Workslop, shadow AI, vibe coding: every AI buzzword is a people problem

Workslop, coined in Harvard Business Review, for AI output that looks finished and quietly pushes the real work onto whoever receives it. Vibe coding, Andrej Karpathy's term for building by accepting what the model suggests without really reading it. The jagged frontier, from the field experiment with BCG consultants, where AI lifted performance on one task and degraded it on a task that looked identical. Shadow AI, where a study of nearly 48,000 workers across 47 countries found more than half hide their AI use from managers.

There is no benchmark for workslop. No model card reports a shadow AI score. Every one of those words describes a decision a person made. Your competitor has the same models you do, from the same handful of vendors, on the same terms, so the model is not the variable. The behaviour is.

Everyone now agrees you should measure fluency. Almost nobody says what fluency is, so it collapses back into one number on a dashboard. That number is the problem, because fluency is not one capability. It is at least five, and they fail independently.

• Prompting and direction: can they describe a task well enough that the tool can do it.
• Critical thinking and judgment: do they verify, and can they recognise a confident wrong answer.
• AI and tool literacy: which tools, how deeply, and do they know where capability ends.
• Workflow integration: is AI in the daily work as a repeatable step, or one clever thing they did once.
• Growth and team influence: do they teach it forward, or is their skill trapped in one person.

The first three are close to the four capabilities in Anthropic's AI Fluency framework, and the ladder underneath them is aligned with the UNESCO AI Competency Framework and the US Department of Labor's AI literacy work. We did not invent this vocabulary. We made it measurable per role.

Now look at what an average does to it. Two people, same overall figure, nothing else in common:

The person on the left is strong at prompting and weak at judgment. They produce a lot, fast, and some of it is wrong in ways that only surface downstream. That is not a training-needs finding, it is a risk finding, and it is exactly the profile that a usage dashboard rewards. The person on the right verifies everything and is safe to trust, but their good work lives in their own head and never becomes a repeatable step anyone else can use.

The average is not a summary of those two people. It is the deletion of the only thing you needed to know about them.

The same applies to the ladder. Absent, experimental, functional, integrated, multiplier are not grades, they are five different states with five different next moves. The one worth counting is multiplier: the people who build things other people then use. Most companies cannot name theirs, which means they cannot promote them, protect them, or copy what they did.

Here is the finding that changed how we built the product. When you ask people properly, the thing blocking more AI use is almost never a mystery, and it is usually not skill. It is one of about seven things, and each one has a completely different fix.

Read that as a budget document. If the real barrier in your operations team is access, training is theatre and a licence is the whole fix. If it is policy, people are not unskilled, they are waiting for permission that nobody has given in writing. If it is time, they are not idle, the workflow is wrong and no course changes that. If it is relevance, they sat through a demo about a job that is not theirs and concluded AI is not for them.

This is why we resist selling a score. A score tells a leader how they are doing. A barrier tells them what to do on Monday, and it is the difference between a training budget and a targeted one.

Alongside the fluency read, there are three things worth knowing about every person's toolset, and their value comes from being asked in the same conversation:

• What they actually use for real work, in their words, not what procurement says they have.
• What they have access to and do not use. You are already paying for this, and it is the cheapest recovery available to you.
• What they wish they had. This is the only honest buy signal in the building, because it comes from the person doing the work rather than the vendor's roadmap.

Set those against what the company actually provisions and the answers stop being opinions and become decisions. Paid and unadopted is a deployment problem, not a purchase. Wanted and absent is the buy case. Used but used badly is the train case. Most AI budgets get argued without any of the three on the table.

Two more reads come from the same conversation and matter more than the score. Whether the person is resistant, willing or eager, because a willing team with an access barrier is the fastest win available to you and a resistant team is a communication problem you have not had yet. And the governance read: whether they paste confidential material into public tools, and whether anything gets verified before it ships. We record that as a behaviour pattern and never the content itself.

One design decision underneath all of it: this is a conversation, not a test, and no real name ever reaches the model. A test measures test-taking, and people manage their answers when they think a score is going to their manager. The honesty is the product. Everything above is worthless if people perform for it.

Look back at the barrier figure and notice the two orange rows. Trust and fear are not fixed by any curriculum. They are what happens when a rollout lands badly: people who believe the tool is auditioning for their job, people who stopped disclosing which tools they use because the first company-wide message sounded like a warning, managers absorbing a verification burden nobody costed, teams told to be AI-first who heard be fewer.

That is not an AI measurement problem. It is a behavioural one with a price attached, and it is what Behavioral Intelligence exists to put a number on, in the unit finance already argues in. It is also why we will not sell the fluency half on its own. Measure capability alone and you get a technically fluent, quietly resentful company. Measure the behaviour alone and you understand your culture beautifully while losing on execution.

There will be another buzzword next quarter. Sort it into a capability signal or a consequence signal and you will already know which of these two things it belongs to, and who owns the fix.

The words will keep changing. The two questions underneath them will not.