Discovery in a mainstream AI Social App should help a person choose an AI personality, not manufacture an implied friendship. This article proposes a transparent recommendation and user-control framework for creator-authorized digital twins that interact through text, voice, live video, photos and short clips. It distinguishes real user choice from vanity engagement metrics and turns the principles of the NIST AI Risk Management Framework and the FTC’s dark-patterns report into practical product tests. The design below is an original proposal, not a claim that any specific app already ships every feature.
Discovery is a relationship decision, not a ranking trick
Discovery screens are often measured by the speed with which a person opens a recommended profile. That metric is easy to inflate. An AI social app can push a familiar-looking avatar, place the largest portrait first, and use language implying a pre-existing relationship. None of those choices proves that the recommendation is useful. A healthier unit of analysis is an informed introduction: the person can identify the AI character, understand the likely reason for its appearance, and decide whether a conversation is worth starting. Separate voluntary follows from system recommendations and paid promotion in both the interface and event ledger. The FTC’s 2022 discussion of dark patterns explains why disguising ads and steering privacy decisions can undermine meaningful choice. The product response is not to eliminate recommendation, but to expose its inputs and boundaries.
A compact reason card should answer three questions
Before a first chat, show what the digital twin represents, what kind of content the creator has authorized, and why this particular account is being suggested. A reason might be a declared interest such as photography, an explicit topic follow, or a user-selected language. Do not claim that two people are personally compatible merely because an opaque model assigned a high score. A practical reason card can say “Suggested because you follow creator interviews,” while a separate AI label clarifies that replies are generated by a digital character. The card should link to creator identity details and distinguish verified ownership from a marketing statement. In moderated experiments, ask participants to paraphrase the reason before they tap; failure to describe it is evidence that the presentation needs revision, regardless of click-through rate.
Treat contact, photos and private memory as separate permissions
A discovery result does not authorize access to a user’s address book, private photographs, microphone or persistent chat memory. These data types have different consequences when shared. A suggested AI personality may need only a broad preference signal to appear in the list; opening a conversation should not silently upload old messages or photographs. Create separate consent states for recommendations, cross-device history, voice interaction, and media exchange, with default settings that can be inspected. Record a timestamped permission version and which feature used it, without putting private content into analytics logs. NIST’s 2023 AI RMF frames risk management as an ongoing process rather than a one-time checkbox. Applied here, a permissions design should be tested after account migration and feature upgrades, not only on launch day.
Offer a meaningful way to correct a bad suggestion
Recommendation quality is revealed when it fails. The user should be able to hide a profile, mute a topic, block a character and report impersonation without losing control of unrelated conversations. These actions must have distinct semantics: “hide” changes local ranking, “block” prevents further interaction under the defined identity boundary, and “report” starts a review case. The recommendation pipeline needs to consume those events quickly, including cached results and notification queues. In an acceptance test, block a character, reopen the app on another device and confirm that direct discovery and suggested notifications respect the rule. Track how frequently a blocked account reappears after a cache refresh, because a zero-looking interface with recurring background suggestions is not a genuine control.
Measure informed starts, reversals and lasting usefulness
An experiment should not reward a design simply because it generates more first messages. Build a funnel from reason-card exposure to profile comprehension, chosen first conversation, voluntary return and user-reported relevance. At the same time, monitor quick exits, hide actions, blocks, complaints and repeated prompts to change permissions. Segment by new versus returning users, accessibility settings and language without publishing small private cohorts. Predefine the observation window so a team cannot move its goalposts after viewing the results. For example, a variant that improves first-chat starts by ten percent but doubles immediate hides might be less useful; this is a hypothetical illustration, not a reported product result. Review the combined outcome rather than privileging one easily optimized number.
Make an auditable launch checklist for editors and engineers
Write the display copy, recommendation features, consent boundaries and rollback conditions into the same launch record. A reviewer should be able to reproduce the suggestion using a synthetic account and identify whether a paid placement is marked. Ask a separate accessibility reviewer to check screen-reader order and whether the AI identity label survives compact layouts. Engineers should version reason templates so an old explanation is not paired with a new ranking model. Editors need to verify that profile photographs and voice likenesses are used with the creator’s authorization. For the first release, use a conservative list of explanation types and expand only after user comprehension and complaint data are stable. The objective is a discoverable AI social experience that grows through clarity and relevant connections, not through concealed persuasion.
Sources
NIST AI Risk Management Framework 1.0 (26 January 2023); FTC Bringing Dark Patterns to Light (September 2022).
