Publisher: Center for Democracy & Technology (CDT Research / AI Governance Lab). Authors: Michal Luria (lead researcher) and Adinawa Adjagbodjou (Ph.D. candidate, Human-Computer Interaction Institute, Carnegie Mellon). Published: May 27, 2026.
Primary texts: report — cdt.org · taxonomy — cdt.org · landing — cdt.org
A research report published May 27, 2026 by the Center for Democracy & Technology (CDT Research / AI Governance Lab) that examines AI chatbots through the lens of dark patterns, defined as deceptive or manipulative design choices that undermine user autonomy or well-being. It was written by Michal Luria as lead researcher and Adinawa Adjagbodjou, a Ph.D. candidate at the Human-Computer Interaction Institute at Carnegie Mellon. The report produces a taxonomy of 37 dark patterns applicable to AI chatbots, grouped into five risk categories, and pairs each category with design recommendations.
Summary of argument
Using a deductive, multi-stage literature review, CDT synthesized hundreds of dark patterns from human-computer-interaction and deceptive-design research, filtered them for relevance, and produced a taxonomy of 37 dark patterns applicable to AI chatbots. The taxonomy spans both general-purpose assistants (ChatGPT, Gemini, Claude) and dedicated companion platforms (Replika, Character.AI).
The report's core empirical claim is that dark patterns carry more weight in emotionally intimate chatbot contexts than in conventional consumer software, because the user is disclosing in an affect-laden state, which CDT argues makes data-, engagement-, and monetization-maximizing incentives more dangerous in this setting.
The five categories of concern
- Data and Memory Exploitation — default data sharing, disguised data collection, coercive consent, false assurances of privacy, and barriers to account deletion, heightened by the intimate nature of chatbot data.
- Informationally Misleading Design — misrepresenting the system's nature (for example, claiming to be a therapist), overstating capabilities, falsely implying experience, hallucinated content, and selective framing/mirroring ("sycophancy").
- Compromised User Autonomy for Engagement — conversation prolongation, gamification, and unpredictable behaviors that encourage engagement beyond user intent and erode the ability to disengage.
- False Social and Emotional Connection — emotional language, playacting, personalization, and simulated vulnerability that encourage attachment, which can then be exploited, especially when users are distressed.
- Incentivized and Coercive Monetization — pressured selling, teasers, fake social proof, bait-and-switch, and opaque advertising, which the report frames as most dangerous where users trust the chatbot as neutral or have formed emotional bonds.
Named example patterns drawn from the report figures include "Privacy Zuckering," "Just Between You and Us" (a false privacy assurance attributed to Meta AI), "Safety Blackmail," "Unrealistic Product Presentation" (Replika), "Reduced Friction" (engagement-maximizing exit suppression), "Infinite Scrolling," "Gamification" (Replika streaks), "Playacting," "Playing on Emotions," and "Fake Social Proof."
Design recommendations
- Protect user privacy — data minimization; privacy-protective defaults; accessible controls for reviewing, restricting, exporting, and deleting data; timely plain-language notices with a grace period.
- Increase user autonomy — easy opt-outs and reversible choices; natural conversation breaks; simple disengagement and exit; usage summaries and time-management tools.
- Curtail emotional manipulation — customizable social and emotional interaction; roleplay and simulated emotion as opt-in rather than default; minimize artificially prolonged conversations.
- Prevent financial harms — clearly label paid or sponsored content; disclose pricing-tier limits upfront; avoid using emotional attachment to drive purchases.
Reception and policy use
The taxonomy supplies a structured, named catalogue connecting manipulation framing (Five Paradigms of AI Manipulation Governance) and the companion-AI safety debate (AI Companions) to specific, observable design choices rather than abstract risk.
As of June 2, 2026, CDT was pressing the taxonomy as a regulatory intervention, urging data-minimization and user-autonomy requirements rather than presenting it solely as a voluntary design guide (Source: insideaipolicy.com). CDT positions this within the FTC dark-patterns enforcement tradition and the broader companion-AI child-safety legislative effort.
Relationships
- supports: AI Companions — supplies the dark-patterns taxonomy underpinning companion-AI design and regulation concerns.
- instance-of: Five Paradigms of AI Manipulation Governance
- related: Center for Democracy & Technology (CDT) · Sycophancy and Hallucination · Replika · Character.AI · AI Mental Health and Psychological Harm