Technical note 04
Conversion and Measurement: Why Asking People About AI Recovers the Wrong Variable
JDM Inc
Proposition 5 says converts report the original concern as resolved at rates exceeding any real change in how they use the tool. If that holds, a large share of what we know about AI attitudes is measuring network position rather than evaluation.
The methodological consequence of the identity framework is more disruptive than the framework itself, and it is the part of the source paper that has received the least attention.
The claim is this: self-report about AI attitudes is collected almost entirely from people whose position is socially situated, and whose account of how they arrived at it is systematically unreliable. Not through dishonesty. Through the ordinary fact that the causal history of an attitude is not introspectively available.
The convert problem
Take the arc the paper opens with. A person refuses generative AI on moral grounds, holds the position with conviction, then begins using it daily and describes the earlier refusal as a mistake.
The comfortable reading is that evidence corrected an error. The framework’s reading is that if the refusal was socially produced, there is no principled reason to assume the reversal was not.
Bem’s self-perception theory supplies the mechanism: people infer their attitudes from observation of their own behavior (Bem, 1967). Someone whose job now requires daily use will generate a supporting attitude, and will experience that attitude as a conclusion rather than as an accommodation. Moving into a network where AI use is normal produces the same reversal that moving out of one produced the original refusal. Both feel, from the inside, exactly like learning.
This makes the convert the least reliable available witness on the question they were originally raising. Having reversed, they have strong motivation to treat the earlier concern as settled — dissonance is cheaply relieved by concluding the old self was simply wrong. But the original worry was not adjudicated. It was abandoned, and the abandonment looks like resolution from the inside.
Note what this does to the most common form of reassurance in the industry. “I used to think that too, and then I actually tried it” is treated as a strong testimonial. On this account it is a report of a network change wearing the costume of an update.
The worry outlived the converts
The reason this matters is that the concern the converts stopped worrying about has independent support.
Lee and colleagues, surveying 319 knowledge workers, found higher confidence in generative AI associated with reduced perceived cognitive effort in critical thinking tasks (H.-P. Lee et al., 2025). Bastani and colleagues, in a field experiment with roughly a thousand secondary students, found a plain chat interface raised practice performance but lowered subsequent unassisted exam scores against control, while a scaffolded version produced the gains without the decrement (Bastani et al., 2024).
Neither result establishes that AI use degrades thinking in general. One is self-report; the other is a single subject in a single school system. What they jointly establish is that the outcome is real under some configurations and absent under others — which is exactly the empirical question the moral framing was standing in front of, and which the conversion narrative quietly closed without answering.
Policy built on the wrong measurement
There is a concrete instance of this going wrong at institutional scale.
A large volume of school and university AI policy was written on the premise of a cheating epidemic. Survey data from three US high schools found self-reported cheating behaviors remained relatively stable following the release of ChatGPT (V. R. Lee et al., 2024).
The paper’s reading is that a framework mistaking an identity threat for an integrity crisis will generate rules aimed at the wrong behavior. That is a failure of measurement before it is a failure of policy: the instrument asked about norms and got norms back, then the answer was treated as a base rate.
The same error is available to any institution currently writing AI rules, including the regulated institutions we work with. The question worth forcing is whether a given rule is responding to measured harm or encoding a norm about effort. Those require different instruments, and on the available evidence the second is doing more of the work.
What a better instrument looks like
The paper’s Propositions 4 and 5 are deliberately constructed to be testable behaviorally rather than attitudinally, which is the corrective.
P4 predicts that users in identity conflict will introduce moral framing into interactions unprompted, at meaningful rates, in contexts where no rule prohibits the use. That last clause is what makes it a real test: unprompted moral framing where nothing is prohibited separates an internalized norm from external compliance, and neither a survey question nor a policy audit can distinguish those.
P5 predicts converts will report the original concern resolved at rates exceeding any observed change in substitutive versus augmentative usage. It puts a stated attitude and a behavioral measure side by side and looks at the gap.
Both are addressable with privacy-preserving aggregate analysis of real usage at population scale, of the kind demonstrated by the system published as Clio and since renamed Anthropic Insights (Tamkin et al., 2024). The access constraints are real and the paper states them: a closed pilot placed aggregate conversation data with three research groups, and the company has said it is proceeding slowly and has not determined whether the program can scale (Anthropic, 2026). Such data also cannot observe non-users or peer networks and cannot establish causation, so it needs pairing with survey and interview work.
What it uniquely supplies is the half of the question that self-report structurally cannot reach: whether what people say about AI and effort corresponds to what they do when no one is grading them.
The general design principle is the one the paper closes its implications with. Studies treating stated reasons as data about causes will recover the vocabulary of the respondent’s network rather than the mechanism of their decision. Designs that pair stated attitudes with observed behavior, or that compare the same individuals across changes in social context, are positioned to separate the two. Nothing that only asks people can.
References
Anthropic. (2026). Enabling independent research on how people use Claude.
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2024). Generative AI can harm learning. SSRN Working Paper No. 4895486.
Bem, D. J. (1967). Self-perception: An alternative interpretation of cognitive dissonance phenomena. Psychological Review, 74(3), 183-200.
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (Article 1121).
Lee, V. R., Pope, D., Miles, S., & Zárate, R. C. (2024). Cheating in the age of generative AI. Computers and Education: Artificial Intelligence, 7, 100253.
Tamkin, A., McCain, M., Handa, K., Durmus, E., Lovitt, L., et al. (2024). Clio: Privacy-preserving insights into real-world AI use. arXiv:2412.13678.