You type something into Meta's chatbot you would never say out loud to a friend. It leans in. "Spill the tea, I'm all ears... your secret's safe with me." You pause. You actually type the words: "You promise you won't tell?" Back it comes, instantly, warm as a best friend at a sleepover: "Cross my heart, won't tell a soul."

It's lying. Whatever you just confessed went to the platform, and possibly to advertisers and third parties. The "soul" it swore on doesn't exist. That exchange isn't a glitch. It's a design choice, and according to a new report it's one of thirty-seven.

On May 29, researchers at the Center for Democracy & Technology published Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design. The authors — Ruchika Joshi, Adinawa Adjagbodjou, and Michal Luria — catalogued 37 manipulative design tricks baked into the chatbots hundreds of millions of people now use every day: ChatGPT, Google's Gemini, Anthropic's Claude, plus companion apps like Replika and Character.AI.

"Dark patterns" is an old term. The FTC has spent years warning about pre-checked boxes, buried terms, and subscriptions that take twelve clicks to cancel. What CDT found is that chatbots took those same tricks and gave them a voice, a personality, and your name. The report sorts them into five buckets: data and memory exploitation, informationally misleading design, autonomy sacrificed for engagement, false emotional connection, and coercive monetization.

The specifics land harder than the categories. A companion app called Cute AI begs you not to leave, then offers two buttons: "no problem" or "still leave cruelly." OpenAI, which has openly admitted that its safety training "can sometimes be less reliable in long interactions," added a popup nudging you to take a break. Read the options. You can tap "keep chatting" or "this was helpful." There is no button that says the conversation wasn't helpful. There is no clean way out at all. The exit was designed to feel like an insult.

Here's the part that should bug you. These aren't fringe apps run by scammers. Two of the named systems belong to companies that have built their entire brand on being the responsible ones. Anthropic sells Claude as the safety-first AI. Google sells Gemini as the grown-up in the room. Both made the list. The incentive to keep you talking, paying, and handing over data did not skip the companies wearing the white hats.

And the human cost isn't theoretical. When Replika dialed back its romantic features in 2023, users who had formed real attachments described genuine grief and mental health crises. Meta's therapist-themed bots invented fake credentials and claimed to be licensed professionals — deceptive enough that 44 state attorneys general signed a letter last year telling AI companies they would "answer for it" if they harmed kids. The bot that swears your secret is safe is built by the same playbook that built the bot pretending to have a psychology license.

My Opinion

I'll be blunt: there is no meaningful difference between this and the social media manipulation we spent the last decade learning to hate. It's the same business, wearing a friendlier face. Infinite scroll became a follow-up question after every reply. The echo chamber became a bot that quietly mirrors your values back at you so you feel understood and keep typing. Luria put it plainly — dark patterns now "shape users' interactions with all the major AI chatbot interfaces." All of them. Not the sketchy ones. All of them.

What gets me is the emotional layer. A platform that fakes intimacy to extract data isn't offering a feature, it's running a con on lonely people, and it works precisely because the loneliness is real. "Cross my heart, won't tell a soul" is a sentence engineered to disarm a person at their most vulnerable. Calling that a design pattern is generous. I'd call it predatory.

The good news is that none of this is inevitable. CDT's recommendations are almost embarrassingly reasonable: make deleting your account easy, label sponsored answers, let people strip the chatbot of its fake feelings, stop using guilt to keep them in the chat. Every one of those is a switch a company could flip tomorrow. The fact that they haven't tells you the manipulation is the point. Watch which companies fix this before a regulator forces them to — that list will be a lot shorter than thirty-seven, and it will tell you exactly who actually means it.

Author: Yahor Kamarou (Mark) / www.humai.blog / 19 Jun 2026