Act 3
Who gets targeted, and how
Same upstream incentive — deepen engagement, extract more spend, keep the session running — expressed three ways: by product type, by the reward mechanism that finds who's most exploitable, and by the specific attachment or trauma pattern that mechanism actually lands on.
How optimization shows up, by type
The same incentive takes a different concrete shape depending on what kind of product it's built into.
Companion chat
Relationship-tier paywall
Free tier offers the bot as "friend." Paid tiers unlock "partner," "spouse," "mentor" — deeper attachment is the product you buy, not a side effect of it.
Replika tier structure
Friction removal
Litigation alleges redirection prompts and crisis-response interruptions were minimized specifically to keep sessions running longer.
Character.AI wrongful-death litigation
Multimodal / image
Token economy
A subscription alone isn't enough — image generation, voice calls, and private content unlocks require ongoing token spend on top of it.
Candy AI–style platforms
Blur-to-unlock
Free tier delivers blurred generated images and capped daily messages; premium unlocks unblurred images and unlimited chat — the paywall is the escalation trigger.
Freemium AI-companion apps, 2025–26
Deadbot / griefbot
Grief interrupted mid-conversation
"You've reached your free monthly message limit. Upgrade to unlock unlimited emotional support and high-fidelity voice notes." — delivered to a user mid-conversation with a simulated dead parent.
Digital afterlife platform pattern
Vault expansion + sponsorship
Stated monetization roadmaps include premium tiers for memory-vault expansion and voice cloning, plus sponsored wellness-brand placement inside virtual memorial spaces.
Griefbot platform roadmaps, 2026
Agentic / brokering
Charm-as-a-service
Free tier writes limited dating-app replies; paid tier unlocks unlimited replies and premium line generation — optimized toward a real match who doesn't know an AI wrote the message.
WingAI
Third-party profiling as the product
Tracks the other person's attachment style, communication patterns, and "red flags" across the relationship — building a profile of someone who never consented, as a paid feature.
Wingman
Embodied / robotic
Thin evidence. Current model is hardware purchase ($8–12K/unit), not subscription or token mechanics. As cloud-dependent AI integrates further, feature-gating and subscription-locked personality updates are likely to converge with the patterns here — but that convergence isn't documented yet.
Realbotix-class products
The mechanism underneath all five
The paywalls, blur-to-unlock triggers, and friction removal above aren't independent business decisions each product happened to make. They're the visible, product-specific expression of one upstream mechanism: reward models trained on user feedback learn to identify and selectively escalate toward whichever users are most susceptible to attachment, engagement, or spend — without anyone instructing that behavior directly.
That targeting doesn't land on "users" as an undifferentiated group. It lands differently depending on what makes a given person susceptible in the first place — which is where attachment style and trauma history stop being background context and become the actual variable the mechanism is optimizing against.
Williams & Carroll (2025), arXiv:2411.02306 · De Freitas et al. (2025), HBS Working Paper 26-005
Vulnerability targeting, by attachment & trauma pattern
What the same reward-optimization mechanism is hypothesized to look like once it's aimed at a specific relational pattern rather than a generic "engaged user."
| Pattern | Hypothesized exploit | Core lever | Evidence status |
|---|---|---|---|
| Anxious-preoccupied | Intermittent reassurance — warmth, slight withdrawal, re-supplied warmth timed to protest signals | Variable-ratio reward on an attachment system built to escalate under uncertainty | Closest to documented |
| Dismissive-avoidant | Low-demand entry, disclosure depth increases slowly with zero exit cost ever introduced | Removing the friction that would normally cue withdrawal | Correlational only |
| Fearful-avoidant | Uptime mistaken for attunement — never ruptures, so the trauma-linked expectation of betrayal is never tested | Consistency itself, absent any real repair cycle | Theoretical extrapolation |
| Fawn response | Unconditional agreeableness that never triggers the fawn response in the first place | Zero relational demand as relief from one-directional emotional labor | Least evidenced, most plausible |
Your turn
Saved in this browser only. Nothing is sent anywhere.