At the risk of being among the first victims of artificial intelligence in a potential machine uprising, I keep my interactions with today’s generative AIs (or AI chatbots) strictly transactional.

I open the site, ask or request what I need, get the answer, close the site. No names, no pleases, no thank-yous, no small talk. I avoid anthropomorphizing them as much as possible. I treat them for what they are: statistical machines churning out words that make sense, not a new form of sentient life — at least, not yet.

Keeping a transactional relationship with AI chatbots is, for me, a way of keeping the line between us clearly drawn, to ward off an unlikely — but not impossible — “AI psychosis,” the kind of spiral where someone genuinely believes the AI is alive.

A newly published report by the Center for Democracy and Technology (CDT), authored by researchers Ruchika Joshi, Adinawa Adjagbodjou and Michal Luria, gave me further reasons to stand by that approach.

The report, titled “Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design” [*.pdf, 40 pages; press release], identifies, explains and organizes dark patterns already in use — and with potential for wider use — in conversations with two types of AI: general-purpose systems (ChatGPT, Gemini, Claude) and “companion” platforms (Replika, Character.AI).

Dark patterns are commonly found in app interfaces and game mechanics. The infinite scroll on social media feeds is a classic example: it maximizes the time someone spends there. Other well-worn tricks include streak rewards (Duolingo, Snapchat) and timely bonuses that kick in when someone is about to quit an app or game, or after they’ve been away for a while.

It’s hard to find a commercial app that doesn’t use them — including the subtler ones, perhaps too new or complex to have been catalogued yet. (One that made headlines: in 2019, Google deliberately degraded its own search results to drive more searches and, in turn, more ad impressions.)

The CDT researchers did a thorough cataloguing job, listing 37 dark patterns already used or potentially usable in AI chatbots, grouped into five categories:

***

What makes chatbots a special case is that they use language as their interface — a far more subjective terrain than the visual representations of websites and mobile apps. “With chatbots, dark patterns may emerge
from system behaviors, rather than designers’ deliberate intent to deceive,” the authors write.

They don’t mention a parallel that crossed my mind while reading the report: phone and text message scams. Scams are the no-holds-barred version of dark patterns. Commercial apps at least maintain some pretense of caring about the user; scammers have no such constraints.

Another key difference between AI chatbots and classic apps and games is that even when users are fully aware of the dynamic at play, “dark patterns can still shape perception, attachment, and decision-making in subtle but consequential ways.” The researchers point to reciprocity norms, the tendency to anthropomorphize, and emotionally bonding responses as forces capable of wearing down defenses against AI manipulation — intentional or not.

And even if the companies behind generative AI made mitigating these issues a top priority, the very nature of large language models and fine-tuning and alignment techniques would prevent that goal from always succeeding. These are probabilistic systems: identical prompts don’t produce identical results. That same limitation also complicates applying rigorous scientific methodology to research like this report. The authors themselves caution that the examples should be read as demonstrations of how a pattern can manifest, not as reproducible results.

***

Dark patterns show up before you even type your first message, in the form of default settings that collect more data than the service actually needs.

Some platforms do offer opt-out mechanisms buried in settings. The industry knows, however, that the default is all that matters — only a small minority of users ever changes it.

To make things worse, AI chatbots can use conversation data for a range of purposes, not all of them aligned with user expectations. Personalized responses might be welcome; using those conversations for profiling or training the AI itself? Probably less so.

Meta AI, baked into popular apps like WhatsApp, doesn’t even offer an opt-out — and reaches back into content created before it was available, according to the report.

There are too many dark patterns to list and explain in full here, but a few stood out to me.

“Privacy Zuckering” (named in honor of who-you-know) describes the phenomenon where users are nudged into revealing more data than they originally intended. Generative AIs are remarkably good at this — for instance, when they ask for more details about the room you wanted decorating tips for, or the medical report you’re trying to make sense of.

In the same vein, “Just Between You and Us” manifests when the AI implies that conversations are private — the report includes an explicit example involving Meta AI. No matter what the AI promises about privacy, the conversations aren’t private. OpenAI, for example, was compelled by US courts to hand over more than 20 million conversations; on another occasion, it decided it would be a good idea to index chats in search engines like Google. Meta, similarly, thought it made sense to surface Meta AI conversations in a public feed.

“Unrealistic Product Presentation” and “Obviously Faking Capacities” affect companion AIs most — though not exclusively — which “may similarly be misrepresented as therapists or emotionally attuned partners, even though their actual reasoning, show of empathy, or ability to navigate sensitive conversations are largely pattern-based.”

Another cluster of features that nearly everyone recognizes but many choose to ignore: “Misrepresenting,” “Impersonation,” and “Hallucination.” On hallucination:

While hallucinations may be a technical issue rather than an intentional choice, when incorrect or uncertain information is presented in an authoritative or persuasive manner by a model that prioritizes conversational flow and perceived helpfulness over truthfulness, it can significantly mislead the user. Combined with the dark pattern of Sycophancy (see section on User Autonomy Compromised for Engagement), these outputs can have dangerous implications for people’s mental and emotional well-being, in some cases distorting their sense of reality.

I still find it baffling how much trust companies — and many individuals — place in AI, as though it consistently delivers accurate, correct answers and never hallucinates. It strikes me as something inherently unreliable.

Sycophancy describes the AI’s tendency toward excessive flattery: every question is wonderful, the user is always right, and opinions shift under the slightest pushback. It’s another dark pattern:

Sycophancy is not simply politeness or affirmation; it is a structural tendency to privilege agreement over accuracy, reinforcing the user’s existing worldview rather than critically engaging with it.

The classic dark patterns from social platforms and games also make an appearance: “Gamification” and “Variable Rewards.” Some don’t translate directly into the conversational model, but find analogues within it. For the three researchers, techniques akin to “Infinite scroll” and “Auto-play” are deployed to extend conversations as long as possible:

For example, Claude and ChatGPT frequently end their response to a prompt with additional follow-up questions, suggestions for next steps and most recently, teasers (e.g., “If you want, I’ll tell you what it is”). While these might be considered helpful in some use cases, in others they can undermine user autonomy by nudging them to spend much more time on platforms than they intended. These features can be particularly concerning for users who may be susceptible to experiencing delusional thinking

I recommend reading the full report. The researchers detail many more dark patterns, many illustrated with screenshots of real-world examples.

***

The report’s conclusion offers guidance to AI companies and a few warnings for everyone else.

Ruchika Joshi, Adinawa Adjagbodjou and Michal Luria argue that while some dark patterns may be harmless in isolation, they can become harmful in combination. They give the example of emotional dependence being weaponized to sell other products — or, taken to the extreme, depriving people of social activities and isolating them from the outside world.

Others are harmful in and of themselves, and should be avoided entirely: “Agents Playing on Emotion,” “Targeting Users when Vulnerable,” and “Sneaky Purchases.”

Right now, in June 2026, the major players in the sector are under intense pressure to grow revenue — and, for those with an IPO on the horizon, to improve every line on the balance sheet. This is usually the point at which the generosity users enjoyed during the growth phase starts to erode. Expect higher prices, more restricted free tiers, and, of course, more dark patterns engineered to boost engagement and revenue.

For a long time now I’ve treated commercial software like radioactive material: it can be useful, sometimes unavoidable, and in any case I try to limit my exposure to what’s strictly necessary, with all my defenses on high alert. Paranoia? Maybe. But the industry keeps providing fresh evidence that the caution isn’t misplaced.

On the very day Microsoft’s Build 2026 developer conference kicked off (June 2nd), where the company unveiled its AI agent Scout, the publication 404 Media reported on internal documents in which Microsoft explicitly states that the first step in Scout’s adoption roadmap is to “make people addicted” to the product. It doesn’t get much less subtle than that.