Oversight readiness

Oversight readiness is whether the workforce of the future will be able to judge the AI work it is nominally supervising. The phrase comes from a Google DeepMind paper published in February 2026, and its value is the tense. Most arguments about deskilling describe something already happening to individuals. This one names a condition an organisation will discover it lacks, at the moment it needs it, some years after the decisions that removed it.

The source#

Nenad Tomašev, Matija Franklin and Simon Osindero, Intelligent AI Delegation, arXiv:2602.11865, submitted 12 February 2026. The paper is mostly about something else. It proposes a framework for how AI agents should break problems into parts and delegate them across other agents and people, covering transfer of authority, responsibility, accountability, role boundaries and trust.

The passage on human capability arrives as a risk the framework has to handle rather than as the subject of the paper, which is part of why it carries weight. The authors are not arguing about deskilling. They ran into it while designing delegation systems.

The mechanism, in their words#

The paper states that unchecked delegation threatens the organisational apprenticeship pipeline. In many domains, expertise is built through the repetitive execution of more narrowly scoped tasks, and those tasks are the ones most likely to be offloaded to AI agents in the short term. If learning opportunities are fully automated, junior team members would be deprived of the necessary experience to develop deep strategic judgement, impacting the oversight readiness of the future workforce.

That is the missing rungs argument, arrived at independently, by people building the delegation infrastructure. The work that trains a person is disproportionately the work worth automating first, because narrow and repetitive is what both descriptions have in common.

The failure state they name is precise. The goal, they write, is to avoid the future in which the human principal is able to delegate, but not accurately judge the outcome. Delegation without judgement amounts to distributing work with a signature attached, rather than supervising it.

The remedy requires not automating something#

Two proposals follow, and both are unusual coming from a frontier lab.

The first is curriculum-aware task routing. Rather than passive approaches such as having humans shadow agents during execution, they propose systems that track the skill progression of junior team members and allocate tasks at the boundary of their expanding skill set, within the zone of proximal development. AI agents co-execute, provide templates and skeletons, and progressively withdraw that support as the junior demonstrates proficiency. It is an apprenticeship, rebuilt inside the routing layer.

The second is blunter. A delegation framework should perhaps occasionally introduce minor inefficiencies by intentionally delegating some tasks to humans that it would not have otherwise, with a specific intent of maintaining their skills.

Read that again with the source in mind. An organisation building agentic systems is proposing that those systems should sometimes route work to a person who is slower at it, on purpose, to keep the person able to do it. The inefficiency is the point, and the cost of retaining a capacity to supervise.

They add a third suggestion for keeping people engaged rather than merely present: requiring human experts to accompany their judgements with a detailed rationale or a pre-mortem of potential failure risks. Writing down why, and what might go wrong, keeps participants in delegation chains cognitively involved.

Where it fits#

Oversight readiness sits between two things this research already names. Synthetic seniority describes what has happened to an individual whose output looks senior while the judgement underneath was never built. Oversight readiness describes what an organisation is left holding when enough individuals are in that position: a supervisory layer that cannot supervise.

It also gives the accountability argument its timing. A moral crumple zone forms when responsibility settles on someone who could not realistically have intervened. Oversight readiness explains how a workforce arrives at that condition gradually, through a sequence of individually sensible routing decisions, none of which looked like a decision about capability.

The regulatory tests now being written assume the readiness exists. The UK standard for meaningful human involvement asks whether a person has the authority, discretion and competence to alter a decision. Competence is the third condition, the one that decays quietly while the other two remain on paper.

What it does not settle#

The paper proposes routing systems that track skill progression, which assumes an organisation can measure what a junior can currently do. Most cannot. Assessment of capability rather than output remains the unsolved part, and a curriculum-aware router without a reliable measure of proficiency will route by proxy, most likely by tenure or by throughput.

There is also a governance question the paper leaves open. Deliberate inefficiency has to be defended in a budget. Somebody must be willing to explain why a task went to a slower human when a faster agent was available, and to keep explaining it in quarters where the cost is visible and the benefit is not yet due.

Key sources

• Tomasev, N., Franklin, M. and Osindero, S. (2026). Intelligent AI Delegation. Google DeepMind, arXiv:2602.11865, submitted 12 February 2026. All three authors are at Google DeepMind. The phrase appears once, in section 5.6, "Risk of De-skilling"; every quotation on this page was checked against the full text on 5 September 2026.