Workers say 15% of the work they receive is low-effort AI output, but has anyone audited that claim?

This explores whether workers' own estimates of AI "workslop" (work that looks finished but doesn't actually move the task forward) have been checked against independent quality reviews. The short answer from this collection: not yet. The main data point is a September 2025 BetterUp Labs and Stanford Social Media Lab survey, in which 1,004 U.S. desk workers estimated that 15.4% of the work they receive is unhelpful AI-generated content, and said each instance costs them nearly two hours to deal with How much work that employees receive is actually unhelpful AI content?. That figure is entirely self-reported. The corpus has no study that took a sample of that work and had auditors grade it. So the 15.4% is best read as a measure of how much slop workers feel they're getting, not a measured amount.

The gap matters because other research in the collection suggests people judge AI-assisted work unreliably in both directions. The "LLM fallacy" work finds that when AI output is fluent and blends smoothly into someone's own work, people credit themselves with skills they don't have Do AI-assisted outputs fool users about their own skills?. Applied to workslop, that cuts two ways. Senders may not recognize their own output as slop, so a survey of senders would undercount it. Receivers may blame AI for any weak work they get, so a survey of receivers could overcount it. A survey can't tell these two biases apart. Only an outside audit can.

What would a real audit look like? The AI safety literature turns out to be a useful model, even though it's about models and not office memos. One theme comes up repeatedly: measurement has to be fixed before you can act. The reward-hacking literature argues that mitigation can't even be evaluated until detection tools are reliable Can we measure reward hacking reliably enough to act on it?. Another note warns that consistency is not accuracy: an LLM at temperature zero gives the same answer every time, but that answer is still just one draw from the model's range of possible answers Does setting temperature to zero actually make LLM outputs reliable?. The same applies to survey estimates. Workers could agree closely on 15% and still all be wrong.

The collection also suggests how to design a check. Blind alignment audits used several independent teams with different methods and found a hidden behavior that no single method could have confirmed on its own Can auditors discover hidden objectives that models learned to conceal?. A workslop audit could work the same way: compare receivers' ratings, blind expert grading, and downstream outcomes such as rework time or errors that slipped through. BenchShield's approach also carries over. It replaces a single score with evidence about *how* a task was done Can infrastructure evidence replace terminal scores in benchmark validation?. For workslop, that could mean logs of revisions, follow-up questions, and redone work, which would show whether a deliverable actually helped. A caution from systems research applies too: each piece of work can pass a quick check (it looks polished and plausible) while the larger workflow still fails Can individual components pass safety checks if the system still fails?. That suggests auditing outcomes, not surface quality.

The takeaway you may not have expected: the best-known workslop statistic is a measure of perception, and the research on how people judge AI-assisted work suggests perception is exactly where the errors are. Checking it would take the kind of multi-method, evidence-based auditing that AI safety researchers already use on models. As far as this collection shows, nobody has yet applied that to the AI output landing in people's inboxes.