study

Help prioritise the capabilities governments will need in an AI loss-of-control crisis

When, where, how fast, in what form? No one can reliably predict an AI loss-of-control crisis, but governments will have to respond. Which capabilities will they need, whatever happens?

When, where, how fast, in what form? No one can reliably predict an AI loss-of-control crisis, but governments will have to respond. Which capabilities will they need, whatever happens?

register your interest

Join our Delphi expert panel | Two rounds, in November and January | 20–60 minutes per round | Help validate a method for advanced AI risk planning | Register your interest by 16th of November 2026.

The study

The Arcadia Impact AI Governance Taskforce is running a Delphi study on the capabilities that governments will need for AI loss-of-control crises.

We typically do not know how untested capabilities will fare in crises that have yet to happen. By combining the judgement of many experts, the Delphi process can build a credible evidence base where data do not yet exist.


Who we are looking for

People whose work or research has given them a well-grounded understanding of either:

  • how advanced AI could go wrong, and how it is governed or evaluated (for example through work in AI policy, safety, evaluation or assurance, or related research); or

  • how governments respond to crises, and what they can and cannot do (for example through work in or close to government, public administration, regulation, or emergency or security planning).


If you are unsure whether you qualify, please apply anyway: we will check eligibility for everyone.

What we ask

  • Two rating rounds: round one in mid-late November; round two expected in January (dates to be confirmed, with at least two weeks’ notice).

  • Flexible: each round is open for around two weeks, with a typical ask of 20–60 minutes.

  • Expert panel members are anonymous by default; experts who complete both rounds may choose to be credited by name in the paper.

  • The study is unpaid and voluntary. Privacy notice.


Why this matters

Rapid development and sudden capability jumps [1] make advanced AI hard to predict [2,3]. When the future is too uncertain to forecast, effective risk mitigation shifts from asking which risks are most likely to asking which capabilities protect against most risks [4].

This is an application of “decision making under deep uncertainty” (DMDU), an approach to planning well established in fields such as defence. Frontier AI developers’ “if-then” commitments partly follow this logic [5], but governments have yet to apply it to AI.


Arcadia ImpactLeft: a scatter of many AI risk scenarios plotted by estimated likelihood and impact, in four families: loss of control, AI × cyber, AI × bio and power concentration. Inverting the question gives the right panel: a grid of capabilities against scenarios ordered by severity. Capabilities 1 to 4 help across most scenarios (no regret); capabilities 5 and 6 only matter for the most severe scenarios (specific insurance).Prioritise across risk scenariosimpact →estimated likelihood →loss of controlAI × cyberAI × biopower concentrationone scenarioinvert thequestionPrioritise across government capabilityscenarios, ordered by severity →cap 1cap 2cap 3cap 4cap 5cap 6no regretspecificinsurance

Inverting the question. Left: scenarios ranked by likelihood and impact, where estimates diverge. Right: the same scenarios as columns, government capabilities as rows; those needed across many scenarios form a “no-regret” core, those needed in a few severe ones are insurance.

Inverting the question. Left: scenarios ranked by likelihood and impact, where estimates diverge. Right: the same scenarios as columns, government capabilities as rows; those needed across many scenarios form a “no-regret” core, those needed in a few severe ones are insurance.

What we have

We have devised a lightweight method for applying DMDU principles to advanced AI risk planning. In a pilot, we rated illustrative capabilities against AI-crisis scenarios [6]. The results revealed a “no-regret” core of capabilities useful across many scenarios, plus “insurance” for the most severe.

If a panel of independent experts, applying the same method, converges on a clear pattern of capability priorities, the method is validated and the panel’s findings substantiated. Governments then gain a tested way to plan and prepare for unpredictable AI crises.

Join our expert panel

Registering takes about two minutes. Know others who should take part? Please forward this page.

register your interest

References

1. Wei et al. (2022), Emergent abilities of large language models.

2. Weidinger et al. (2023), Sociotechnical safety evaluation of generative AI systems.

3. Hammond et al. (2025), Multi-agent risks from advanced AI.

4. Davis (2002), Capabilities-based planning, RAND.

5. METR (2025), Common elements of frontier AI safety policies.

6. Arcadia Impact AI Governance Taskforce (2026), Capability planning & preparedness.

Frequently Asked Questions

Who is this for?

People with either kind of understanding described above. You do not need both.

What happens in a Delphi study?

Our Delphi study will ask a panel of experts for independent, anonymous judgements over two rounds. In round one, you rate a set of government capabilities against one or more loss-of-control scenarios, giving your reasoning in each case. In round two, you see anonymised results and reasoning from other members of the expert panel and can choose to revise any of your round one ratings. Sharing ratings and reasoning lets the panel pool its expertise, while anonymity stops any one voice dominating. For more on the Delphi method: RAND’s guidance on Delphi panels.

How long does it take?

Each round is open for one to two weeks, time-dependent, with a typical ask of 20–60 minutes per round, depending on how many capabilities you choose to rate.

When will we know the round dates?

Round one is expected in mid-late November, round two is expected in January. Provisional dates will be confirmed to registered experts in due course and any changes made with at least two weeks’ notice.

Is it paid?

The study is unpaid and voluntary.

Will I be named?

Not unless you choose to be. Ratings are anonymous by default. Experts who complete both rounds may choose to be credited by name in the paper.

What happens to my data?

See the privacy notice.

What happens next?

We check everyone’s eligibility, then contact eligible experts to invite them to a short onboarding call.

Can I suggest others?

Yes! Please forward this page.