When AI output fails downstream, the cause is almost always the same: employees using AI beyond the boundary where it performs reliably, without knowing that boundary exists.
MIT Sloan’s research on AI-assisted work identifies a consistent finding: performance drops measurably when people use AI outside its capability frontier. The capability frontier is the threshold beyond which a given AI tool moves from reliable output to confident-sounding but unreliable output. Employees working inside that frontier see genuine productivity gains. Employees working outside it produce what researchers, journalists, and practitioners have started calling workslop: output that looks finished, moves through the system at speed, and fails when it reaches someone who has to act on it.
The Human Performance Intelligence framework treats workslop as a human performance problem. Understanding why requires looking at what actually happens at the capability frontier, and what the human system needs to manage it well.
What the Capability Frontier Is and Why It Moves
Every AI tool has a domain where it performs reliably and a domain where its outputs become probabilistic in ways that are not visible on the surface. The boundary between those domains is not fixed. It shifts with the task, the prompt structure, the available context, and the complexity of the output required.
This creates a specific challenge for human performance. A person learning to use a hammer develops a stable mental model: the tool works this way, in these conditions, and fails in these other conditions. A person learning to use an AI writing or analysis tool is working with a boundary that moves. An approach that produced reliable output yesterday may produce plausible-but-wrong output today, on a slightly different task, without any visible signal that something has changed.
MIT Sloan’s research on generative AI and productivity found that performance dropped when workers moved to tasks at the edge of the tool’s capability, even among experienced users. The drop was not because those users became careless. It was because the feedback loop that would normally signal a boundary had been crossed was missing. The output still looked like output. The errors were content errors, not formatting errors, and content errors require domain knowledge to catch.
How Workslop Accumulates at the Team Level
Workslop is an individual-level phenomenon that creates team-level costs. A single employee producing unreliable AI outputs is an error that can be caught and corrected. A team where several employees are operating near or beyond the capability frontier without shared quality standards creates a different problem: the error rate becomes structural, and the correction burden accumulates downstream.
MIT’s research on knowledge collapse adds a dimension to this. When employees outsource cognitive work to AI without maintaining their own engagement with the underlying reasoning, the domain knowledge required to catch capability-frontier errors degrades over time. The person best positioned to identify that an AI output has gone wrong is someone who understands the subject well enough to recognize a plausible-but-incorrect claim. If that domain knowledge is being displaced by AI use rather than supported by it, the quality control layer weakens precisely as the volume of AI output increases.
HBR’s research on AI brain fry documents the fatigue that builds in employees who spend significant time reviewing AI outputs they cannot fully trust. The cognitive load of sustained critical review at scale is real, and it has measurable consequences: more errors pass through, decision quality declines, and employees begin to disengage from the review task itself. The human performance cost of unmanaged workslop is not just downstream correction time. It is the depletion of the attentional resources required to catch errors in the first place.
The Human Performance Conditions That Contain It
The Human Performance Intelligence framework identifies three conditions that determine whether workslop accumulates or stays contained within a team.
The first is capability clarity: employees need a working model of where their specific AI tools perform reliably and where they do not. Building that model requires regular updating as tools evolve and as tasks shift, and it requires organizational support for the kind of experimentation and shared learning that builds calibrated judgment over time.
The second is psychological safety around quality. Teams where employees feel comfortable flagging AI outputs that seem wrong catch errors earlier and correct them with less friction. Where that safety is absent, employees pass on outputs they are uncertain about rather than risk the social cost of raising a concern. The quality control layer depends on the relational conditions that make honest quality assessment feel safe.
The third is review design. Workslop accumulates fastest in workflows where AI-assisted work moves directly to the next stage without a structured review step. Adapting review processes to match the pace and volume of AI-assisted production is a management and workflow design task. It requires understanding where the capability frontier sits for specific tasks and building the review step at that boundary rather than assuming errors will surface naturally downstream.
What HPI Measures at the Capability Frontier
Human Performance Intelligence approaches workslop through the intersection of its four pillars: human performance, workplace wellbeing, people management, and AI-enabled work. The capability frontier is a human performance variable, shaped by the domain knowledge, attentional capacity, psychological safety, and management practices of the people working with the technology.
Measuring workslop risk inside an organization requires looking at all four dimensions. Where is domain knowledge being maintained, and where is it degrading? Where are employees experiencing the cognitive fatigue that reduces error-catching capacity? Where are management practices creating the conditions for quality, and where are they inadvertently creating pressure to move fast at the expense of accuracy? Where are the workflows that move AI output downstream before a qualified human has reviewed it?
These questions do not have technology answers. They have human system answers, and finding them is the starting point for building the conditions that keep workslop contained.