A small warning, then a gut punch
We are heading towards a Minority Report moment. Not the bit where the cops arrest you for thinking about a crime. The bit where the interfaces show up when they’re needed, do the job, and disappear. Tom Cruise waving his arms at a screen that knows what he wants before he’s decided what he wants. That moment.
The premise is this: AI generates the interface on the fly. For the user. For the task. For the context. Based on the previous interactions. Possibly based on the user’s mood. The perfect interface for the task at hand, made up on the spot, used once, thrown away.
Where does AI learn what good interfaces look like
So where, exactly, is the next generation of interface-generating AI supposed to learn what good interfaces feel like? From scraped, decade-old Bootstrap tutorials from 2015? From SEO articles about UX written by AI for AI? From slop-shots optimised for likes rather than use?
It’s not the internet anymore. That is for sure. The cleanest, most useful, most opinionated source of human-generated interface design data left is us: the app creators, the web developers, the software engineers, and this legion of new “app creators” designing their perfect interface or tool to help them do their job better.
The word for it is vibecoding. You ask Claude, Cursor or Lovable to build you a little dashboard. A tracker. A calculator. A custom CRM view. A small bespoke tool for the one specific problem you have. You iterate on it. You tell it what feels right, what doesn’t, what to move, what to remove. You make a thing that does exactly what you need.
And every single one of those interactions with the AI, the “no, the button a bit larger” the “move the counter more to the top” the “make it a sortable table not a list”, is the highest-quality and most direct signal for the AI of what humans actually want their interfaces to do and to look like.
Real users. Real tasks. Real preferences. Real correction loops. No focus groups. No usability studies. No designer interpreting business requirements through three layers of management. Raw statistical data.
That is gold. That is exactly the data the next generation of AI needs to learn how to generate interfaces on the fly. And we are producing it, voluntarily, at scale, for free.
Here’s the gut punch
If you are a business owner, and you have been quietly delighted to find that you can use AI to build your own small tools instead of paying an agency or a freelance developer, congratulations. I genuinely think you’ve discovered something amazing. We, old school developers, we do this already for a few years now.
But you are, accidentally, training the AI to do what you do in your business.
Think it through. Right now you build, let’s say, custom dashboards for accountants. You charge a fair price. People hire you as an accountant because you know the domain and experience, and now you have tools to supercharge your services.
You discover Claude. You start using it to scaffold your work. Faster, easier. You ship more. You ship better. Great. More service to sell, less time to spend.
In the background AI is always learning and getting better, learing from your input and from all accountants all over the world. Your potential client won’t ask you when they have a question. Of course they’ll ask the AI for help first. And the AI will gladly build a fully personalised dashboard based on your training data and that of everyone else in your field. You trained it. You used it for a while. And now you’re not in the picture anymore.
The AI does not need to recommend you. It does not need to find you in a directory. It does not need to call you for a kickoff meeting. It already knows what the best dashboard for your client’s specific case should look like. It learned that, in part, from you.
You thought you were saving money. You were training your own replacement.
Who’s next
This isn’t just web developers, although we developers, as early adopters and power users of AI, are on the cusp of being phased out. It’s any service business that solves specialised, repeated problems. Copywriters. Bookkeepers. Translators. Marketing consultants. Tax preparers. A particular sort of small-scale technical consultant. SEO specialists. PR generalists.
we developers, as early adopters and power users of AI, are on the cusp of being phased out
The pattern is the same: if your job involves taking a domain you know well and shaping it into a deliverable a non-expert can use, AI is being trained on every interaction you have with it. Each one makes the model a bit better at what you do. And eventually (not tomorrow, not next year, but eventually), the customer skips you and asks the model directly.
The cruel twist is that the more competent and specific your AI prompts are, the more valuable your contribution to that future model. The best practitioners in any field are, structurally, the ones giving the most signal about how to replace themselves.
Should we stop using AI then?
No. Honestly, no. I’m not going to stop either. The productivity boost is too real, the work is too much fun, and the alternative is being the only person in the industry who doesn’t use these tools. That ends badly too.
Choose what you publish. Choose what you put into the prompt window. Save your secret sauce for offline. Diversify your skills and your fields. That’s a key part of staying useful. AI doesn’t understand the world the way it is; it sees patterns. You, as a human, can make the connections between different fields, the people you work for, the tasks at hand. Keep the bigger picture. Don’t let the AI do the thinking for you. Stay the boss.
That’s the part to lean into. That’s the part the model can’t replicate. Yet.
English 
Dutch