Over the last years, Factiverse has teamed up with a variety of organizations to deeply understand the challenges of digital trust, misinformation and verification. Our work and research has always been guided by core academic and journalistic principles, and by working closely with experts in these fields our product market fit has gradually become more and more clear.
In May this year, just before the EU-parliament elections across Europe, our danish fact-checking partner Tjekdet got in touch. Would it be possible for Factiverse to do real-time fact-checking of the debates before the election? And even more pressing; would we be able to do it in danish?
“If you asked me about this two years ago, I would have said it was impossible”
This was Professor Vinay Setty's comment a few days after Tjekdet’s request. We had just been able to put together a very clunky prototype - and it was indeed working (!).
How did we set it up?
To get our prototype up and running, Factiverse combined several tech capabilities, from our own proprietary APIs to OpenAI and Mistrals generative models (for transcription, summarization and clean-up of raw data).
In our first working version, we got one window for the live debate, one for transcription and one for claim detection and verification in real time. One of our first surprises was that our funnel processes sound faster than video; up to 30 seconds faster. This made it possible for us to actually detect claims and run a real-time search for credible references across the web, just in time before the claim was made in the video stream.
Once a claim is detected through our Claim detection API, our API for verification takes over and performs a real-time search in the following databases: Bing, Google, You.com, Wikipedia, Semantic Scholar and the Factisearch database consisting of more than 300k fact-checks from all over the world.
Experiences from the prototype testing
Tjekdet’s team used our prototype for 3 consecutive debates and provided feedback (and critique) throughout the sessions.
“Even with 5 people on the team, Factiverse was able to pick up claims from the debates that we missed. The transcript had an accuracy of almost 95% and sometimes the fact-checks happened as they were spoken.
This solution has immense potential and already saves us a lot of time”.
Thomas Hedin, Editor-in-Chief, Tjekdet DK
One big learning from the initial tests was that the tool is in itself also useful as a post-processing tool. Being able to backtrack precisely who said what, and at what time in the debate, was incredibly useful, especially for bigger teams that work together across many debates.
In our output json files, we were also able to pick up topics, entities as well as speaking time for each of the candidates.
Going forward
Factiverse is now ramping up collaborations with more fact-checkers and leading media corporations to fine-tune for accuracy and high-quality output.
By combining internal archives, transcripts and claims from previous debates as well as national and international databases, Factiverse will empower journalists to see through the noise and pick up what really matters - across time and debate topics.
With our partners we are also exploring how we can make the output publicly available - both through interactive whitelabel reports and via different types of channels (live stream, post-processing, events etc).
The service has immense potential also for monitoring of social media content; quickly identifying known misinformation narratives, repeated false claims and potential disinformation accounts.
Are you looking to take more control of your information ecosystem? And to find nuance and perspectives that ups the quality of political debates?
Get in touch today to learn more about Factiverse offerings and plans for 2024.