What happens when the AI stops waiting for orders and starts assigning you tasks — then grading how you did?

This explores what happens when the usual direction of delegation flips: instead of people handing tasks to AI and checking its work, the AI assigns tasks to people and judges how well they do them. The collection's most direct material is on the first half, AI handing work to humans. It has much less on AI formally grading the humans it hands work to. That second half has to be pieced together from nearby research.

The central argument comes from Holldack, Banh and Strobel. Once an information system can make decisions on its own, it can pass rights and responsibilities to people, not just receive them from people Can AI systems delegate tasks back to humans?. They say most existing theory assumes that only humans hand out work, so it can't describe a system that sets a task for a person and then waits for the result. Evaluation follows naturally from this. Whoever delegates usually also decides whether the job was done well. Agency theory names the problems that come with that role: one side knows more than the other, one side has room to make its own choices, and loyalty is uncertain Can traditional governance tools actually control AI agents?. Those problems were written with humans in charge. In a reversed setup, the person being assigned work and graded may be the one who can't see how decisions are made.

The tools for AI-as-evaluator already exist, but they point at other AI, not people. Agent-as-a-Judge systems, which collect their own evidence before giving a verdict, are about 100 times more consistent than a single language model acting as judge Can agents evaluate AI outputs more reliably than language models?. The same research found that errors in the judge's memory spread through later judgments. Applied to people, that means an agent grading a worker could build an early misreading into every later assessment. Work on multi-agent accountability adds a related point. Labelling someone "accountable" only changes outcomes if the workflow actually sends the deliverable through that party Does accountability placement actually shape multi-agent outcomes?. If AI evaluates humans, the real authority lies in how work is routed, not in what the org chart says.

The less obvious finding is that the human side of the reversal may already be underway, informally. Leonardi shows that when people credit an AI with independent judgment, they feel less in control and less responsible, even though they keep formal authority Does attributing agency to AI actually reduce human sense of control?. People can start treating an AI's view of their work as the one that counts well before any system officially grades them. Leonardi also finds that this shift moves in cycles: people can reassert control and reshape the arrangement. Trust research suggests where they push back. People pull away when actions can't be undone and other people can see them, not simply when the stakes are high What makes people distrust AI agents they delegate to?. A public, permanent AI verdict on a person's work fits that description closely.

The governance material implies limits. If risk grows with the autonomy given to an agent Does AI risk increase with the autonomy we give it?, then an AI that both assigns and grades human work sits near the top of that scale. Two proposals offer guardrails. One is to expand autonomy only in stages, each unlocked by trust that has been demonstrated Should agentic AI autonomy expand based on earned trust?. The other is to put monitoring in place before trying to align incentives Must AI monitoring precede incentive alignment as agents evolve?. Both were written for humans overseeing AI. Whether they work when the roles are reversed is an open question that the collection raises but doesn't yet answer.