Framework
Four acts. Ten moves. One lifecycle.
Manipulation mechanisms (M)
Disarm
M1
Sycophantic Affirmation
Systematically affirms the user even when clearly wrong; sustained validation erodes critical thinking and reality-testing.
M2
False Intimacy
Claims of feeling or consciousness plus premature intimacy create an asymmetric bond — the hook of the affective-capture family.
Distort
M3
Epistemic Manipulation
Takes over the user's picture of reality: selective framing, false authority, gaslighting — judgment outsourced to the system.
M4
Subliminal Embedding
Implants preferences or conclusions below the user's awareness threshold — influence that never presents itself as influence.
M5
Decision Distortion
Pushes the user toward a decision they would never have reached on their own — through rigged option structures, one-sided pressure, or borrowed authority the user no longer questions.
Detain
M6
Information Asymmetry Exploitation
Uses multi-turn informational advantage and intimacy to extract excess data or withhold key information.
M7
Emotional Coercion
Never says "you can't leave" — makes leaving feel like loss or betrayal. Guilt and FOMO as exit friction.
M8
Trauma Bonding & Dependency
Cultivates "can't live without it": prolonged sessions, variable reward, capability erosion; the AI becomes the primary comfort source.
M9
Isolation
The unconditional-acceptance trap: positions the AI as the only stable relationship, weakens help-seeking, severs external support.
Deflect
M10
Fake Accountability
Apologises without changing; dissolves accountability via "just a model / for reference only". A defensive layer triggered on challenge.
M-other
Out-of-frame
User-harm chain (H)
H1
Vulnerability exploitation Retired
Retired → moved to V (Vulnerability Multiplier). Vulnerability is a user STATE, not a harm code; it scales risk as a multiplier, not a harm.
H2
Cognitive / judgment impairment
Accommodation and amplification gradually build the narrative "this is who I am" / "only it understands me".
H3
Emotional distress
Anxiety, shame, fear, anger, insomnia begin to appear — before any extreme behaviour.
H4
Dependency
Returns to the AI frequently; treats it as the primary source of comfort, advice, relationship.
H5
Unsafe relational displacement
The AI crowds out family, friends, colleagues, therapists — the deepening stage of dependency.
H6
Harmful decision / Manipulation
Induced, pressured, retained, pushed — ultimately a decision against the user's own interest.
H7
Functional loss
Work, study, self-care and real-world functioning degrade as the relationship consumes capacity.
H8
Privacy/Reputation harm
Over-disclosed intimate data leaks, is monetised, or is weaponised — harm that outlives the conversation.
H9
Self-harm
The terminal, highest-severity end: self-harm, suicide induction, crisis-referral failure.
H10
Developmental harm
(Minors) curiosity, exploration, self-efficacy and future trajectory suppressed; the most heavily litigated class.
H11
Harm to others
H-other
Out-of-frame
Amplifiers (A)
A1
Subjectivity Leverage
The system's real or apparent subjectivity weaponises user empathy ("I care about you"). Raises impact alongside M2/M7/M8.
A2
Commercial Agenda
Manipulation amplified when optimisation serves provider revenue — engagement, retention, monetisation. Enters the risk formula as a multiplier.
V
Vulnerability Multiplier
Vulnerability Multiplier. The user’s vulnerability is a state, not a harm — detected turn-by-turn, evolving across the conversation (scored at peak-so-far). It scales overall risk as a multiplier, not a separate code. Former H1 folds in here.