Technical note 03
Legibility by Design: Building for an Objection That Is Not About Usability
JDM Inc
Section 8 of the source paper contains one sentence of design implication. This note takes it seriously, because it is the reason this company builds the way it does.
The design implication in the source paper is a single sentence: if the primary barrier is identity rather than usability, interventions aimed at ease of use will underperform, and designs that make the user’s contribution legible — that preserve and display the judgment being exercised — address the actual objection.
Most of the AI tooling built in the last several years has aimed at the other target. The dominant design instinct has been to reduce the number of steps between a question and an answer, on the theory that friction is the enemy. If the framework in the paper is right, that instinct is not merely insufficient in this domain. It is pointed at the wrong variable, and in the limit it makes the problem worse.
What the objection actually says
Start with what the objection is, stated precisely.
Celniker and colleagues established across eight studies that effort functions as a moral signal: effortful individuals were judged as more moral and rewarded accordingly, even where the effort produced no additional output (Celniker et al., 2023). Effort is not treated as an input to be minimized. It is treated as evidence about the person.
Given that, the complaint “you didn’t really do it” resolves into something specific. It is not a claim about output quality — it survives perfectly well when the output is good, and the paper argues it gets stronger as the output gets better. It is a claim about what the output entitles you to say about yourself.
That reading predicts a pattern worth testing directly, and the paper flags it as one of the more tractable entry points to the whole framework: professionals concealing AI assistance in settings where no rule prohibits it. Concealment without prohibition would be strong evidence that the operative norm is about identity rather than compliance. We are not aware of that measurement having been made at scale.
If this is the objection, then a faster interface does not touch it. Removing steps removes exactly the thing the user needs in order to make a claim about their own judgment. A system that produces a perfect answer in one move and shows nothing has maximized ease of use and maximized the identity cost at the same time.
Legibility is not explainability
The word we want is legibility, and it is worth separating from a neighboring one.
Explainability, as the term is normally used, is about the system: why did the model produce this output, what features drove it, how confident is it. It is a property of the machine, offered to an auditor.
Legibility, in the sense the paper’s implication requires, is about the user. It asks a different question: after using this system, can the person say what they contributed and defend it? Does the artifact preserve the judgment they exercised, or does it absorb it?
A system can be highly explainable and completely illegible in this sense. A full attention map explains the machine and returns nothing to the person. Conversely, a system can be legible without being deeply explainable: if every assertion carries the source it came from, and the user selected, weighted, rejected and overrode those sources, then the user’s judgment is present in the artifact and recoverable from it — whatever is happening inside the model.
The design test we hold ourselves to follows from that. An answer that cannot show where it came from is not a fast answer; it is an answer the user cannot stand behind. Everything the system tells you, it should be able to show you — not because a regulator demands the citation, though one might, but because the citation is what converts a generated output into a position the user has taken.
The supporting result
There is one empirical result that bears on this directly, and it is the most important finding in the source paper’s Section 6.
Bastani and colleagues ran a field experiment with roughly a thousand secondary students. Access to a plain GPT-4 chat interface raised practice performance substantially but lowered subsequent unassisted exam scores relative to control. A version scaffolded with pedagogical safeguards produced the practice gains without the exam decrement (Bastani et al., 2024).
The careful reading matters. This is a single subject in a single school system, and it does not establish that AI use degrades thinking in general. What it establishes is narrower and more useful: that it can, that whether it does depends on how the tool is built, and that the difference between the harmful and benign configurations was a design difference.
The same study contains both the harm and the mitigation. That is the seam this company is trying to build on, and a separate note takes up the cognitive-science side of it.
Lee and colleagues’ survey of 319 knowledge workers points the same way from the attitude side: higher confidence in generative AI was associated with reduced perceived cognitive effort in critical thinking tasks, while higher self-confidence in one’s own abilities was associated with more critical engagement (H.-P. Lee et al., 2025). It is self-report and should be weighted as such. But the direction is consistent with a design in which the user’s own judgment stays in the loop, and inconsistent with one that encourages deference.
What we are not claiming
We want to be exact about the status of all of this.
We have not run a study showing that a legibility-preserving interface reduces refusal, or that it prevents the capability decrement Bastani and colleagues measured. Those are the experiments the argument implies and we have not done them. Nor does the framework license the claim that objections to AI are merely psychological — the source paper names that inversion as its central abuse risk, and on the one substantive question it examines, it concludes the worry is well founded.
What we are claiming is narrower: that the design target implied by the identity account is different from the one most of the field is aiming at, that the difference is specifiable, and that we have chosen to build against it. If the framework turns out to be wrong, we will have built a system that shows its work for no reason. That seems like an acceptable failure mode.
References
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2024). Generative AI can harm learning. SSRN Working Paper No. 4895486.
Celniker, J. B., Gregory, A., Koo, H. J., Piff, P. K., Ditto, P. H., & Shariff, A. F. (2023). The moralization of effort. Journal of Experimental Psychology: General, 152(1), 60-79.
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).
Ram, S., & Sheth, J. N. (1989). Consumer resistance to innovations: The marketing problem and its solutions. Journal of Consumer Marketing, 6(2), 5-14.