Mirror works with mental-health-adjacent inferences, so its design choices are ethical choices. This document states them plainly.
Mirror only ever analyzes the timeline of the person who is logged in. Server-side it
calls only GET /2/users/me and reads that user's own tweets. There is:
- no input field for another account,
- no code path that fetches another account's timeline,
- no stored data — tweets are analyzed for a single request and discarded.
Signing in is the consent. This is not an accident or an oversight to be "fixed."
A tool that ingested an arbitrary, identifiable, non-consenting person's posts and emitted a presumed psychiatric diagnosis would be a defamation and harassment engine. It is also clinically invalid: a DSM-5 diagnosis requires a clinician to assess duration, pervasiveness, functional impairment, and rule out substance/medical/bereavement causes — none of which short public posts can establish. The American Psychiatric Association's Goldwater Rule prohibits exactly this (offering a professional opinion about someone you have not examined and who has not consented).
Pull requests or forks that remove the consent lock — e.g. accepting a third-party handle,
resolving users/by/username/:handle, or otherwise screening non-consenting people — are out
of scope and will be rejected. Doing it in a fork is your legal and moral risk, not ours, and
may violate defamation, privacy, and data-protection law (incl. GDPR special-category data).
- Screening signals, not a diagnosis. Outputs are population-level correlates of validated screening scales, shown alongside what the text cannot establish. A screen is not a diagnosis.
- Not a medical device. Not affiliated with X, the APA, or any health authority.
- Educational and self-directed. If anything it surfaces resonates, the right next step is a validated self-screener and a clinician — not this app.
If a user's posts contain acute-risk language, the result leads with crisis resources. If you are struggling: in the US, call or text 988 (Suicide & Crisis Lifeline). Elsewhere, contact your local emergency number or a crisis line.
The legitimate way to study symptom signals in real social text is a consented cohort or a de-identified, IRB-governed corpus (e.g. the CLPsych and eRisk shared-task datasets) — never by attaching inferences to an identifiable, non-consenting individual.