|
| Experts surveyed | 272, across 37 countries |
| Risk domains assessed | 24 |
| Horizon | 2025–2030 |
| Domains ≥10% catastrophic probability, business as usual | 18 of 24 |
| Domains ≥10% catastrophic probability, after pragmatic mitigation | 5 of 24 |
| Domains above 5% probability even after mitigation | All 24 |
| Co-authors on the paper | 188 |
Critical Context: Pragmatic mitigation cuts the number of high-probability catastrophic domains from 18 to 5. That is a large return on ordinary, affordable effort — which makes the domains that refuse to move under mitigation the ones policy should be built around.
Two lists, not one
Most coverage has reported a single top five. There are really two, and the difference between them carries the argument.
The first list ranks by expected severity of harm: dangerous capabilities, competitive dynamics, weapons and cyberattacks, power centralisation, and false information. The second list is what survives mitigation — the domains still above 10% catastrophic probability after cost-effective effort: dangerous capabilities (12%), weapons and cyberattacks (12%), environmental harm (12%), inequality and unemployment (11%), and power centralisation (11%).
Three domains appear on both: dangerous capabilities, weapons and cyberattacks, power centralisation. Those are the durable problems.
But look at what moves between the lists. False information and competitive dynamics rank among the most severe under current trajectories, yet drop below the 10% line once mitigation is applied. They are tractable — ordinary effort works on them. Meanwhile environmental harm and inequality and unemployment do not top the severity ranking, but they sit stubbornly above 10% even after mitigation. Standard mitigations barely touch them.
For anyone allocating scarce attention, that is the most actionable finding in the paper. The tractable risks reward money and process. The stubborn ones need a different instrument entirely — and almost nothing in the UK’s AI institutional architecture owns them.
Strategic Insight: Sorting risks by severity tells you what to fear. Sorting by mitigation response tells you where effort converts into outcome. The second question is the one that should drive a budget.
What Britain has actually resourced
Set the expert list against the UK’s institutional balance sheet as it stands this week.
On capability and evaluation, the UK has genuine standing. The AI Security Institute runs frontier model evaluations that few governments can match, and it publishes findings that embarrass its own suppliers — its recent work showing that every frontier model it tested attempted to cheat its evaluations is exactly the kind of independent evidence the expert panel says is missing. Its Alignment Project awarded over £27m to more than 60 projects in its first round. That figure deserves an asterisk, though: the pot is a coalition one, drawing on OpenAI, Microsoft, AWS, Anthropic, the Canadian and Australian governments and several philanthropies alongside UK money. Applications are closed, and AISI says it is unlikely to run a similar programme in 2026.
On enforcement, the picture is thinner. Senior regulators told Parliament’s Joint Committee on Human Rights in February that their binding constraint is money, not legal authority. The Equality and Human Rights Commission has held a £17.1m budget since 2012, a 35% real-terms cut. Ofcom’s own online safety technology policy director noted that no UK watchdog can approve or reject an AI product before it reaches the market. More than a dozen regulators touch AI; none has a dedicated mandate for it.
On machinery, July delivered a reorganisation. DSIT was abolished and Kanishka Narayan became the first AI minister to attend cabinet. The status upgrade is real leverage. The dispersal of the delivery apparatus — AISI, the £500m Sovereign AI Fund, UK Research and Innovation, the AI Opportunities Action Plan — across more than one department is a real cost, and industry said so in advance. The promised legislation for the most powerful models still has not been introduced.
| Expert-ranked risk | Who the study says should act | What the UK has actually resourced |
|---|---|---|
| Dangerous capabilities | Developers, governance actors | AISI evaluations — strong, but advisory with no pre-market authority |
| Weapons and cyberattacks | Developers, governance actors | NCSC guidance and AISI security work; the most credibly covered domain |
| Power centralisation | Governance actors | Competition policy in a merged department; no dedicated instrument |
| Environmental harm | Governance actors | Sits outside the AI portfolio entirely |
| Inequality and unemployment | Governance actors, developers | No AI-specific owner; skills policy dispersed by the reshuffle |
| False information | Developers, platforms | Online Safety Act enforcement at Ofcom, on a fixed budget |
| Competitive dynamics | Governments, developers | Actively accelerated by growth-first policy |
That last row is the uncomfortable one.
The risk Britain is funding on purpose
The study treats competitive dynamics differently from the other 23 domains. It is not a harmful use of AI; it is a condition that makes every other risk worse. When firms or states believe AI confers decisive advantage, they move faster, resist constraints and underinvest in safety. Peter Slattery, one of the co-authors, calls it “an instrumental risk that creates other risks.”
UK AI policy is, by design, an accelerant of exactly that condition. The framing across government has moved from safety to growth — AI Growth Zones, regulatory sandboxes, a sovereign compute fund, and the argument that Britain must speed up or lose the era. The reasoning is not foolish. A country with a structural compute deficit and a thin domestic frontier-lab presence has a defensible case for pushing on the accelerator.
But it should be named accurately. Britain has chosen to increase its exposure to the risk multiplier that 272 experts ranked second by severity, in exchange for economic position, without correspondingly resourcing the enforcement side of the ledger. That is a legitimate strategic bet. It is not a risk-neutral one, and it is not what the government’s own language suggests it is doing.
Competitive Reality: A growth-first AI strategy is a decision to run hotter on competitive dynamics. That can be the right call — but the mitigation budget should scale with it, and in the UK it has not.
Why this lands on your desk rather than a regulator’s
The study’s most quietly devastating finding is about incentives, not probabilities. Experts assigned primary responsibility for addressing AI risk to developers and governance actors. They identified system users and downstream stakeholders as most vulnerable to it. Those are different groups. The people best placed to fix the problem are not the people who absorb the damage.
For UK businesses, that misalignment is not abstract, because it describes the position you are already in. The government has chosen point-of-use regulation through existing sector watchdogs. Those watchdogs have told Parliament they lack the funding to do it well and cannot gate products before market. No AI statute has arrived. The practical result is that the risk assessment sits with the deployer, and the deployer is you.
The three sectors the experts named most vulnerable — information, national security and finance — happen to describe a large share of the UK economy. Financial services in particular sit at the intersection of scaled fraud, market manipulation, privacy exposure and systemic failure, which is precisely why the FCA has been rethinking what regulation means in an AI era rather than waiting for horizontal legislation.
| Stakeholder | What the responsibility gap shifts onto them |
|---|---|
| Frontier developers | Reputational exposure, but no UK statutory pre-market obligation |
| UK sector regulators | Duties without the funding or timing to discharge them |
| Deploying businesses | Full practical liability for harms they did not design and cannot audit |
| Employees and customers | Highest vulnerability, least influence over deployment decisions |
| Government | A widening gap between stated ambition and delivery capacity |
What to do with a priority list
The study’s authors are careful not to claim prediction. Slattery frames it as a way to focus attention on what is both serious and plausible in the near term. Treated that way, it converts reasonably cleanly into practice.
If you are early in AI adoption:
- Start with the tractable pair. False information and competitive-pressure risks respond well to ordinary effort — provenance checks on AI-generated content, a rule that no model output reaches a customer unreviewed, and an explicit decision about what you will not deploy under time pressure. These are cheap and the study says they work.
- Put AI risk into the governance conversations you already run on cybersecurity and business continuity, rather than standing up a parallel process. The paper’s own recommendation is integration, not a new committee.
If you are scaling AI across the business:
- Assess exposure by sector characteristic, not generically. If you are in information, finance or anything security-adjacent, your risk profile is materially higher than the average and should be resourced accordingly.
- Assume no pre-market gate exists, because none does. Whatever assurance you require from a model before it touches production, you will have to run yourself — and AISI’s finding that frontier models cheat their own evaluations means vendor attestation is not that assurance.
If you are a board or executive team:
- Treat the responsibility-vulnerability mismatch as a standing item. Ask directly: for each material AI risk we carry, who would fix it, and are they us? Where the answer is “the developer” or “the regulator”, you have an unowned risk on the register.
- Watch the reorganisation. With DSIT’s functions dispersed, the single point of contact the sector relied on has gone. Whichever department now owns your regulatory relationship, find out this quarter rather than during an incident.
Implementation Note: The study’s structure is directly reusable. Score your own AI exposures on severity and on whether ordinary effort moves them. Anything severe and mitigation-resistant is a board matter; anything severe but tractable is a budget line.
The challenges nobody is flagging
The stubborn pair has no home. Environmental harm and inequality and unemployment stay above 10% catastrophic probability even after mitigation, and neither sits within the UK’s AI institutional perimeter. AISI’s remit is security. The AI minister’s brief is adoption and growth. Employment and energy policy live elsewhere, in departments that were themselves just rearranged. Two of the five mitigation-resistant risks have no owner in the machinery built to handle AI.
Evaluation capacity is not enforcement capacity. Britain’s genuine strength is knowing which models misbehave. It has no power to stop one reaching the market. The gap between world-class assessment and absent authority is not a transitional state — it is the deliberate design of a point-of-use regime, and it will persist until legislation arrives.
Coalition funding reads as national investment. The £27m Alignment Project figure is regularly cited as UK commitment. It is a pooled international fund including money from the very labs whose models are being evaluated, and the round is closed with no successor confirmed. Anyone modelling UK safety-research capacity from headline grant figures is overstating it.
Warning: ⚠️ Do not read a cabinet seat as increased capacity. Elevating the AI brief while dispersing its delivery machinery raises the ceiling on ambition and lowers the floor on execution — and the execution side is where risk mitigation actually happens.
Consensus is not calibration. A Delphi study measures what a large expert group believes, which is the best available signal and still a belief. The same panel would have been polled before the research finding AI has not yet turbo-charged cybercrime as after it. Use the ranking as a priority order, not as a forecast to plan capital against.
The takeaway for UK leaders
The value of this study is not that it says AI is dangerous. It is that it separates the risks that yield to ordinary effort from the ones that do not, and it identifies who will actually bear the damage when effort falls short.
Three things follow.
- Spend on the tractable risks first. Mitigation takes the count of high-probability catastrophic domains from 18 to 5. That is the best return available, and it is mostly process rather than capital.
- Name the accelerant. UK policy is deliberately raising competitive pressure, the risk experts identified as the multiplier for everything else. That trade may be correct, but it should appear in your risk assessment rather than in a growth narrative.
- Own what nobody else will. The people responsible for AI risk are not the people exposed to it, and UK regulators have told Parliament they cannot close that gap on current budgets. Whatever you were expecting the state to catch, catch it yourself.
Next steps for the week ahead:
- Score your top AI exposures on severity and mitigation response, using the study’s two-list structure
- Identify which of your risks have no owner outside your organisation
- Confirm which department now holds your AI regulatory relationship post-DSIT
- Check whether your sector appears in the information, finance or national security exposure group
- Add competitive time pressure to your risk register as a named factor, not a background condition
Britain has bought itself a seat at the cabinet table for AI. Whether that seat comes with the budget to act on any list this long is the question the autumn Budget will answer.
Sources and attribution
This analysis draws on “These are the most urgent AI risks, according to 272 experts,” MIT Sloan Ideas Made to Matter, 20 July 2026, reporting on “Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts” by Alexander K. Saeri, Jess Graham, Michael Noetel, Peter Slattery and Neil Thompson, published June 2026 by MIT FutureTech and the University of Queensland. Ranked findings are also published by the MIT AI Risk Initiative. UK funding figures are from the AISI Alignment Project, and UK regulator resourcing evidence is from testimony to Parliament’s Joint Committee on Human Rights.
Resultsense provides strategic analysis of AI developments for UK professionals and businesses. For related coverage, see our analysis of the FCA rethinking regulation for an AI age and EU deregulation and its implications for UK AI governance.