Leading with Complexity and Wisdom for AI Safety Leaders

AI Safety organisations are growing fast. Many of their leaders are coming from technical or policy backgrounds with no management or soft skill training. 79% of AI Safety organisations have no dedicated people leadership (RAISEimpact). A survey of 25 EA leaders (assuming there is a strong overlap between AIS and EA in terms of mindset; without assuming that both groups are exactly the same) found that many of the top sought-after skills are management, not research (Meta-Coordination Forum 2024). Yet only 3 of 20+ AI Safety training programs are for non-researchers (Chris Clay, EA Forum, Aug 2025).

The result is that many brilliant people run teams with a leadership style derived from how they would do science. They treat most problems as being reducible to its components, i.e. complicated (solvable with enough analysis). They become the single point of coordination because they can't delegate what they don't understand as a skill (Mahajan, EA Forum, June 2026), avoid difficult people decisions because intellectual cultures frame directness as unkind (RAISEimpact; EA Forum coaching trial, 2022), and confuse intellectual rigour with management competence (Kat Woods, EA Forum, Aug 2025).

Existing programs address important parts of this, such as RAISEimpact (raiseimpact.org) building operational systems inside orgs, and WorkStream (workstreamnonprofit.org) providing operations consulting. But none of them address the modality and quality of judgment itself, i.e. the capacity to reason wisely under radical uncertainty, distinguish which parts of a problem need analysis and which parts need complexity approaches, as well as when to consult the team ("crowd wisdom") or AI on high-stakes decisions and trade off multiple perspective against ones own biased judgment.

This project fills that gap by combining two scientific frameworks for organisational and leadership settings:

• Cynefin (Dave Snowden): Being used in org settings for over two decades, it distinguishes Complicated domains (where expert analysis works) from Complex domains (where cause and effect are only visible in retrospect). Most challenges in scaling AI Safety orgs, from team dynamics, strategic direction, culture, hiring judgment, are in fact Complex. Leading them as if they are Complicated is a category error that is most common in organisations (including AI Safety orgs).
• Wise reasoning (Igor Grossmann, University of Waterloo) provides empirically validated exercises that measurably improve judgment. His RCT with 555 participants showed that the technique of "distanced self-reflection" produces significant gains in intellectual humility, perspective-taking, and recognition of uncertainty (published in Psychological Science, 2021).

The program has two phases:

• Phase 1: A one-day intensive online workshop where 8–10 AI Safety leaders, preferrably with a technical or scientific background, work on a real decision they're currently stuck on. Participants experience what it feels like to reason differently using the above frameworks.
• Phase 2: Four 1-on1 coaching sessions per participant over three months, focused on specific management challenges. The workshop insights should become a sustained practice.

Planned Outputs:

1. a tested workshop design that could be replicated across the field
2. pre/post assessment data (incl. using Grossmann's Situated Wise Reasoning Scale (SWIS) adapted for management contexts)
3. a public write-up of what worked and what didn't, shared on Substack, LessWrong, or other suggested outputs; so others can build on it

The field's capacity to reduce x-risk depends on its organisations working well. Right now, many are working suboptimally, because talented researchers and policy people have been placed in management roles without the judgment and soft skills associated with good managers.

As mentioned above, leadership patterns such as defaulting to analytical problem-solving, avoiding complex people decisions, and centralising coordination in the founder, lead to many problems. New AI Safety orgs commonly have no governance backbone, no delegation structure, and no cross-functional cadence (Mahajan, EA Forum, June 2026). Leaders spend 75% of their time on low-value tasks they agree should be delegated (EA Forum coaching trial, 2022). Frustrations accumulate silently until they erupt (RAISEimpact).

This matters for x-risk in three direct ways:

1. Talent retention: When a good alignment researcher leaves because their manager can't give useful feedback, navigate conflict, or set clear priorities, it is a direct loss of safety capacity. The AI Safety talent pool is too small to waste on management dysfunction.
2. Decision quality under uncertainty: AI Safety leaders make high-stakes decisions under radical uncertainty. These are Complex-domain decisions in the Cynefin sense. Leaders who default to the Complicated-domain approach (more analysis, more data, more certainty before acting) either make decisions too slowly or make the wrong kind of decision entirely. Grossmann's research shows that the capacities needed for Complex-domain judgment (intellectual humility, perspective-taking, recognition of uncertainty) are trainable (Grossmann et al., Psychological Science, 2021). The MATS 2026 survey confirms this urgency: as AI automates execution, "research taste, strategic judgment, and human-facing capabilities" become the differentiating skills (MATS Research, March 2026). This program builds exactly those.
3. Absorbing scale wisely: The ecosystem is about to receive significantly more funding, with the bottleneck not being money but management capacity to deploy money well. The FTX period showed what happens when organisations scale faster than their governance and leadership capacity (incl. wise judgment, how to build trust and keep others accountable) can support (EA Forum). Published results from RAISEimpact show that targeted interventions can move team satisfaction from 4.9 to 7.9 in eight months (raiseimpact.org), demonstrating that these problems are tractable. This program complements that operational work by building the underlying judgment capacity that makes good management decisions possible in the first place.

The theory of change: leaders who can distinguish complicated from complex + who reason wisely under uncertainty (increased meta-cognitive skills) → better-functioning AI Safety orgs → more effective work on reducing x-risk

Minimum ($15,000) - Core pilot with 6 participants:

• Program design and preparation: $3,000
• Workshop facilitation (1-day intensive): $1,500
• 1:1 coaching (6 participants × 4 sessions): $5,000
• Participant recruitment, coordination, scheduling: $1,500
• Pre/post assessment and evaluation write-up: $2,000
• Contingency: $1,000

This covers the essentials: a well-designed pilot workshop, meaningful follow-up coaching, and a proper evaluation. No co-facilitator, no scientific advisory, smaller cohort.

Ideal ($27,000) - Full pilot with 8-10 participants, co-facilitator, and scientific advisory:

• Program design and preparation: $4,000
• Workshop facilitation (1-day intensive): $2,000
• 1:1 coaching (8–10 participants × 4 sessions): $7,000
• Co-facilitator (workshop delivery + observer role + coach): $5,000
• Scientific advisory: Igor Grossmann (measurement design, program review, impulse talk as part of the workshop): $3,000
• Participant recruitment, coordination, scheduling: $2,000
• Pre/post assessment and evaluation write-up: $2,500
• Contingency: $1,500

The ideal budget adds three things:

1. A co-facilitator from collectiv:a, an organisational development cooperative, who brings 5–10 years of professional leadership development and transformation consulting experience and can observe group dynamics during the workshop.
2. Igor Grossmann for scientific advisory and providing a theory impulse during the workshop about wise judgment under radical uncertainty, as well as supporting to co-design a proper measurement framework using validated instruments (such as the Situated Wise Reasoning Scale).
3. A larger cohort of 8–10 participants for more robust pilot data.