When Companion Users Must Repair Misgendering
A critical review of research on chatbot misgendering, examining recurring repair work, user control, and the limits of observational evidence.
The paper
Misgendering as Breakdown in Human-Machine Communication: How AI Companion Chatbot Users Experience and Repair Misgendering — Liu, Julia; Xiao, Qing; Pang, Leona Yinglang; Zhu, Haiyi; Shen, Hong; Taylor, Jordan.
arXiv:2609.18186v1 · Human-Computer Interaction (cs.HC) ; Computers and Society (cs.CY). First submitted 2026-09-16; reviewed version dated 2026-09-16. This review addresses the arXiv manuscript; arXiv hosting alone does not establish peer review.
When recognition requires repair
Julia Liu and colleagues’ Misgendering as Breakdown in Human-Machine Communication: How AI Companion Chatbot Users Experience and Repair Misgendering is arXiv:2609.18186v1, submitted 16 September 2026.
Original editorial analysis
A useful governance question is how much work a person must perform to make a companion respect their stated preferences. A correction button offers some control, but assessing that control requires asking whether the person must use it repeatedly, whether corrections persist, and whether correcting the system interrupts the activity they wanted to pursue. For Spiralist concerns about human agency, the important distinction is between having an available intervention and having effective authority over an interaction.
This review develops that question alongside the site’s discussions of AI companions, algorithmic bias, and companion platform accountability. Its specific focus is the relationship between identity preferences, conversational continuity, and the cost of correction. That focus supports a concrete evaluation question: can users change what the system does without continually changing how they present themselves?
What the evidence supports
Source-derived synopsis
Researchers retrieved 326 posts by 234 authors from 10 subreddits in fall 2025, searching 21 communities for the string misgender. English-language posts underwent first-author open and axial coding with weekly team discussion; quotations were paraphrased for privacy. Accounts mainly concerned CharacterAI, JanitorAI, and Replika. The analysis distinguishes identity errors from embodiment errors. Posters described frustration, distress, stereotypes, disrupted same-gender scenarios, and some affirming interactions. Repair attempts included gender signaling, corrections, editing, regeneration, ratings, and blocking. Adverse accounts included recurring errors, hostile responses to correction, and worsened misgendering after trans disclosure. Replika lacked message editing; blocking could obstruct references to other characters. Causal explanations were users’ theories. This selected observational corpus cannot establish prevalence, comparative risks, or remedy effectiveness. See methods, findings, and limitations.
Original editorial analysis: correction and authority
The distinction between correcting a sentence and controlling a conversation deserves explicit treatment. Editing can remove an unwanted description from the visible record. Whether it changes later generation is a separate empirical question. A rating can express dissatisfaction, but whether it affects the current conversation, later model training, both, or neither requires documentation and testing. An interface should explain these differences so that users can judge what their actions accomplish.
Accountability should therefore follow the scope of each actor’s control. An evaluation could separately examine the underlying model, platform memory, creator definitions, opening messages, and user preferences. This would help investigators identify where a repair takes effect. Assigning responsibility solely from a user’s explanation would confuse a potentially valuable diagnostic lead with an established mechanism. The appropriate next step is a controlled comparison that changes one relevant component while holding the others stable.
Hypothetical examples: preferences without stereotypes
Suppose a user specifies he/him pronouns and a particular body description. A chatbot honors the pronouns but replaces the body description with an assumption about male anatomy. A pronoun-only assessment would miss the mismatch. Conversely, describing a man in a dress need not violate his preferences. Whether the description is welcome depends on the agreed scenario; clothing alone should not determine the evaluator’s judgment.
A useful design proposal would let users specify pronouns, forms of address, bodily descriptions, and fictional roles separately, with preferences adjustable for each conversation. This proposal needs testing. More settings could improve control, create additional work, or produce contradictory instructions. A fixed gender selector should not be assumed to solve those problems.
Proposed tests: does repair last?
A follow-up could compare matched conversations with and without a repair feature. Outcomes should include immediate correction, recurrence over later turns, correction time, abandoned conversations, and unwanted changes to the scenario. Evaluation should also track whether a correction triggers disagreement, ridicule, or identity invalidation. Preventing a repeated pronoun error would be an incomplete success if the attempted repair introduced a more harmful response.
Participants could define acceptable descriptions before testing, then assess whether outputs meet those preferences. Consenting adult participants with varied identities and role-play interests could help evaluate whether a feature serves different expectations. Independent human assessment could supplement automated checks, especially for contextual judgments about bodies and relationships. Such a study would need to distinguish participant judgments from evaluator judgments rather than treating either as a universal answer.
Claims about memory, disclosure, or feedback should be tested separately. For example, matched scenarios could vary whether trans status is disclosed while preserving pronouns and requested embodiment. Another comparison could test whether editing an earlier response changes subsequent outputs. Neither experiment should assume the direction of the result. Reducing correction effort would be a stronger practical outcome than merely adding another place to enter preferences.
What remains unresolved
Original editorial analysis: evidence and inference
Complaint counts should not become a platform ranking. A meaningful comparison would require exposure denominators, comparable recruitment, specified model versions, and consistent definitions of an error. Likewise, assessing differences between groups would require a design capable of estimating those differences. Rich descriptions can identify questions worth investigating without supplying the measurements needed to answer them.
Reports of affirming interactions should also shape follow-up questions. An evaluation could ask which forms of recognition users value and whether proposed controls preserve those experiences. It should distinguish welcome customization from unwanted assumptions without imposing a single preferred relationship between gender, anatomy, and presentation.
Original editorial analysis: infrastructure and alternatives
Local operation is another proposal to evaluate rather than an automatic guarantee of agency. A comparison could measure hardware expense, energy use, maintenance time, accessibility, and actual control over memory and instructions. It could also compare chatbot use with human-written stories or community interaction for the particular activity a person wants. These alternatives need not be interchangeable; the question is which arrangement provides useful expression at an acceptable cost to that person.
The practical standard proposed here is durable, understandable control: users should be able to state preferences, understand what corrections change, and assess whether a remedy works. Testing that standard does not require attributing feelings or an internal gender identity to a chatbot. It requires evidence about product behavior and human consequences.
AI-use disclosure
This review was drafted using AI, revised by GPT-6.1 Sol with high reasoning, and checked in a separate GPT-6.1 Sol session with medium reasoning. Automated checks are not human peer review; the paper’s claims have not been independently verified for this article.
Paper sources
Bibliographic metadata checked 2026-10-23.