The Work You Can Stand Behind
Most people are handing over AI output they wouldn't personally sign. The scarce thing now, frankly, is judgement: putting your name on the page. Four questions to test exactly that.
The memo arrives at 4pm, and it looks right: clean headings, a confident summary alongside numbers arranged with the superior authority of things that have been verified. You open it properly and something drops. The figures don’t reconcile. The recommendation is built on a market that isn’t yours. The second half, you realise, simply restates the first in slightly different clothes and adds nothing. Forty minutes later you’ve confirmed only one thing: there’s nothing here to build on, and you still have to write the memo yourself.
BetterUp Labs and Stanford’s Social Media Lab named it in Harvard Business Review in September 2025: workslop, meaning AI-generated content that looks like work and does none of the work’s actual job. Their survey of 1,150 US desk workers found 40% had received some in the previous month, each instance taking an average of one hour fifty-six minutes to untangle. The cost runs to around $186 per person per month, they calculated, which at 10,000 employees passes $9 million a year.
My first assumption, reading about AI-generated slop washing through workplaces, was that it was other people’s problem. Other people’s laziness.
Then the Work AI Index arrived on 10 June 2026 to rudely dispel my judgy assumption on other people’s laziness. Glean’s Work AI Institute, alongside researchers from Stanford, Berkeley, and five other universities, surveying 6,000 digital workers across the US, UK and Australia found that 69% of AI users admit to submitting work they haven’t verified or don’t fully understand, work they couldn’t confidently stand behind. The researchers called it ‘botshitting’. (I know. I’m screaming.) Heavy users are 64% more likely to do it than light users. The best AI users are the worst offenders.
Read the two studies together and something uncomfortable comes into focus: people are handing over the part of the job they were actually being paid for, and most haven’t noticed.
The time-saving headlines leave this out. The same Work AI Index that found workers saving 11 hours a week also found them spending 6.4 of those hours botsitting, babysitting and repairing the output. Only 13% said their organisation was performing significantly better for any of it. The work didn’t vanish; it moved from making the thing to checking the thing, which is, frankly, the harder job.
Ethan Mollick, who teaches at Wharton and has been watching this longer than most, calls the most capable AI users inside large companies “secret cyborgs” - people who hide how much the machine does because admitting it feels, to them, like admitting they’re replaceable. The truth runs the other way actually. The person who can catch what the machine got wrong is the one who is not replaceable.
That catching has an old name. Judgement.
Judgement is accountability with the consequences still attached, and AI can’t hold those for you. Sure, an AI model can draft the board report but can it sit in the room when someone leans forward and asks where the 14% came from?
The consulting world has a name for the constraint that replaced volume: verification, which sounds technical but actually means “whose fault is this” and the answer, still, is a person’s.
This is why the work you can stand behind is becoming the scarce thing, while everything that merely looks finished gets cheaper by the month. Anyone can generate a polished page now. Far fewer people can tell you which line in it would fall apart under a hard question, and fewer still bother to check before they press send.
The signature test
So before anything the machine helped you write or create leaves your desk, ask four questions. Call it the signature test, because the real question under all of them is whether you would sign it.
Can I explain how every number got here? If a figure appears and you cannot trace it back, it isn’t yours yet.
Would I defend this out loud to the person in the building who knows the most about it? Reading your own work well is easy. Surviving the one real expert in the room is something else entirely, and you know it.
Did I check the single fact that, if wrong, collapses the whole thing? Every piece has one (usually the one you were most relieved not to double-check). Find it, check it against a named source.
If this went out under my name and turned out wrong, would I be embarrassed by something I could have caught in five minutes?
Four questions. Most workslop fails the first one.
What to do this week
Pick the next deliverable you would normally let AI mostly finish: a report, a client email, a deck. Before it goes anywhere, do three things.
Find the load-bearing fact and verify it against a named source you can point to. One fact, properly checked, beats a polished page nobody stands behind.
Time your botsitting honestly. For one day, note how long you actually spend fixing AI output against producing with it. If you are anywhere near the 6.4 hours a week the research found, you’re not saving time, you are relocating it.
Write one thing this week with no AI at all, start to finish. Do it to check your writing muscle still works, because the muscle that judges work is the same one that does the work, and it goes slack when you only ever check.
The uncomfortable part is less the workslop than the 69% of us already attaching our names to it and feeling, apparently, more productive for doing so. The tools will keep improving at producing the page. What they cannot answer is the question waiting at the end of anything that actually matters: when someone asks you to stand behind this, do you know if you can?