AI is meant to take over routine work, cut costs and speed up decisions. But better technical performance does not by itself produce a better social order. To understand resistance to AI, we therefore need to ask: What improves, for whom, and which future possibilities for action are given up in return?
Consider a public agency introducing AI to process applications. The system summarises documents and prepares decisions. Staff still review and sign off on them. Applicants could receive answers sooner, while staff could gain more time for difficult cases.
Management could, however, use all the time saved to increase caseloads. The opportunity for review would then shrink, while responsibility remained unchanged. If staff object to this change, they are not necessarily disputing the usefulness of AI. They are disputing whether the chosen use of that benefit amounts to an improvement.
An old confusion
From 1811 onwards, English textile workers destroyed weaving and knitting machines. These machine-breakers, also known as Luddites, fought against falling wages and the devaluation of skilled work, among other grievances, in the early nineteenth century. They were themselves familiar with machinery. Their conflicts therefore cannot be reduced to fear of unfamiliar technology.
The historical comparison does not justify destroying machines. It brings a distinction into view: a way of producing goods can become more productive while those who previously made their living from it lose their livelihoods. The technical gain does not answer the social question.
The public-agency example concerns a different conflict. Yet pointing to more completed cases would miss the objection here too. Shorter waiting times, sound decisions, viable working conditions and public expenditure all belong in the assessment. None of these concerns disappears because the software works faster.
How today’s gain changes tomorrow
The Path Ethics analysis follows this conflict beyond the immediate distribution of benefits. It asks whether the new way of working also preserves the capacity to recognise its later effects and continue shaping how the system is used.
If staff conduct fewer independent reviews and junior colleagues have little opportunity to learn professional judgement, the agency might initially accomplish more while becoming less able to make its own decisions over time. As its expertise declined, so might its influence over the technology provider. The right to sign off would remain, but exercising it responsibly would become harder. This is a possible development to investigate, not an outcome already proven by the introduction of AI.
The technology itself is not exempt from scrutiny. The information it makes accessible and the corrections it permits shape the possibilities for action. Its purposes may also be problematic: a professionally competent agency could reliably operate a procedure whose criteria do not do justice to applicants. Good oversight does not make an unsuitable purpose right.
For alignment, this means that reliably directing AI towards existing requirements is not enough. The collaboration must also allow those requirements to change in response to experience that provides grounds for revision.
What the introduction would have to demonstrate
In the agency example, staff would have to be shown to be capable of recognising and correcting errors before the system’s use was expanded. Independent information, sufficient review time and opportunities to learn therefore belong in the deployment plan. If higher caseloads can be achieved only by removing these conditions, the target must change.
Applicants’ experiences must also be able to feed back into the process. If complaints show that certain life circumstances are systematically misclassified, correcting individual decisions is not enough. The categories or rules governing use must then be reviewed. Switching systems or discontinuing their use must remain possible without bringing essential services to a halt.
This requires neither unchanged occupations nor agreement with every criticism. Resistance can also defend the advantages of a particular group. But anyone eliminating tasks must explain how learning, responsibility and occupational transitions will be supported in future. A general promise of prosperity is not enough.
Acceptance of AI is therefore not the goal by which its introduction should be judged. What matters is whether its benefits are embedded in an arrangement that enables those affected to continue exercising judgement, acting and bringing about change. Whether the technology meets this requirement must become evident over time, not merely in its performance record.