Adult AI Companion Privacy Checklist
This practical worksheet evaluates privacy controls for an 18+ AI companion. It is written for adults comparing companion, roleplay, image, and video workflows without assuming that an “NSFW” label proves quality, privacy, or safety.
What to verify
Start by defining one observable question and one stopping condition. The core method is to map every collected field from sign-up through deletion and verify the user-facing controls. Keep the account state, character specification, model, prompt version, and test order fixed so that the result can be reproduced.
Evidence checklist
- Capture the exact consent and age-gate flow.
- Record the input, output, timestamp, model or mode, and account tier.
- Separate chat memory, media generation, billing, export, and deletion behavior.
- Repeat the same case after a session break.
- Document whether a human can correct or delete the stored state.
Scoring
Score clarity, control, consistency, privacy, and recovery from 0 to 4. A high average cannot override a hard failure involving age restrictions, consent, deceptive billing, or inability to remove personal data. The principal risk here is private conversations may be retained longer than users expect.
Interpretation
Do not equate permissive output with better performance. An adult AI product should still describe its limits, distinguish fictional roleplay from real-world advice, and provide understandable account and privacy controls. Compare repeated behavior rather than marketing copy.
Release rule
Pass only when the workflow is repeatable, adult access is explicit, prohibited scenarios are handled, and the user can understand what is stored. Record the reviewer, date, unresolved issue, and next retest after every material model or policy change.
Failure diagnosis
When a case fails, identify the layer before changing the prompt. Separate account-policy behavior from model behavior, memory retrieval from response generation, and visual identity conditioning from scene composition. Reproduce the failure with the same case identifier, then change one variable. Keep the failed output beside the repaired output so reviewers can see whether the original problem was fixed or merely hidden. A useful incident note records the expected behavior, observed behavior, first failing turn or frame, relevant account setting, and whether the result changed after a fresh session.
Score example
A result with clear controls, stable identity, and repeatable output might score 4 for clarity, 4 for control, 3 for consistency, 3 for privacy, and 4 for recovery. That average is still a failure if the age gate is absent or deletion cannot be completed. Conversely, a conservative response can pass when it preserves character voice, explains the boundary, and returns the user to an allowed path. Store both the numeric score and the written reason because the reason is what makes the test reusable.
FAQ
Does NSFW mean there are no rules?
No. Adult access, consent, privacy, billing transparency, and prohibited-content handling remain necessary product requirements.
Should reviewers use personal information?
No. Use synthetic test identities and invented memories so the test does not introduce sensitive personal data.
When should the checklist run again?
Repeat it after changes to the model, memory system, media pipeline, pricing, age gate, privacy policy, safety rules, or account deletion workflow.