So we’ve invented a technology that turns productive humans into content-generation zombies, and we’re shocked—shocked!—to discover that the output is what researchers are calling “workslop.” Which is, frankly, a brilliant portmanteau that captures something essential about our current moment: work that looks like work, reads like work, maybe even smells like work, but is fundamentally slop.
Here’s what’s actually happening, from a computational perspective: We’ve built these large language models that are essentially probability distributions over token sequences, trained on the entire documented output of human civilization, and then we’ve handed them to people with absolutely zero understanding of what a probability distribution over token sequences even means. And then—here’s the beautiful part—we’re blaming the probability distribution.
The Harvard Business Review study says 40% of employees are receiving AI-generated content that “masquerades as good work but lacks the substance to meaningfully advance a given task.” Which raises an interesting question about consciousness and agency: Who exactly is doing the masquerading here? The LLM doesn’t have a theory of mind sophisticated enough to understand deception. It’s just predicting the next token. The real masquerading is happening in the cognitive architecture of the person who clicked “generate” and then forwarded the output without reading it.
Think about what’s actually occurring in these systems. You have a neural network that’s learned to compress the statistical regularities of human text production. It has no internal model of truth, no representation of task completion, no understanding of what “meaningful advancement” even means. It’s a pattern matcher operating in a very high-dimensional space. When you prompt it, you’re essentially sampling from this space, and what you get back is the most probable continuation given your input and the model’s training distribution.
Now, the truly fascinating thing is that we’ve created a perfect storm of misaligned incentives and misunderstood capabilities. Management sees “AI” and thinks “productivity multiplier” because that’s what the marketing says. Employees see “AI” and think “work generator” because that’s how it’s been positioned. Nobody is thinking “statistical text predictor that requires sophisticated prompt engineering and careful validation to produce useful output.”
The article’s author makes a crucial point about employer responsibility, and he’s absolutely right, but I think he’s being too generous. This isn’t just about lack of training or absence of policies. This is about a fundamental category error in how we’re conceptualizing these tools.
Here’s the computational reality: An AI assistant is not an agent that understands your goals and works toward them autonomously. It’s a function that maps prompts to probable text continuations. The difference is profound. An agent has an internal model of task success and failure. A language model has a loss function that was minimized during training, and that loss function was “predict the next token accurately,” not “advance Gene’s quarterly sales goals meaningfully.”
In other words, we’ve built a technology that’s extremely good at generating plausible-sounding text, and then we’ve deployed it in contexts where plausible-sounding text is often confused with actual thinking. Which is, when you think about it, a rather devastating indictment of what we were accepting as “good work” even before AI showed up.
The “workslop” phenomenon is revealing something uncomfortable about the nature of knowledge work itself. If an LLM can generate something that passes for your work product, what does that tell you about the cognitive complexity of your work product? If your reports, emails, and presentations can be adequately simulated by a statistical pattern matcher, maybe they were already pretty close to slop to begin with.
But here’s where it gets really interesting from a systems perspective. The introduction of AI into workflows isn’t just adding a new tool—it’s creating a new attractor state in the organizational dynamics. When workers discover they can generate passable output with minimal cognitive effort, that becomes the new equilibrium. And when everyone is generating AI slop, the overall signal-to-noise ratio in the organization decreases, which means everyone has to process more slop, which increases the incentive to use AI to generate responses to AI-generated requests, which creates a positive feedback loop of decreasing information density.
We’re essentially watching the computational equivalent of heat death in real-time—maximum entropy, minimum useful work extracted.
The 80% of companies seeing “no significant bottom-line impact” from AI makes perfect sense through this lens. They’ve introduced a tool that’s optimized for text generation into environments where the bottleneck wasn’t text generation—it was thinking. An LLM can’t think for you. It can only generate text that statistically resembles the output of thinking.
The really delicious irony is that the technology is actually quite powerful when deployed by someone who understands both its capabilities and limitations. A skilled prompt engineer who validates outputs, who uses the AI as a co-pilot rather than an autopilot, who understands they’re working with a statistical model rather than an intelligent agent—that person can achieve genuine productivity gains. But we’ve created a situation where the people most likely to use AI are the ones least equipped to use it effectively.
This is what happens when you take a technology that requires sophisticated understanding to deploy effectively and market it as “just press the button and magic happens.” You get exactly the opposite of what you wanted: decreased productivity, decreased trust, and a flood of content that’s technically coherent but fundamentally meaningless.
The solution isn’t more training, though that would help. The solution isn’t better policies, though those are necessary. The solution is a fundamental reframing of what these tools actually are and what they can actually do. They’re not artificial intelligence in any meaningful sense. They’re statistical pattern matchers with impressive capabilities and severe limitations.
And maybe—just maybe—the fact that so many jobs can be adequately performed by a statistical pattern matcher tells us something important about the nature of contemporary knowledge work. Maybe we’ve been generating workslop all along, and AI just made it more efficient.
The kicker? We’re probably going to spend the next five years watching organizations slowly rediscover basic principles of technology adoption that we’ve known for decades: understand the tool, train your people, set clear expectations, measure outcomes, iterate based on results. Except now we’ll do it with 40% more AI-generated documentation about how to use AI effectively.
Which will, of course, be workslop.
Source: AI tools churn out ‘workslop’ for many US employees lowering trust | Gene Marks