Handoff

Handoff is when a group of humans grant AIs a position of trust, decision-making, or both.

A paradigm example would be a frontier AI company handing off to their own AIs. These AIs might be in charge of: prioritising between research agendas; deciding how to develop and deploy successor AIs; managing relations with governments (foreign and domestic); managing relations with clients, suppliers, competitors, and the public; managing philanthropic ventures; etc.

Handoff varies along several dimensions:

1. Trust. If the AIs wanted to "screw us over", how badly could they do so?
2. Decision-making. How much scrutiny are humans still applying to the AIs' decisions?
3. Scale. Did a single individual hand off? A team within a company? An entire company? A government? A coalition of governments? Humanity as a whole?
4. Scope. Which decisions are AIs making? Just R&D, or all decisions within the company?
5. Capability. How capable are the AIs at handoff? Greenblatt argues for handing off near the minimum viable capability level (“Min-H”), since less capable AIs are less likely to be scheming and easier to oversee.
6. Reversibility. If humans decide to reverse handoff, how easily could they do so?
7. Transparency. Who knows that about handoff and its various dimensions?

The central questions include:

1. How desirable is handoff in different scenarios, given the AI's capabilities and propensities, and the wider strategic landscape? In these scenarios, what happens after handoff?
2. How can we make good decisions about whether, when, and how to hand off? What evidence would justify a handoff?
3. How can we increase the chance of good handoff, and decrease the chance of bad handoff?
4. What are viable alternatives to handoff?

See also:

1. Types of Handoff to AIs (Kokotajlo, 2026)
2. How do we (more) safely defer to AIs? (Greenblatt & Stastny, 2026)
3. AI 2040: Plan A, Alignment Roadmap supplement, Phase 4.