Two people sit in front of the same language model. Both type a question. The first writes: “Give me a summary of this topic.” The second writes: “I’m working on a comparison between X and Y for audience Z. What are the three most structurally relevant differences for this audience, and where are the limits of the comparison?”
Both get an answer. The first sounds plausible; it is generic, superficial, interchangeable. The second is precise, differentiated, useful. Both users consider their answer good. The difference is not in the tool. It is in the question.
This is not an edge case. It is the structural principle of AI use.
A Mirror That Amplifies
The common metaphor is the hall of mirrors: AI shows you who you are. That is too passive. A mirror shows what is there. AI amplifies. It reinforces what the user brings. Good questions produce brilliant results; poor questions produce plausible-sounding mediocrity.
The horror is not in seeing yourself. It is in not recognising yourself in the output while being unable to judge its quality.
The so-called Dunning-Kruger effect describes a possible distortion of self-assessment under low competence: missing knowledge can itself make it harder to judge one's own performance. AI gives this effect a technological lever. Those who cannot ask good questions get poor answers but consider them good. This creates a competence illusion that stabilises itself.
Access Is Not Freedom
Anyone with internet access can use a language model. No admission, no exam, no prerequisites. Path Ethics calls this *nominal autonomy*: the formal freedom to do something.
*Structural autonomy* is something else. It describes the actual ability to use the tool competently: to ask the right questions, to critically evaluate the output, to know the system’s limits.
Path Ethics developed this distinction for a different context. A platform worker who takes on jobs through Fiverr has nominally maximal autonomy: no supervisor, flexible hours. Structurally, they are minimally autonomous: the algorithm determines their visibility, customer ratings determine their future.
AI use follows the same pattern. Nominal autonomy is identical: everyone has access. Structural autonomy diverges maximally.
The Invisible Divide
This is where the societal stakes lie. The divide does not open between AI users and non-users. It opens between competent and incompetent users. And it is invisible, because both groups use the same tool.
A company deploying AI without the ability to judge output quality systematically produces mediocre results at high speed. The speed is misread as a productivity gain. This is not hypothetical: in the group of executives and IT/technology professionals surveyed by Pluralsight, a substantial share reported abandoned or failed AI initiatives associated with skills gaps. In a Gartner survey of employees at organisations already using AI in supply-chain work, 94 percent expressed openness to AI; 36 percent were able, according to the survey, to integrate it into existing workflows.
At a societal level, this becomes a division harder to address than the job losses everyone discusses. Job loss is visible, measurable, politically negotiable. The competence divide is invisible because those affected do not know they are affected.
What This Means for the Debate
The public AI debate orbits two positions: AI destroys jobs (pessimism) or AI creates more than it destroys (optimism). Both miss the problem.
The relevant question is not whether AI creates or destroys jobs on balance. The relevant question is: who can use the tool in ways that strengthen them, and who is pushed further back by their own incompetence without noticing?
Regulation and competence-building address different problems. For the competence gap described here, regulation alone is insufficient; building users' own judgment acts directly on their structural autonomy.
Path Ethics frames this as a diagnostic question: does an intervention support the feedback architecture of the affected system, or does it replace it? Labelling requirements and bans can limit risks, but they do not replace users' own judgment. Competence-building, by contrast, directly targets the ability to assess outputs independently.
The Missing Counter-Thesis
A one-sided analysis would be incomplete. AI also enables an exceptionally broad democratisation of cognitive tools. Those who previously lacked access to lawyers, translators, tutors, or advisors now have it. In lower-income countries, where AI exposure is lower, AI optimism dominates (Stanford HAI, 2026).
Fear is a phenomenon of industrialised nations. Opportunity is global. Both are true. Path Ethics calls this a *mixed coupling profile*: constructive coupling (democratisation) and destructive coupling (competence amplification) in the same system. Real systems almost always have both.
The fault line does not run between AI advocates and AI opponents. It runs between those who understand the competence divide and those who do not see it.