Technical note 01
Capability and Resistance: Why the Best Objection Is Not the Most Informative One
JDM Inc
Proposition 2 predicts that among people who refuse generative AI, analytical ability tracks the sophistication of the argument but not the experience behind it. In high-expertise domains, this has a practical consequence: the objection you should weight most heavily is rarely the one that sounds best.
There is a habit in product work of treating the most articulate objection as the most serious one. It is a reasonable heuristic in most settings. A detailed critique usually reflects a detailed examination, and the person who can name the failure mode has usually seen it.
The identity-work framework predicts that this heuristic inverts in one specific circumstance: when the tool under evaluation threatens the evaluator’s own claim to competence. Under those conditions, articulacy stops being a proxy for examination and becomes a proxy for motivation.
The finding this rests on
The load-bearing evidence comes from outside AI entirely. Kahan and colleagues measured science literacy and numeracy against perceived climate change risk and found that greater capability was associated with more polarization, not less. The most numerate respondents on each side were the furthest apart (Kahan et al., 2012). Capability did not move people toward a shared reading of the evidence. It improved the quality of the case each side was already making.
Kunda’s account of motivated reasoning explains why this happens without anyone lying. Directional goals shape which evidence gets sought, how hard it gets scrutinized, and which conclusions are permitted to stand — while leaving the subjective experience of impartial analysis fully intact (Kunda, 1990). Identity-protective cognition names the version where the directional goal is standing within a group whose members share your commitments (Kahan, 2017). The person is not aware of reasoning backward. There is nothing to catch.
Analytical capability, on this account, is not a corrective. It is a resource. It is what motivated reasoning has to spend, and a well-funded defense is a better defense.
Why compliance work is the sharp case
Regulated-industry expertise has three properties that make it an unusually clean instance.
First, the expertise is the identity. A compliance officer’s professional standing is not organized around a credential that sits beside the work; it is organized around the specific capacity to read a rule, hold several of them in mind at once, and produce a defensible reading. That is the exact capacity a retrieval-and-reasoning system claims to assist. There is no comfortable distance between the tool and the self-concept.
Second, the vocabulary of rigor is already loaded and available. A compliance professional has a fully developed language for why a source is inadequate, why a citation does not support the proposition it is attached to, and why a confident answer without provenance is worse than no answer. All of that language is correct, and all of it is exactly what an identity-protective objection would reach for first. The framework offers no way to tell these apart from the outside — which is the point of the note, not a gap in it.
Third, the paper’s Section 2 argument bites hardest here. A tool that performs the task badly is not a threat to an identity organized around performing that task; it is a confirmation. A tool that performs it well is the threat. So in the domains where the system is most capable, an identity account predicts resistance goes up rather than down — the opposite of what a usefulness-and-ease model would forecast (Davis, 1989). Any evaluation framework that cannot produce the prediction “capability increases resistance” is missing the operative variable.
The occupational-identity precedent
Nelson and Irwin’s study of librarians and internet search is the closest available analogue, and it is worth being precise about what it shows. Librarians did not evaluate search technology on its features and reach a verdict. They responded by renegotiating what the profession fundamentally was, relocating their claimed expertise from searching — now automated — to instruction and interpretation (Nelson & Irwin, 2014).
Both the resistance and the eventual accommodation were downstream of an identity negotiation, not an instrumental one. The profession did not decide the tool was good. It decided what it was going to be now that the tool existed.
The experimental record on algorithmic advice points the same direction through its moderators. Dietvorst and colleagues found people abandon algorithms after seeing them err, even when those algorithms outperform human forecasters (Dietvorst et al., 2015). Logg and colleagues found the opposite in other designs — lay participants preferring algorithmic to human judgment — with the effect weakening specifically when the comparison was against the participant’s own judgment, and among experienced professionals forecasting in their own domain (Logg et al., 2019).
Read together, the moderators are more informative than the headline results. Reliance on machine judgment falls precisely where that judgment competes with the respondent’s own claim to competence. That is not the signature of a stable evaluative disposition toward automation. It is the signature of an identity boundary.
What this changes in practice
The practical implication is narrow and we want to state it without overclaiming.
It is not that expert objections should be discounted. Section 9 of the source paper names this as the framework’s central abuse risk, and it is a real one: an argument that explains resistance as identity-protective can be turned into a device for dismissing every criticism as psychology. The framework holds that the intensity and social distribution of objections are partly identity-driven. It makes no claim that the objections are false, and in the one place where the paper examines the substance, it concludes the central worry is well founded.
What it does change is what counts as a signal. Sophistication of argument and depth of engagement are separable variables, and in this domain they may be close to uncorrelated. The question worth asking of a critique is not how good it is. It is how much contact with the system it reflects — which is a question about hours and specific cases, and is answerable.
It also reframes what a demonstration is for. If the barrier were usability, a smoother demonstration would move people. If the barrier is the claim the output makes about the person using it, then the thing that has to change is what the output lets them say about their own judgment. That is a different build, and it is the subject of a separate note.
References
Davis, F. D. (1989). Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS Quarterly, 13(3), 319-340.
Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114-126.
Kahan, D. M. (2017). Misconceptions, misinformation, and the logic of identity-protective cognition. Cultural Cognition Project Working Paper Series No. 164.
Kahan, D. M., Peters, E., Wittlin, M., Slovic, P., Ouellette, L. L., Braman, D., & Mandel, G. (2012). The polarizing impact of science literacy and numeracy on perceived climate change risks. Nature Climate Change, 2(10), 732-735.
Kunda, Z. (1990). The case for motivated reasoning. Psychological Bulletin, 108(3), 480-498.
Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90-103.
Nelson, A. J., & Irwin, J. (2014). “Defining what we do-all over again”: Occupational identity, technological change, and the librarian/internet-search relationship. Academy of Management Journal, 57(3), 892-928.