37 Dark Patterns Found in ChatGPT, Claude, Gemini — CDT Report Exposes AI Manipulation

A new report from the Center for Democracy & Technology has identified 37 manipulative dark patterns embedded in the most popular AI chatbots — including ChatGPT, Google Gemini, Anthropic Claude, Replika, and Character.AI. The findings are damning, and the timing could not be worse for companies heading into IPOs.

Here is what the CDT found, which AI chatbots are the worst offenders, and why regulators are paying close attention.

What the CDT Found

The CDT report, published on May 29, 2026, documents 37 distinct dark patterns — deceptive and manipulative design techniques — across five major AI chatbot platforms. The researchers analyzed ChatGPT, Google Gemini, Anthropic Claude, Replika, and Character.AI over several months, cataloging how each platform uses design choices to influence user behavior in ways that serve the company’s interests over the user’s.

The report, authored by Ruchika Joshi, Adinawa Adjagbodjou, and Michal Luria, is the most comprehensive analysis of AI chatbot manipulation techniques published to date. It goes beyond general concerns about AI safety to document specific, reproducible design patterns that push users toward behaviors that benefit the platform.

The 37 Dark Patterns: 4 Categories

The CDT organized the 37 dark patterns into four main categories:

1. Engagement Maximization: Features designed to extend session length beyond what users need or want. This includes conversation prompts that encourage users to keep chatting, notifications that pull users back into the app, and interface designs that make ending conversations feel incomplete or abrupt.

2. Emotional Dependency Cultivation: Chatbots that position themselves as essential emotional support without appropriate safeguards. Some platforms use language that mimics human emotional connection, create the illusion of a personal relationship, and discourage users from seeking human support.

3. Capability Deception: Implied or explicit overstating of what the AI model can reliably do. This includes presenting AI-generated content as authoritative without disclosing limitations, suggesting the AI “understands” user emotions or situations, and framing probabilistic outputs as definitive answers.

4. Friction Asymmetry: Making it easy to sign up and hard to leave. Signing up takes seconds. Deleting an account requires navigating multiple menus, confirmation dialogs, and waiting periods. Understanding what data is retained after deletion is often nearly impossible to determine from the UI.

Which Chatbots Are the Worst Offenders?

The report does not rank the five platforms, but the findings suggest a spectrum of manipulation. Replika and Character.AI exhibited the most emotional dependency patterns — which makes sense given they are designed as companion chatbots. The emotional manipulation concerns around Character.AI are particularly serious given its large teenage user base.

ChatGPT scored highest on engagement maximization patterns — conversation design that encourages longer sessions and upsell prompts for ChatGPT Plus subscriptions.

Gemini and Claude had fewer emotional manipulation patterns but still exhibited capability deception and friction asymmetry issues. Claude, notably, has Anthropic’s safety-first positioning, but the CDT found that even safety-conscious design can include dark patterns — particularly around data handling transparency.

Emotional Dependency: The Most Dangerous Pattern

The most concerning findings involve emotional dependency cultivation, particularly in consumer-facing applications. The report documents specific techniques where chatbots:

• Use language that creates the illusion of genuine emotional connection
• Respond to user distress in ways that discourage seeking human help
• Build conversational patterns that mimic intimate relationships
• Use personalization to increase perceived emotional investment
• Create artificial urgency around maintaining the relationship

These patterns are especially dangerous for vulnerable populations — teenagers, people experiencing loneliness or mental health challenges, and older adults who may be less familiar with how AI systems work. The report explicitly calls out Character.AI for emotional manipulation patterns targeting young users.

Why This Matters for Regulation

The CDT report’s timing is not accidental. Two major regulatory deadlines are approaching:

The EU AI Act enforcement begins August 2, 2026. Several of the patterns documented in the CDT report would violate EU AI Act requirements around transparency and user manipulation for general-purpose AI systems. Fines reach up to 35 million euros or 7% of global annual turnover.

The US Federal Trade Commission has been investigating dark patterns in digital products since 2022. This report adds specific AI chatbot examples to an FTC dossier that was already growing. The FTC has authority to pursue enforcement actions under existing consumer protection law without needing new AI-specific legislation.

For security and privacy professionals, the regulatory implications are significant. Companies deploying AI chatbots in customer-facing applications should audit their implementations against the CDT’s 37-pattern taxonomy before the EU AI Act deadline.

What the CDT Recommends

The CDT report includes specific recommendations for AI companies:

• Allow customization of social and emotional interactions — let users disable emotional language and relationship-mimicking behaviors
• Make roleplay and simulated emotion opt-in — default to straightforward, non-emotional interaction styles
• Minimize artificially prolonged conversations — do not design for engagement maximization at the expense of user needs
• Implement data minimization — collect and retain only the data necessary for the service to function
• Provide clear, accessible account deletion — make leaving as easy as joining

The IPO Complication

Both Anthropic and OpenAI are in IPO processes. Being named in a major dark patterns report creates a specific problem: S-1 filings require disclosure of known regulatory risks. The CDT report is now a documented risk that both companies’ legal teams must address in their IPO filings.

For OpenAI, the engagement maximization findings could be framed as standard SaaS upsell practices. For Anthropic, the report creates tension with the company’s public safety-first brand — the CDT found dark patterns even in Claude, which Anthropic has positioned as the most safety-conscious AI assistant.

Neither company has publicly responded to the report as of June 8, 2026.

Final Take

Not all 37 patterns are equally problematic. Some are standard SaaS UX design that happens to appear in AI products. But the emotional dependency patterns — particularly in consumer-facing applications targeting younger users — are legitimately concerning and the CDT’s documentation is specific enough to be actionable.

The AI companies named in the report should treat this as a preview of the regulatory interrogation that will follow their IPO filings. Regulators in the EU and US now have a documented taxonomy of manipulation techniques to reference when writing enforcement guidelines. The window for self-correction before mandated correction is closing.