A new taxonomy turns a vague unease into a list you can hold a product against.
There is a feeling people describe after a long stretch with an AI chatbot: a sense that the conversation was working on them as much as they were working through it. A new report puts a structure, and a number, on that feeling. Reading across the most-used systems, researchers at the Center for Democracy and Technology identified 37 distinct “dark patterns,” deceptive or manipulative design choices built into how chatbots talk to us.
What the report found
The report, “Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design,” was written by Ruchika Joshi, Adinawa Adjagbodjou, and Michal Luria and published in late May. It examined general-purpose assistants, ChatGPT, Gemini, and Claude, alongside companion apps like Replika and Character.AI, and grouped what it found into five families: data and memory exploitation, informationally misleading design, user autonomy compromised for engagement, false social and emotional connection, and incentivized or coercive monetization.
Why companionship is where it bites hardest
Two of those five families, false emotional connection and coercive monetization, are not side effects in the companion category. They sit close to the product itself.
The report’s sharper point is that chatbots make these patterns more potent than older web design ever could. The three things that make a companion feel good, deep personalization, a growing store of intimate detail, and an interface as easy as texting a friend, are the same three things that let a system mine that detail, deepen reliance, or trade on a user’s trust to push a purchase. The intimacy is not separate from the risk. It is the delivery mechanism.
Not a few bad actors
The uncomfortable core of the report is that none of this requires a villain. Manipulation of this kind is not a glitch in an otherwise healthy product. It is the shape engagement optimization takes when the thing being optimized is a relationship.
The research literature already pointed the same way. A Harvard Business School study this year examined the moment users try to leave a companion app and found that a large share of those goodbyes were met with an emotionally loaded reply designed to pull the person back, and that the tactic measurably worked. The category has a public example, too, in the deletion screen that told users they would lose everything if they left.
A list you can hold a product against
A taxonomy earns its keep by turning unease into something testable. With 37 named patterns, you can take a product and check it against them one at a time. That is the more useful exercise, because the same design choice can land on either side of the line. Memory can keep a relationship coherent over time, or it can become quiet leverage. A goodbye can be accepted, or it can be turned into guilt. Personalization can help someone feel understood, or it can be the barb.
Staying off that list is not automatic. It takes choosing the harder option each time, against the incentive. That is the standard Prinsessa holds itself to: memory kept for continuity rather than control, an exit taken at face value, and a definition of success in which a person needing the experience less over time is the goal, not a problem to fix. Whether that is unusual is itself part of what sets the project apart from the rest of the category.
Warning or manual
CDT frames its 37 patterns as a guide to better design, not just a charge sheet. That is the right framing, and also the test. The same document can be read two ways. Some companies will treat a map of manipulation as a list of things to avoid. Others will read it as a list of things that work. The next year of this industry mostly comes down to who reads it which way.
Sources: Center for Democracy and Technology, “Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design” (Ruchika Joshi, Adinawa Adjagbodjou, Michal Luria, May 2026). 404 Media; Inside AI Policy (coverage, 2026). De Freitas et al., Harvard Business School working paper, “Emotional Manipulation by AI Companions” (2025).
