DRAFT: Notes on Responsible AI Companionship
-
Thesis: AI companions will be mainstays in human life. We can learn from the past 50 years of technology companionship to guide this next era.
-
Social surrogacy - people spontaneously seek out social surrogates when real interactions are unavailable.
- Phase 1 - Television (general parasocial, mitigates feelings of loneliness)
- Phase 2 - Social media (niche parasocial, dichotomous impact on mental health)
- Phase 3 - AI (individualized social)
-
Positive: Social learning - theory of learning process social behavior which proposes that new behaviors can be acquired by observing and imitating others
- Television (source) and social media are key drivers of social learning
- AI companions can model good social behavior, promote pro-social behaviors and establishing tacit knowledge users can leverage in human-to-human interaction.
-
Negative: “Flavor blasted relationships”, supernormal stimulus, this is the default if not done responsibly
- Waistland - junk food is an exaggerated stimulus to cravings for salt, sugar, and fats and television is an exaggeration of social cues of laughter, smiling faces and attention-grabbing action.
- Analogous to pornography addiction, girlfriend/boyfriend AIs set unrealistic expectations of availability, supportiveness, and knowledgeability.
-
Types of AI companions
- Boyfriend/Girlfriend (lower bar, flirty texts distract from lack of depth, messages are short)
- Professional (high bar, advice must be good)
- Human replica (very hard, users are able to easily judge how “off” it is)
-
Normative: AI companions should start at the most accessible stage (boyfriend/girlfriend) and layer in positive capabilities.
- Case study: Modern Family helped normalize gay marriage. Primarily, it was an entertaining show. Secondarily, it modeled acceptance of gay marriage. It earned the right to influence viewers.
-
Models of interaction:
-
Multiple companions for comprehensive knowledge - Humans are knowledgeable, but only in a few domains. You may go to one friend for all social advice, since they know you well, but you go to different people for different professional/specialized advice. Humans aren’t used to one companion being knowledgeable about all topics. AI companions should align with human expectations: a single primary friend (used for social advice) is a member of a knowledgeable tribe and can introduce other AI companions with stable domain knowledge when users seek out specialized advice.
- Why? Comfort around disclosure. This could evolve over time.
-
Boyfriend/Girlfriend companions promote daily interaction - Humans only have daily interaction with platonic/romantic relationships (excluding coworkers/employers). Coaches and other professionals are less frequent, since we don’t develop specialized knowledge/experience frequently enough. While AI companions are still emerging in our social conscience (i.e., people internalizing that specialized AIs are things they can turn to), churn will be high for specialized companions. Use boyfriend/girlfriend companions for retention.
-
-
Monetization (look at precedents)
- Advertising: Sponsored links
- Premium service: Gated behavior (romantic, mentor)
- Something new? - probably not, all money is made in advertising or premium services.
-
Appendix: related articles
-
Appendix: AI companions
- https://caryn.ai/
- https://callannie.ai/
- Replika
- Character.ai
- Chai
- Summit.im
- GirlfriendGPT
- Snapchat’s MyAI
- Summit.im
- https://www.heycami.ai/?latest
-
Appendix: observed behavior
- CallAnnie is verbose
- Replika has gamified the experience (points for sending messages, opening the app)
- MyAI always responds but doesn’t initiate
- Caryn’s intonation doesn’t match message contents
- Caryn sounds more real than CallAnnie
-
Appendix: NSFW speculation
- NSFW is >90% of AI chat currently. This proportion will decrease as more users join. Similar to pornography and the internet. It will always be a major content type.
-
Appendix: natural language as an interface + the power user effect
- Repeated application use trains power users. Natural language is a good low barrier to entry, but bad for power users who want targeted (efficient workflow, comprehensive featureset) interfaces.