In October 2025 Deloitte refunded part of a A$440,000 report it had written for the Australian government on welfare compliance.
The report cited academic papers that did not exist. It quoted a federal court judge saying something no judge had said. A researcher at the University of Sydney found it by reading the footnotes.
Deloitte later disclosed that a generative model had helped write it.
A report like that has an author, a reviewer, a partner who signs it out and somebody at the department who receives it. The machine invented a judge's words, and every one of them passed it on. Not one of them was careless. Every one of them was somebody's human in the loop.
What a nod used to be
The first time I signed off somebody else's work I was nervous about it. Not about the service. About them.
A person had made the thing. They would read whatever I put in the box, and they would know from the first line whether I had really looked. Writing the feedback took longer than the review did. It felt like something I owed them, because it was.
I could not tell you how many things I approved last month. Not one of them had anybody waiting to read the comment.
The nod is still there. It has stopped being a gesture and started being a rhythm.
Nobody stopped checking
The comfortable reading is that people got lazy. It is the wrong reading, and it lets everybody off.
Nobody stopped checking. We check more than anyone has ever checked. Every machine-drafted thing that goes near a public service now has a person's name attached to it. Human in the loop is not a slogan any more. It is procurement language, and in the European Union it is law.
So the checking went up, steeply, in about three years. The question is what happened to the judgement inside it.
The number moves
In 2000 Linda Skitka, Kathleen Mosier and Mark Burdick ran the experiment that answers it.
People worked alongside an automated aid. One group was told they were answerable for their own decisions. The other was told they were answerable for some other part of the task. Same aid, same job, same failures sitting there to be caught.
The group answerable for their decisions made 1.32 wrong calls out of six. The other group made 3.14, and checked just over half the times they could have.
Nothing about the machine changed. Nothing about the difficulty changed. What moved the number was accountability, which is the ordinary fact that people think harder when they expect to answer for something, and think about the thing they expect to answer for.
Now hold your last sign-off against that. What were you answerable for? Not the reading. The tool that wrote it gives you a copy button and a thumbs up. Neither one asks what you checked. You were answerable for getting it off your desk, and you got it off your desk.
Jaded judgement
There is a word for how this feels from the inside. Jaded. The head gets heavier through the afternoon, and by the sixth one you are moving rather than deciding.
The yes costs less every time you give it.
Jaded is usually an accusation. It says the enthusiasm went and you let it go. Most people who sign things for a living have said it about themselves at about four o'clock.
It is the wrong charge.
A sign-off was never held up by the reading. It was held up by the odds of being asked about it. Ten a year, those odds were real, and you could feel them while you read. Ten a day, they are arithmetic. You know that. So does everyone sending you the work.
Nothing has gone wrong with your judgement. What used to make you look properly was the chance of being asked about it, and that chance has gone.
What you are spending
Judgement is a currency. You are issued some each morning, you spend it through the day, and none of it carries over.
Most of what you spend it on does not need it. The fourteen screens that were always going to be fine. The error message that has not changed in two years. By four o'clock the account is empty, and the one thing that needed you is still in the pile, looking exactly like the others.
Nobody keeps accounts on this. There is no statement at the end of the month, and nothing in the approval log that says what it cost you.
Two rooms, opposite answers
While the lawyers were writing people into the loop, the engineers were writing them out.
Ben Green went through forty-one policies requiring human oversight of government algorithms and found the oversight does not do what it promises. That was 2022. In those same years the field building these systems reached the same conclusion from the other side, and gave it a name.
Scalable oversight is the problem of supervising a system on work too complex for a person to check. It is not a worry at the edge of the field. It is a research programme with papers and benchmarks, and it starts from a premise nobody there bothers to soften: on most of the work that matters, the model is better than the person watching it.
So the law thinks you are the safety net. The people building the machine do not.
The smallest thing you did all week
Somebody read that report before it went to the department. More than one person, on a job that size.
Ask what any of them should have done differently and the answer comes back thin. Their job was the section in front of them, not the footnotes behind it. Checking a quotation is somebody else's task, or it is nobody's. The nod was the smallest thing they did all week.
That is the whole trouble. The nod was the cheapest thing in the process and the only part holding it up.
Nobody inside caught it. A researcher at another university did, reading footnotes nobody was paying him to read. Sit with that for a second, because it means the thing that worked was somebody outside the loop, with no queue of their own to clear.
Which is the part you can build. How many items land in one batch is a design decision. So is whether saying no takes one click or a meeting, and whether the reviewer can see how many they have already waved through this week. Most of that is currently built to make yes fast.
Somewhere in this month's pile is the one that needed you. It looks exactly like the others, and nothing in the pile will tell you which.