My participation in the UNESCO Youth Hackathon

This summer I took part in the UNESCO Youth Hackathon 2026. The theme was "Play Your Part: Youth Designing the Future of Media and Information Literacy", and one of the tracks asked for innovative solutions to the challenges AI poses, through media and information literacy. I read that line and knew immediately what I wanted to build.

I didn't win. Out of 2,028 projects from 129 countries, mine wasn't among the four winners, and I have a certificate of participation to show for it. But this hackathon taught me more than most of the projects I've finished, so I want to tell it honestly.

Why this theme mattered to me

More and more people get their answers from an AI now. They don't open ten links any more. They read one confident paragraph with a few citation chips and move on. I use these tools every day, and I know how convincing they sound, including when they're wrong.

That is a sensitive subject, and it's easy to get wrong in both directions. If you say "AI lies", you sound like someone who hasn't used it. If you say "AI is reliable", you hand your judgement to a machine. And whoever builds a tool to sit between the two quickly faces the hardest question of all: who decides what is true?

I wanted to help people trace what they read, not to tell them what to believe.

Source-Trace

So I built Source-Trace. It's a browser extension that reads an AI answer on Perplexity or ChatGPT and marks every claim in place: green when a source backs it, orange when the source is weak or unverifiable, red when nothing visible supports it. A "deep trace" goes looking for independent sources on the web and sums up what each one says. It works in English, French and Spanish, and a privacy mode keeps everything on the device.

I was confident. Very confident. The first commit went in on 1 July, and in five weeks I had an extension, a FastAPI backend, a shared contract between the two, tests, a privacy policy and a version ready for the Chrome Web Store. I would open Perplexity, ask something, watch the claims light up, and think: this is useful, people need this.

The lie detector I wanted to sell

In my head, the pitch was bigger than an extension. I wanted to position Source-Trace as a lie detector for AI. Then grow it into a mobile SDK, so that any app with a chatbot inside, a bank, a health service, a school platform, could show its users which parts of an answer stand on something solid.

There was a tension I didn't want to look at. In my own design document, the first rule was "coach, not oracle": the tool never says true or false, it only describes the sourcing you can see. So I was writing code that refused to judge, and dreaming of a pitch that promised to catch lies. I told myself it would sort itself out with time.

Check this out

Here's the pitch I actually sent to UNESCO: Source-Trace in under three minutes, filmed vertically like a proper influencer. Spoiler: no trophy. But I still think AI answers deserve someone checking their homework.

How the models knocked me out

What I underestimated was speed.

While I was building a tool to trace AI answers, the models were learning to do it themselves. They search the web before answering. They cite their sources inline. Each new generation is better at saying what it doesn't know instead of inventing it. The gap I was trying to fill was closing while I filled it.

The clearest sign was in my own code. My deep trace worked by asking a language model to search the web for independent sources. The very capability that made my tool necessary was also what made it work, and it was getting better faster than I could build. In early August I cut the deep trace from 180 seconds down to 31, and fixed a bug where the model's own citation tags leaked into my results. The model was already citing its sources. I was rebuilding, one layer up, something it was learning to do natively.

That's the knockout. Not a competitor, not a bug, not a lack of time. The ground moved under the idea.

What I'm keeping

I'm not bitter about it. I'm a little proud, actually, of how far five weeks went.

And some of it still holds. A model that cites its sources doesn't mean you've read them. The habit of opening the link, finding a second source and noticing when a claim stands on nothing is media literacy, and no model update will build that habit for us. If anything, the more fluent the answers get, the more that habit matters.

The real lesson is about building on AI. If your product fills a gap in what the models can do today, ask yourself where they'll be in six months. Build on what stays true when they improve, not on what they can't do yet.

Source-Trace is open source, under the MIT licence, on GitHub. Have a look, take it apart, reuse what you want.