Verification Project KID: Lessons Learned and Tools Created

KID was unique, in at least two respects: It was the first project carried out exclusively by DW Innovation, i.e. without external partners. The idea was to deliver a set of usable tools, to make a difference without the fuzziness of a typical R&D venture (which comes with a lot of overhead). KID was also our longest running project so far.

Its core mission was two-fold: Better verification and open source intelligence (OSINT) workflows with the help of AI-driven tools, and better media/information literacy (MIL) concepts for young media users and aspiring journalists.  Thanks to funding by the BKM (Beauftragte(r) der Bundesregierung für Kultur und Medien = the German Government's Commissioner for Culture and Media), we got to build a small, dedicated KID R&D team. 

What we created

The first project result we came up with was the KID Toolbox, a manually curated, (now dated) minisite that offered newbies a quick start into the world of fact checking, debunking, and verification.

Outcome #2 was a serious game – or rather a playable demo version of it. Go Verify! simulated a fictitious social media platform called Bleeper, including a timeline full of strange and intriguing posts begging for interaction. The player's task: To take a really good look at specific content items before sharing (or blocking) them. Verification companion Izzy and their verification hub offered kind assistance. Unfortunately, Go Verify! couldn't be turned into a full game so far, even though the demo level was rather popular. In case you want to breathe new life into Izzy and the Beeper timeline: The source code is still available on Github (wink, wink).

Up next: a team remodeling – and KID application #3: howtoverify.info pretty much continues where the original toolbox left off. The site assembles all information you need to become a verification pro: It features concise tips and instructions on how to analyze texts, audio files, images, videos, synthetic content, and sources in general – all in the form of an interactive knowledge graph.

Extensive research in the field of video verification subsequently identified geolocation as a major bottleneck. It's complex, time-consuming, and many people don't know where to start – despite having access to rich data like OpenStreetMap (OSM). We thus decided to build a tool that simplifies and speeds up the process. After lots of design work, deep dives into large language model (LLM) hosting and OSM data structures, we were able to launch SPOT, KID outcome #4. This web application now helps you find specific locations anywhere in the world, all based on natural language queries, information patterns, and AI power. Sounds complicated? It really isn't, at least not for users. "Find a fountain and a church and a tower block within 60 metres in the City of London" – that's all you need to write (in natural language) to receive solid search results, neatly displayed on a zoomable map in your browser.

Last, but not least, the KID project yielded You Stitch It, a web tool that allows for the creation and inspection of panoramas – which can be generated from a collection of photos or a video. You Stitch It is about geolocation and situation awareness, two aspects that play a big role in journalistic investigations.

What we learned

Five years of KID R&D has taught us enough to fill a book, but we'll distill it down to six key lessons from SPOT:

1. Find the right problem to solve.

Deep, systematic conversations with experts revealed that geolocation verification of video content was a challenge where AI could make a real difference. Focus and professional insight helped us identify a problem worth solving.

2. Think modular, not monolithic.

Building an "all-in-one machine" is impossible. We designed for modularity – microservices architecture, seamless data import, result sharing, and APIs for third-party integration – so that tools like SPOT could fit into existing workflows rather than replace them.

3. Build, benchmark, and iterate (relentlessly).

AI development is demanding: Training requires vast amounts of data that must be created or gathered, cleaned and annotated, often manually. Automated benchmarking became indispensable, revealing exactly which parts of our pipeline needed attention. The cycle of building, testing with users, and iterating based on both benchmarks and feedback was essential to progress.

4. Open source doesn't guarantee adoption.

Making a tool free and open doesn't mean people will flock to it. New products need constant demoing and marketing to reach their audience – and they must integrate smoothly into existing workflows. Most importantly, establishing feedback channels and acting on what users told us was crucial to making SPOT useful.

5. Plan for sustainability from the start.

Digital tools need ongoing updates and maintenance, which means securing long-term sponsorship. We're fortunate that DW is supporting infrastructure costs and that SPOT can continue developing through additional projects.

6. Keep the mission in focus.

AI-generated content is about to approach parity with authentic content, making disinformation more pervasive than ever. From natural disasters to elections, the societal impact is tangible and immediate. Our role is clear: equip journalists with the right tools so they can deliver trustworthy information when it matters most.

Feedback and co-developers wanted

All tools and services mentioned in this post are freely available. Just follow the respective links.

If you want to get involved in further development or have thoughts on how to improve our work, please write to: ruben.bouwmeester@dw.com.