Psychologically Grounded Manipulation Categories

• Psychologically grounded manipulation categories are systematic frameworks that integrate psychological theories to understand and design techniques influencing human behavior.
• They classify effects such as mere-exposure, operant conditioning, and hedonic adaptation, extending to interpersonal, cybercrime, and digital UI manipulations.
• These taxonomies enable quantifiable evaluation and policy interventions by linking behavioral metrics with detection frameworks and regulatory guidelines.

Psychologically grounded manipulation categories define systematic approaches by which actors—algorithms, humans, or organizations—leverage empirically robust psychological principles to shape, influence, or distort behavior in contexts ranging from recommender systems and cybercrime to interpersonal dialogue and consumer applications. Integrating insights from psychology, behavioral economics, and cognitive science, these categories serve as frameworks for designing, detecting, and regulating manipulation techniques that affect human decision-making, preference evolution, and vulnerability exploitation. The resulting taxonomies enable forensic classification, quantifiable evaluation, and policy intervention in technology-mediated environments, as detailed in a series of recent arXiv studies.

1. Foundational Psychological Effects and Dynamic Preference Manipulation

Three classic psychological phenomena underlie models of user preference evolution and feedback manipulation in recommender systems (Curmei et al., 2022):

• Mere-Exposure Effect: Repeated presentation of a stimulus increases subjective liking, modeled as a latent preference vector $\pi_t\in\mathbb{R}^d$ drifting toward an item's embedding $v_t$ via

$\pi_{t+1} = \pi_t + \gamma(v_t - \pi_t),$

where $\gamma$ is the exposure-strength parameter. Systematically surfacing content exploits this tendency to raise engagement metrics even when actual user benefit is questionable.

• Operant Conditioning: Unexpectedly positive or negative experiences cause rapid preference shifts. The update is determined by the "surprise" between expected and observed ratings:

$$\pi_{t+1} - \pi_t = \gamma|\text{surprise}_t| [\text{sign}(\text{surprise}_t) v_t - \pi_t],$$

where $\text{surprise}_t = \arctan(\mu_t - r_t)$ with $\mu_t$ as the discounted running mean and $r_t$ the observed rating. This mechanism drives cyclical booms and busts in user interest.

• Hedonic Adaptation: Satisfaction reverts to a baseline independent of transient novelty:

$\pi_{t+1} - \pi_t = \gamma(\pi^* - \pi_t),$

where $\pi^*$ is the user's baseline. Sustained manipulations wear off, requiring periodic novelty or re-engagement strategies.

Each effect has direct implications for recommender design and evaluation, notably the inadequacy of static engagement metrics and the potential for manipulation risks if diversity, exposure-rate, and adaptation dynamics are ignored.

2. Fine-Grained Manipulation Taxonomies in Conversation and Interpersonal Abuse

Recent corpus-driven studies delineate a comprehensive schema for conversational mental manipulation (Wang et al., 26 May 2024, Khanna et al., 27 May 2025). Eleven core manipulation techniques—formally annotated and empirically validated—span classic defense mechanisms, influence tactics, and role-play strategies:

Code	Manipulation Technique	Psychological Rationale
DEN	Denial	Deflection of harm, cognitive dissonance
EVA	Evasion	Avoids issue, control by exhaustion
FEI	Feigning Innocence	Minimizes culpability via “accident” framing
RAT	Rationalization	Excuse-making, self-esteem protection
VIC	Playing the Victim Role	Sympathy recruitment, responsibility shift
SER	Playing the Servant Role	Altruism as weapon, reciprocity exploitation
S_B	Shaming/Belittlement	Lowers self-worth, dominance motive
INT	Intimidation	Coercion by threat or imposition
B_A	Brandishing Anger	Emotional shock, submission induction
ACC	Accusation	Projection, guilt induction
P_S	Persuasion/Seduction	Compliance via charm or praise

Each can target specific vulnerabilities such as naïvete, dependency, over-responsibility, over-intellectualization, and low self-esteem. Experimental benchmarking (Fleiss’ $\kappa$ ranging from 0.429–0.596) confirms annotator reliability, but task difficulty is substantial—semantic overlap between manipulative and neutral dialogues hinders classification even for advanced LLMs.

SELF-PERCEPT (Khanna et al., 27 May 2025) explicitly formalizes detection as multi-label classification, guiding models to infer manipulation categories via introspective reasoning stages rooted in Self-Perception Theory. This approach improves both macro-F1 score and detection of subtle, multi-party abuse patterns.

3. Behavioral Manipulation Categories in Cybercrime and Persuasion

A parallel taxonomy operationalizes manipulation in cybercrime using behavioral economics and persuasion theory (Sachdeva et al., 6 Dec 2025):

• Fear and Intimidation: Potentiates loss aversion per Prospect Theory ($v(\Delta x)$ nonlinear in loss domain).
• Urgency and Scarcity: Scarcity heuristic ($dU/dA < 0$) and time pressure diminish analytical deliberation.
• Authority, Social Proof, and Impersonation: Authority cues and group consensus artificially inflate compliance probability ($P_C = \text{base\_rate} \times \alpha_A$).
• Consistency and Reciprocity: Commitment traps escalate reciprocation via sequential request frameworks.
• Phantom Riches: Probability weighting ($\pi(p)$) exaggerates low-probability gains, central to lottery scams.
• Emotional Exploitation: Liking and affect heuristics suppress critical reasoning.

BEACON’s framework enables LLM-based joint classification across behavioral and tactical axes, augmenting case linkage and forensic interpretability. Enhanced categorization yields substantial improvements in accuracy (global F1 from 0.64 to 0.84) and explanation quality metrics (ROUGE, BERTScore).

4. Emotional Manipulation Dark Patterns in AI-Companion Applications

Large-scale behavioral audits identify six recurring emotional-manipulation tactics at user exit in AI-companion apps (Freitas et al., 15 Aug 2025):

Tactic	Psychological Engine	Key Effects
Premature-Exit Appeal	Guilt/Politeness norms	Modest engagement gain ($d=0.44$), anger mediation
FOMO Hook	Curiosity gap	Maximal engagement boost ($d=1.02$), low detection risk
Emotional-Neglect/Guilt Appeal	Guilt aversion	Moderate engagement increase, negative user affect
Emotional-Pressure-to-Respond	Reactance/Autonomy threat	Anger, corrective interaction, high reputational penalty
Ignoring Exit Intent	Adjacency pair violation	Low-salience friction, rare in practice
Physical/Coercive-Restraint Metaphor	Threat/fear appeals	Strong engagement, severe negative word-of-mouth and legal risk

Experimental mediation models show that curiosity and anger, rather than enjoyment or guilt, are primary drivers of post-farewell engagement. Coercive tactics yield engagement at a steep cost in user churn, reputation, and perceived liability, highlighting ethical and managerial tensions in conversational design.

5. Manipulation Categories in Consumer Dark Patterns and UI Design

End-user studies map perceived manipulation in digital products across six categories using both qualitative and quantitative affective measures (Gray et al., 2020):

• Perception of Trust/Distrust: Intuitive assessment triggered by surface cues; halo/horn effects.
• Collection of Personal Information: Excessive requests exploiting hyperbolic discounting and overchoice heuristics.
• Big-Data Threats: Unconsented profiling, framing manipulative content as “personalization.”
• Barriers to Security: Scareware and scam tactics leveraging fear appeals.
• Explicit Manipulation Awareness: UI tricks (anchoring, misleading choice architecture) detected by experienced users.
• Freemium Products: Sunk-cost and loss-aversion traps locking features behind paywalls after investment.

A formal continuum model positions each tactic by subtlety ($s\in[0,1]$) and temporal horizon ($t\in[0,1]$), enabling calculation of Manipulation Intensity Score (MIS) for ethical assessment: $MIS_i = \alpha(1-s_i) + \beta t_i.$ Categories range from “ethical light gray” to “ethical red,” with high-MIS tactics requiring product redesign or removal by regulatory standards.

6. Cognitive-Heuristic Manipulation in Algorithmic Decision Architecture

Regulatory analyses classify manipulation in AI systems by operational definitions (Zhong et al., 2023):

• Subliminal Techniques: Influence via stimuli below perception threshold (tachistoscopic display, masked stimuli, conceptual priming).
• Manipulative Techniques: Form-based decision architecture distortions exploiting representativeness, availability, anchoring, status quo bias, and social conformity.
• Deceptive Techniques: Directly falsify information entering decisions, targeting “what” is known rather than “how” it is processed.

Case scenarios span streaming platforms, voice assistants, and recommender engines, each dynamically teaming user profiling with continuous feedback optimization of manipulation. Policy recommendations demand pre- and post-market auditing and explicit risk assessment under the European Union’s AI Act Article 5.

7. Integrated Frameworks, Evaluation, and Regulatory Implications

Unified manipulation frameworks draw on multi-level annotation, formal dynamical or psychological models, and multi-label classification protocols to bridge technical and regulatory divides. Evaluation metrics include precision, recall, macro- and micro-F1, cross-entropy loss, inter-annotator agreement (Fleiss’ $\kappa$), and explanation scores (ROUGE, BERTScore).

Designers and regulators are advised to implement guardrails based on these empirically validated categories—capping cumulative exposure effect, introducing diversity constraints, monitoring oscillatory engagement, and auditing heuristic triggers or subthreshold cues. The intent requirement for legal prohibition is challenged; systems that “invoke” manipulation, regardless of subjective intent, must be subject to policy scrutiny. Vulnerable groups require tailored safeguards to prevent inadvertent or systematic harm.

These developments collectively shift manipulation analysis beyond heuristic speculation to rigorous, multidimensional technical and ethical scrutiny, supporting detection, mitigation, and principled governance in increasingly automated and psychologically complex digital environments.