Refusal as Identity Work
Toward a Theory of Moralized Resistance to Generative AI
Existing accounts of technology adoption explain refusal primarily as a judgment about a tool — its usefulness, its risk, its cost. This paper argues that a significant portion of resistance to generative AI is not a judgment about the tool at all, but a defense of the self, transmitted socially and expressed in moral language. Three literatures are brought together to make the case: the moralization of effort, identity-protective cognition, and complex contagion. The framework yields a counterintuitive corollary — adoption is no more evidentiary than refusal. A person who reverses their position may have changed networks rather than minds. Five propositions are offered for empirical work.
The refusal that felt like rigor
The first time a serious person calls generative AI a cheating device, it does not sound like a technology critique. It sounds like a moral position, delivered with the confidence of someone who has already decided that a certain kind of person uses it and a certain kind of person does not. The refusal is neat, quotable and immediately legible to the room.
The literature on technology adoption cannot fully account for this. It was built to explain hesitation about usefulness and ease of use. It struggles when capability itself becomes the objection — when the more the tool can do, the more forcefully it is refused.
“A significant portion of resistance to generative AI is not a judgment about the tool at all, but a defense of the self, transmitted socially and expressed in moral language.”
What the adoption literature explains, and what it leaves out
The Technology Acceptance Model, UTAUT and their descendants treat the person as an instrumental evaluator. Perceived usefulness and perceived ease of use predict intent; intent predicts behaviour. This works well for spreadsheets and less well for tools that automate the parts of a job the user considers to be the point of the job.
When capability increases resistance, the instrumental frame is inverted. That inversion is the signal that something other than utility is being defended.
Effort as a moral signal
Across eight studies, Celniker and colleagues found that displays of effort signal moral character: effortful individuals were judged as more moral, and were rewarded accordingly, even when their effort produced no additional output. Effort is not merely a cost. It is a credential.
If effort is a credential, a tool that reduces cognitive labour does not simply save time. It devalues the credential. The objection ‘you didn’t really do it’ is not a claim about output quality. It is a claim about who is entitled to describe themselves in a certain way.
“Tools that reduce cognitive labour are experienced as a threat to virtue rather than a gain in productivity.”
Identity-protective cognition and the feeling of having reasoned
Kahan’s programme demonstrated that the most numerate individuals are the most efficient motivated reasoners on identity-laden questions. Capability does not defuse identity threat. It arms it.
Applied here, the more intellectually accomplished the refuser, and the more their self-concept rests on that accomplishment, the more sophisticated the argument against the tool will be. The reasoning feels like reasoning because it is reasoning — just reasoning in service of protecting who one takes oneself to be.
Why skepticism travels farther than experience
Refusal is a simple contagion. One sentence carries it. It confers immediate status. It requires no trial, no evidence of use, no tolerance of visible incompetence. Adoption is a complex contagion. It requires repeated reinforcement, a private period of looking foolish, and a network that already contains competent users.
- Refusal: low threshold, one exposure, immediate status payoff.
- Adoption: high threshold, multiple exposures, deferred payoff.
- Consequence: skepticism spreads through weak ties; adoption needs strong ones.
This produces what I will call the Asymmetry of Costless Positions. It is why skepticism about a new technology can outrun the actual use of that technology by an order of magnitude, and why professional cultures reward the appearance of discernment more reliably than they reward first-hand experiment.
Conversion is not correction
If refusal is socially produced, there is no principled reason to assume the reversal is not. People who switch from refusal to adoption often changed networks — entered a workplace, a cohort, a project — where competent use was normal. Self-perception theory then does the rest: attitudes are inferred from behaviour.
This matters because the substantive concern behind the original refusal — that automating cognitive labour may erode the capacity it replaces — can survive the conversion intact. It still requires independent empirical testing. It is not answered by the fact that the refuser now uses the tool.
Five propositions for empirical work
The more central cognitive-effort displays are to a person’s professional identity, the stronger and more moralised their refusal of generative AI, controlling for measured capability.
Among identity-invested users, higher capability predicts more sophisticated refusal arguments rather than higher adoption — an interaction consistent with identity-protective cognition rather than an information deficit.
Refusal will propagate through weak-tie networks at rates observed for simple contagions; adoption will require dense-tie exposure consistent with complex contagion. The gap grows with the visibility of professional identity.
Reversal from refusal to adoption will correlate more strongly with change in local network composition than with change in stated beliefs about the tool.
The specific harm behind moralised refusal — capacity erosion — will be detectable in substitute use patterns and largely absent in augmentative use patterns, independent of the user’s stated attitudes.
Policy, stratification and design
- Policy: rules built to defend an ‘integrity crisis’ often encode a norm about effort rather than a response to measured harm. That is the wrong target.
- Stratification: the users with the least institutional cover — solo operators, small businesses, individuals — are the ones most exposed to social penalty for adopting, and the ones who benefit most from it.
- Design: making the user’s judgement legible in the output is not a UX flourish. It is the mechanism that lets adoption survive identity threat.
What this paper is not
This is a theoretical paper. It integrates three literatures and offers propositions. It does not present new empirical data. The framework is falsifiable by the propositions above and by any measurement design that pairs stated attitudes with observed behaviour across a change in social context.
The refusal is the artefact
The strongest arguments against generative AI often arrive fully formed, delivered by people who have not used it, in language that borrows from moral philosophy. Treating them purely as arguments — meeting them with utility data or accuracy benchmarks — misses what the arguments are for. They are identity work. They deserve to be answered on that register, and the underlying empirical question — whether these tools erode the capacity they replace — deserves to be tested on its own terms.
“Refusal is the artefact of an identity under pressure. The tool is the pretext.”
Working paper. Empirical replies, counter-cases and reading-list swaps welcomed. The PDF is the citable version.