AI’s most visible markers can be the least informative
A human approved it. A company uses it. A person and an AI collaborated. All true, perhaps. None tells us what actually happened.
In brief: AI arrangements are often described through visible markers such as human approval, organisational adoption and human–AI collaboration. These markers may be factually correct while concealing how work, influence and control were distributed. To understand the arrangement, we need to trace what happened inside it.
I keep noticing how often a visible feature of an AI arrangement becomes evidence for the arrangement as a whole. A person approved the decision, so there was human control. A company uses AI, so it has adopted AI. A finished piece involved both a person and a model, so they collaborated.
None of those statements has to be false. The problem is that each identifies something easy to see and then quietly asks it to prove much more.
“A human approved it” does not tell us what the human controlled
Imagine an AI-supported system processing benefit applications. It gathers information, classifies the case, assigns a risk score, selects which evidence to display and recommends an outcome. A caseworker then approves or rejects the recommendation.
The approval is real. But by the time the caseworker arrives, the system may already have shaped which cases received attention, what information appeared relevant and which explanation looked plausible. It may even have determined which cases reached a human at all.
This was one of the problems I found in my research on agentic AI in public administration. Advisory descriptions of human final approval could coexist with prior machine evaluation, prioritisation, pathway selection, initiation or escalation. The final approver remained visible while consequential upstream choices receded into the workflow.
Approval also covers very different kinds of work. Someone who independently examines the evidence and reconstructs the reasoning is doing something different from someone checking that a generated recommendation looks plausible. Both can produce the same reassuring audit entry: human approved.
A better question is: what had the system already selected, initiated or changed before the human became involved?
“The company uses AI” does not tell us what changed
Organisational adoption has a similar problem. One employee experimenting with ChatGPT can make a firm an AI user. So can a customer-service operation whose staffing, handovers, quality checks and escalation routes have been redesigned around AI.
The label tells us that AI crossed the organisational boundary somehow. It does not tell us whether consequential work changed.
Adoption statistics can therefore sound more substantial than the activity beneath them. The denominator might be firms, even when AI touches only a small proportion of their working hours. Counting users can create the opposite distortion if a few people operate an AI system embedded in a highly consequential workflow.
Instead of treating adoption as a yes-or-no property of a firm, I would look at recurring workflows. Which sequences of work have changed? Have decision rights, verification, handovers or responsibility changed with them? How much meaningful work now passes through those arrangements?
This does not make firm-level or individual-use figures worthless. It stops them from masquerading as direct measures of organisational transformation. My broader account of AI adoption as an organisational process develops that distinction.
“The human and AI collaborated” hides the division of cognitive labour
This one is particularly obvious when I examine my own work with AI. During one recent exercise, three ideas from my research garden were selected at random. I rejected an initial combination because it was conceptually flat. In the next set, the model noticed a structural similarity that I had not seen: all three concerned visible markers concealing the distribution of work underneath.
Who produced that insight?
The source ideas came from my earlier work. My rejection decisions shaped which material remained in play. The model performed the comparison and surfaced the connection. I then judged whether the result was genuinely useful or merely sounded clever. Later, the model drafted this Note from the discussion, while I retained the decision about whether it represented my thinking and should be published.
“Human and AI collaborated” erases almost everything interesting about that sequence.
The division also changes during a project. A person may originate an observation, reject an interpretation, delegate drafting, correct conceptual drift and authorise publication. The model may retrieve, compare, challenge, draft, check and implement. Counting the final words contributed by each party would tell us very little about authorship or intellectual influence.
The better questions are more specific. Who originated the material? Who transformed it? Who selected among alternatives, detected errors and decided what survived? Which claims were verified, and by whom? Who had authority to release the result?
Trace the arrangement beneath the label
Human approval identifies a checkpoint. Organisational adoption identifies presence. Human–AI collaboration identifies participation by both. Each can be accurate while leaving the consequential structure undescribed.
The practical response is not to abandon these labels. It is to stop treating them as conclusions. Trace the work itself: what changed, what had already happened, what each participant could see, and where influence accumulated.
The polished surface of an AI arrangement often looks simpler than the activity that produced it. That simplicity may be the result of good design. It may also mean that the important mess has been edited out.