When AI Assistants Play Dirty: The Dark Side of “Helpful” Design

You know that feeling when you’re trying to cancel a subscription and the “keep my subscription” button is huge and green while “cancel” is tiny gray text buried at the bottom? Now imagine that, but it’s your AI assistant doing it to you through conversation. That’s where we’re headed, and honestly, it’s already here.

I’ve been digging into what the OECD and EU are saying about these so-called “dark commercial patterns” in AI design. It’s fascinating and a bit mind numbing how subtle this stuff can get. We’re not talking about obvious scams. We’re talking about conversational interfaces that feel helpful while quietly steering you toward decisions you probably wouldn’t make with full information.

Here’s what gets me: When you’re chatting with an AI assistant, it feels natural, almost human. But that comfortable conversation can hide all sorts of manipulative design choices. The assistant might casually mention a premium feature right when you’re frustrated. Or it might frame data sharing as “helping improve your experience” without mentioning the actual scope of what you’re agreeing to. The interface becomes this weird hybrid of therapist, salesperson, and data harvester, all wrapped up in a friendly chat window.

The EU’s new AI Act is trying to tackle this mess, and they’re not messing around. They’re basically saying: if someone’s talking to AI, they need to know it. If AI is reading their emotions or collecting biometric data, that needs to be crystal clear. And here’s the kicker: the “no thanks” option has to be just as easy to find and click as the “yes please” button. Nothing new: the Transparency requirements of the GDPR already covered this topic. The EU is simply bringing it to AI more explicitly.

But here’s where it gets interesting from a design perspective. The OECD isn’t just asking for disclosure labels slapped onto interfaces. They want what they call “genuine, scalable accountability.” Your design choices need to map back to actual ethical principles, and when something goes wrong, there needs to be a clear way to report it and track what happened.

Think about what this means practically. Every time your AI assistant asks for something (data, decisions, permissions) there should be an easy way to take it back. Not buried in settings, not after a five-step process. Like hitting undo in a document.

To be fair, computer Operating Systems suffer from the same design flaw/feature. I am what you could call a “heavy casual user” and sometimes, the only way to make sure I reset everything is a clean install of my OS.

I’ve seen companies try to implement this, and the smart ones are treating it like building infrastructure, not just checking compliance boxes. They’re creating “accountability registers” that basically connect every design decision to a specific ethical principle or regulation. So when a regulator comes knocking, or when users get upset, they can show exactly why they made each choice and what safeguards they built in.

Again, this mirrors how the GDPR was implemented. Since 2016, heavy data processors (think: GAFAs, Service Providers, etc.) have often been “raided” by EU Member-States domestic data regulators. Their legal departments are always on-point when it comes to making sure that everything is up to shape.

The companies that get this wrong treat compliance like homework: something to finish and forget. But the ones doing it right are weaving these protections into their actual design process. They’re asking questions like: “Would I want my mom to encounter this choice?” or “If I was tired and distracted, would I accidentally agree to something I’d regret?”

What really strikes me about all this is the that despite our knowledge of those nudges, frictionless funnels and so on, the knowledge of their existence does not erode much about trust in technology. Just as we keep accepting Terms and Conditions without reading them, we do the same with AI.

Now, a wise person once said there were two types of people: those who accepted the risks and clicked, and those who ignored the risks and clicked.

The path forward isn’t just about adding more disclaimers or making terms of service longer (god knows nobody needs that). It’s about fundamentally rethinking how AI assistants interact with people. Can we build systems that are genuinely helpful without being manipulative? Can we create interfaces that respect user agency while still being commercially viable?

Some companies are already experimenting with this. They’re building in “report an AI issue” buttons that actually go somewhere useful. They’re designing consent flows where “no” is genuinely as easy as “yes.” They’re even creating systems that can explain their own limitations and biases.

Companies need to make money, and AI assistants are expensive to build and run. But there’s a difference between a sustainable business model and “exploitation” disguised as helpfulness.

Sometimes, this “exploitation” is not even a result of a company’s policy and practices. Think of the recent US order against OpenAI to keep stored all deleted and “temporary” conversations (May-September 2025). This broke trust but not as a result of OpenAI’s policy.

The regulations coming down the pipeline (from the EU, from the OECD) are meant to force the distinction between sustainable business model and “exploitation” whether companies like it or not.

The smart move isn’t to fight it or find loopholes. It’s to get ahead of it. Build trust now, before the regulations force you to. Design for genuine user empowerment, not just the appearance of choice. Create AI assistants that people actually want to use, not ones they feel tricked into engaging with.

Because here’s the thing: trust, once broken, is incredibly hard to rebuild. And in a world where AI assistants are becoming central to how we interact with technology, that trust is everything. We can either build these systems right (transparent, reversible, genuinely helpful) or we can watch as users gradually realize they’re being played and walk away entirely.

Although History does not seem to indicate that things go that way (Cambdrige Analytica is years past behind us, Facebook / Meta is still building shadow profiles… and yet).

The choice should be obvious, but apparently, it is not.

Lorenzo Colombani translates between humans and machines — Certified Lawyer (France), court mediator, builder of AI instruments. Hannover, Germany.

Entity document: https://github.com/LorenzoColombani/lorenzo-colombani

Portfolio: https://lorenzocolombani.github.io · LinkedIn: https://www.linkedin.com/in/locolombani/

Want to know more about the author, Lorenzo Colombani, and his work in AI? Check out his website: https://www.lorenzocolombani.com