GlazkoCha2025AutoethnographicInsights
Kate Glazko, JunHyeok Cha et al. (2025) Autoethnographic Insights from Neurodivergent GAI "Power Users"
Bibliographic info
Glazko, K., Cha, J., Lewis, A., Kosa, B., Wimer, B. L., Zheng, A., Zheng, Y., & Mankoff, J. (2025). Autoethnographic Insights from Neurodivergent GAI "Power Users." Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25), Yokohama, Japan. ACM. DOI: 10.1145/3706598.3713690.
Commentary
The authors' claim is that GAI restores agency because the user generates the prompt and thus controls the task and amount of it. However, the evidence provided shows otherwise. For example, almost all the samples require fixing, re-prompting, and adjusting, whether the flat dialect, incorrect labeling of the feeling, or legal summary that one author does not trust and needs to verify with an actual lawyer. Thus, it is not the tool itself that allows these users to feel secure, but their expertise in questioning its results. The entire argument needs to be reconsidered: the agency mentioned in the paper is not something that GAI grants but something the user possesses and uses, in terms of "spoons."
This is precisely why the sample becomes so crucial for the study. It is not just a limitation, but the eight computer scientists are exactly those users who have enough skills to audit the model. Therefore, the study can only observe the version of GAI usage that actually works. The person who relies on the model to the same extent but cannot determine when it goes wrong and needs protection the most is structurally excluded from the analysis.
Excerpts & Key Quotes
The Machine That Names Your Feelings for You
- Page 9:
"Sometimes I don't even know what emotion exactly I am feeling... sometimes it will pick words that better describe my insides."
Comment:
It is easy to read this as an assistive device helping someone communicate, but there is something more troubling underneath. The model is not conveying the author's feelings; it is telling him what those feelings are. And it has no access to his inner world, only to the vast average of how people on the internet tend to describe theirs. What it offers back is a script: the standard, most statistically likely emotion for a situation like his. For someone who genuinely struggles to name their inner states, that script can feel more articulate than anything they could produce, which is exactly the danger. The one thing a person cannot check the model against is their own mind. So when the generic label fits well enough, it quietly replaces the harder, messier work of feeling out what is actually there, and the author has no way to notice the substitution, because the model's version arrives sounding clearer than his own
Preferring the Machine to the Friend Who's Available
- Page 10:
"the friends I have in real life–they can't keep up as well."
Comment:
Most of the paper depicts GAI as a replacement for support that does not exist, no accessible peer, no accessible psychologist, mandatory reporting requirement making confession risky. This is not the case here, and it is a lot more disturbing one, peer is accessible but neglected, because machine is more patient, less tiresome and always impartial. Authors see this as an advantage and, indeed, it is for the user. What is left unexplored is the price to be paid on the other side of the balance sheet, relationship that improves because it asks for nothing and if an assistance that you do not have to pay back makes all the others seem like too much effort.
Forced Personal Responsibility
- Page 14:
"GAI use places the burden of adaptation onto 'power users' themselves, reinforcing normative expectations rather than challenging the systems necessitating the adaptions."
Comment:
This is the paper that confesses itself, and it should be pushed further than the authors do. Each and every successful result in the experiment comes from facilitating the ability of one neurodiverse person to fit into the oppressive environment. The system that forced all this remains unchanged; there is nothing to question about it. Thus, not only does the tool fail to help overcome the root cause of the problem, but it also decreases its urgency: the better people cope on their own, the less attention is drawn to the problem. In GAI in this case, we see not so much an accessibility solution as a way of coping with the issue privately.