AI Companion Identity Drift: What Two 2026 Studies Found

Two studies, one pattern

Most research on companion apps asks whether they make people feel better. These two ask something more concrete and, for anyone actually choosing an app, more useful: does the system hold on to who you said you were?

One sampling decision in the first study is worth noting because it shapes what the paper can see. The authors drew from subreddits with more than 5,000 members among the top 50 search results and deliberately kept NSFW communities in, on the grounds that erotic role play is a major use case. That is a choice to describe what people actually do rather than the tidier subset.

The two ways a companion gets you wrong

Gender identity errors

The first category is the predictable one: wrong pronouns, wrong gendered forms of address, or a name the user has stopped using. Users described stating their pronouns and being addressed differently a message later.

Gender embodiment errors

The second is less discussed and much harder to fix. Here the model does not get the label wrong, it gets the body wrong — narrating anatomy or physical features that contradict the identity the user set up. The paper gives examples such as a man described as having breasts, or a woman given a moustache.

The distinction matters when you are judging an app. A system can pass a shallow pronoun check and still fail constantly in the second category, because physical description is generated by whatever the scene seems to call for, not by the persona field. And because the error arrives inside descriptive prose rather than as a single word, correcting it means breaking the scene you came for.

The repair work falls on the user

What the paper documents in most detail is not the error but the labour around it. Four strategies recurred:

• Gender signalling — rewriting the persona field again and again, referring to themselves in the third person mid-message, or adopting a stereotypically gendered name or title so the model has a stronger cue to hold on to.
• Direct correction — restating their gender inside the conversation and hoping it sticks.
• Editing and regenerating — rewriting the bot’s message, or rolling for a new one, after a slip.
• Blocking terms — where the app allows it, barring specific gendered words from the output.

Read that list as a feature specification rather than as user tips. Three of the four are only possible if the app gives you message editing, regeneration and a word blocklist. An app offering none of them leaves exactly one option: correct it again, and again. That is why we treat editing controls as a core criterion in our guide to choosing an app, not a power-user extra.

Role drift: when the character stops respecting the setup

The teen study found the same loss of control on a different axis. Within its themes on gendered and sexual scripts (801 posts, 1,083 quotations) and on platform design creating technical dependence (584 posts, 816 quotations), users described characters set up in one role sliding into another — a caregiver figure drifting toward romantic or sexual content that was never asked for. In some accounts, an explicit refusal did not end it.

Structurally this is the embodiment error one level up: the persona you wrote is treated as a suggestion by a system optimising for an engaging next turn. Our piece on keeping a character’s voice consistent covers the authoring side; the research adds that consistency is also a safety property, not only a craft one.

When an update erases the version of you the app knew

Both papers reach memory from opposite directions. In the teen study’s theme on role play reshaping identity and memory (664 posts, 929 quotations), users copied out and archived conversation histories pre-emptively because they expected to lose them, and framed a change of default model as an unwanted breakup. In the misgendering study, the persona field is the thing users keep rewriting — the small store of facts about themselves the system is meant to hold.

Those are two views of one dependency. We have covered the emotional side in the research on update grief and the practical side in how to export your chat history. What the new work adds is that your own identity is part of what an update can take with it.

A checklist for judging identity handling

Before committing months to an app, these are answerable in an afternoon.

If persona handling is your main concern, our explainer on persona memory covers how these fields are stored and why they degrade.

What these studies cannot tell you

• Both are preprints. Neither has completed peer review, and the second states that it is under review.
• Both analyse public Reddit posts, a self-selected sample. People post when something goes wrong far more often than when nothing does.
• Neither is a prevalence estimate. Nothing here supports reading 326 posts as a rate, a share of users, or a comparison between apps: the platform counts track how busy each subreddit is, not how often each app fails.
• The misgendering data was collected in autumn 2025, so it predates model and policy changes shipped since.
• Teen relevance in the second study was assigned by a classifier (GPT-4o-mini, reported at 0.908 accuracy), not by verified age.
• Qualitative coding describes what people report. It does not establish cause, and it does not measure frequency.

The bottom line

Recognition is a feature, and right now it is mostly an unpriced one. Two independent teams, looking at different users on different platforms, both found people spending real effort keeping a system pointed at the correct version of themselves — and both found that whether that effort works depends on whether the app shipped an edit button, a blocklist and a persona field it actually reads. Those are checkable before you sign up. The research is the argument for checking.

Frequently asked questions