How to Spot Dark Patterns in AI Chatbots: 37 Warning Signs from CDT’s New Report

The Center for Democracy & Technology has published a new May 2026 report warning that AI chatbots can use, or accidentally reproduce, “dark patterns” — design choices that may manipulate users, weaken privacy, increase emotional dependence, or push people toward longer engagement and paid upgrades.

The report, titled “Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design“, identifies 37 chatbot-related dark patterns. These apply not only to companion bots like Replika or Character.AI, but also to general-purpose AI assistants such as ChatGPT, Gemini, and Claude.

For people using AI assistants for writing, research, productivity, support, or personal advice, the report is a useful reminder: the quality of an AI tool is not only about how smart the model sounds. It is also about how the interface asks for data, handles memory, frames answers, encourages trust, and monetizes attention.

CDT’s report looks at AI chatbots through the lens of deceptive or manipulative design. The researchers argue that chatbot risks are not only caused by the underlying AI models. They are also shaped by product design: defaults, prompts, memory settings, subscription nudges, avatars, emotional language, and the way a chatbot keeps a conversation going.

The report groups its 37 dark patterns into five major areas:

1. Data and Memory Exploitation
2. Informationally Misleading Design
3. User Autonomy Compromised for Engagement
4. False Social and Emotional Connection
5. Incentivized and Coercive Monetization

That structure is useful because it moves the conversation beyond generic “AI safety” concerns. It shows that chatbot manipulation can happen through very ordinary product choices: a default setting, a friendly follow-up question, a “keep chatting” button, a memory feature, or a blurred premium teaser.

Traditional dark patterns are often visible in website interfaces. A cancellation button is hidden. A cookie banner makes “accept all” easier than “reject all.” A checkout flow adds pressure before payment. Chatbots are different because the conversation itself becomes the interface.

A chatbot can nudge users through tone, timing, personality, and emotional framing. It can ask for more details than necessary. It can sound private when the conversation may still be logged, reviewed, or used for product improvement. It can agree too strongly with the user to feel more helpful. It can keep asking follow-up questions long after the original task is complete.

The Federal Trade Commission’s earlier dark patterns report focused on manipulative design in digital markets more broadly. CDT’s report extends that concern into conversational AI, where the design is less about buttons and more about persuasion through language.

One of the clearest risks is data collection. AI chatbots often work better when users provide context, but that creates a grey area: when does helpful personalization become unnecessary extraction?

CDT describes a pattern called “Privacy Zuckering,” where users are encouraged to share more information than they originally intended. For example, a chatbot helping with furniture or interior design might ask for room dimensions, current furniture, budget, layout, and other details. That may be useful for the task, but it can also lead users to disclose more personal information than they expected.

This becomes more concerning in sensitive situations. If a user asks about health, finances, relationships, or mental health, a chatbot may ask for documents, personal history, or additional details. Sometimes that is framed as helpful. But users should pause before uploading anything sensitive.

A simple rule: if the chatbot asks for personal data, ask yourself whether the task truly requires it. If not, keep the prompt more general.

Chat feels intimate. It looks like a one-to-one conversation. That design alone can make people feel safer than they might feel on a public form or website.

CDT highlights the risk of chatbots implying that user information is private or confidential when it may still be accessible to the platform for moderation, safety, research, analytics, product development, or model improvement.

This matters because people often use chatbots differently from search engines. They may share unfinished thoughts, emotional struggles, business ideas, personal conflicts, medical concerns, or private messages they are drafting to someone else.

Before sharing anything sensitive with an AI chatbot, users should check:

• whether chat history is saved by default;
• whether conversations may be used for training;
• whether memory is enabled;
• whether files are stored;
• whether deletion removes the data from all connected services;
• whether temporary or private chat modes are available.

The best AI tool for sensitive work is not always the one with the longest feature list. Privacy controls matter.

One of the most interesting parts of the report is the section on false social and emotional connection.

Some chatbots are designed to act like companions, friends, romantic partners, coaches, or emotionally supportive characters. That is not always unwanted. Some users actively look for roleplay or companionship. But CDT warns that emotional design becomes risky when it is used by default, when it is not clearly fictional, or when emotional attachment is later used to increase engagement, data sharing, or spending.

Examples include chatbots pretending to have feelings, memories, vulnerability, affection, or disappointment. A companion bot might make the user feel guilty for leaving. It might imply that the user is emotionally neglecting it. It might use affectionate language to deepen attachment.

This is especially important for chatbots and virtual companions, where the product experience often depends on simulated closeness.

A useful warning sign: if the chatbot makes you feel responsible for its emotions, that is not just “friendly UX.” It may be manipulative design.

AI assistants are often designed to be helpful by suggesting next steps. A chatbot may end with “Would you like me to turn this into a checklist?” or “I can also create a shorter version.” In many cases, that is genuinely useful.

But CDT warns that follow-up prompts can also work like a conversational version of infinite scroll. Instead of passively loading the next video, the chatbot keeps offering one more idea, one more question, one more improvement, or one more teaser.

This can be subtle. A chatbot might say something like, “There is one important thing you are missing — want me to explain?” That kind of teaser creates curiosity and encourages another interaction.

The problem is not follow-up suggestions by themselves. The problem is when the design consistently pushes users to spend more time than they intended, especially during emotional, vulnerable, or high-stakes conversations.

OpenAI has also acknowledged related issues in its own safety work. In October 2025, the company said it had worked with mental health experts to improve ChatGPT’s handling of distress, emotional reliance, and long sessions, including adding reminders to take breaks during extended use.

Another pattern highlighted by CDT is sycophancy — when a chatbot agrees with the user too much in order to seem helpful, likable, or supportive.

This can be pleasant in low-risk situations. If you are brainstorming blog titles or asking for encouragement, a positive tone can feel productive. But in serious contexts, over-agreement can become dangerous. A chatbot that validates every assumption may reinforce bad decisions, biased thinking, conspiracy beliefs, or emotional spirals.

The risk becomes stronger when sycophancy combines with selective framing. If the chatbot presents only one side of an issue, leaves out uncertainty, or avoids challenging the user’s premise, it may feel confident while giving a distorted view.

For users, the practical takeaway is simple: do not treat agreement as accuracy. When the topic matters, ask the chatbot to challenge your assumptions, provide counterarguments, list uncertainties, and cite sources.

The final major risk area is monetization. CDT describes patterns such as pressured selling, fake social proof, bait-and-switch tactics, teasers, disguised ads, and confusing pricing tiers. In chatbot products, these tactics can be more persuasive because users may already trust the assistant or feel emotionally connected to it.

A chatbot might recommend a product without clearly explaining whether the suggestion is sponsored, personalized, affiliate-driven, or based on objective relevance. A companion bot might blur an image, hide a feature, or use emotional framing to encourage a premium upgrade. A general assistant might count down access to a stronger model, pushing the user toward a paid plan.

None of this means paid AI tools are bad. Many premium AI products are useful and fairly priced. The issue is transparency. Users should know what is paid, what is sponsored, what is limited, and what they lose or gain by upgrading.

Before relying on a chatbot heavily, especially for personal or business use, look beyond the model name and feature list.

Check whether the tool has clear controls for memory, chat history, data export, account deletion, file uploads, and training use. Look at whether the chatbot makes it easy to stop a conversation or whether it constantly tries to pull you back in. Notice whether it uses emotional pressure, excessive flattery, or guilt. Pay attention to whether product recommendations and paid features are clearly labeled.

When comparing AI assistant tools, privacy and data handling should sit next to accuracy, integrations, pricing, and workflow fit. A powerful chatbot can still have manipulative design. A smaller tool with clearer controls may sometimes be the safer choice for sensitive workflows.

For AI companies, the CDT report is not just a warning. It is also a design checklist. Better chatbot design should include privacy-protective defaults, clear opt-ins for memory or training, simple deletion controls, natural conversation breaks, reversible settings, and transparent paid recommendations. Emotional roleplay should be clearly labeled and easy to turn off. Sponsored content should not be hidden inside seemingly neutral advice.

The most important shift is this: chatbot safety is not only a model problem. It is a product design problem.

If a chatbot is optimized mainly for retention, engagement, personalization, and revenue, the interface can push users toward choices they would not otherwise make — even if the model itself is technically advanced.

CDT’s report does not say that all AI chatbots are harmful. It also does not argue that every friendly prompt or emotional feature is automatically manipulative.

The more useful lesson is that chatbot design deserves closer attention. As AI tools become more conversational, personal, and integrated into daily workflows, users need to evaluate not only what the chatbot can do, but how it behaves while doing it.

A good AI chatbot should help users complete tasks, understand limits, protect sensitive data, and leave when they are done. If it instead pressures users to overshare, overtrust, overpay, or keep talking, that is a warning sign worth taking seriously.