Categories
Reading time
One of the challenges facing the AI team at Full Fact is the pace of change: over the last few years, there have been new models, libraries and concepts published every day and any of these could be useful in our work. We're a small team, so it's a particular struggle to keep up to date. Thankfully, Ian Ozsvald stepped in to help. Ian is an independent Chief Data Scientist and one of the many hats he wears as the organiser of "playgroups". For these, he brings together small but diverse groups of experts across the AI / data science community to share ideas and do a deep dive into matters of interest.
Recently, Full Fact was happy to host one of Ian's playgroups as a hackathon. We presented the group with an overview of the fact checking process and a list of project ideas addressing the pain points and bottlenecks we work with. In between developing new tools from scratch, there was plenty of time to share ideas, predictions and coffee. And by the end of the day, the group had produced a dozen working prototypes and demoed them to one of our fact checkers for some instant feedback. The tools included:
- a browser plug-in to inform users when they visit pages that Full Fact has written about;
- a chatbot that can answer questions on topics that Full Fact has investigated, but without going too far off-topic;
- some visualisation showing the (lack of) consistency of AI chatbots;
- a tool to help fact checkers carry out initial research using trusted sources; and
- a Claude Skill to quickly test if a draft fact check article meets the requirements of our internal style guide, and suggest any required edits.
One benefit to Full Fact is that we can already see the scope of many potential tools that we can go and build to support fact checkers here and in other organisations -- though of course much engineering remains to be done.
Two brief anecdotes summarise our learnings from the day. First, one of the participants asked Claude to "write a tool to help fact checkers choose which claims to check." This is a perennial problem and our existing tools are useful but fall far short of perfection. Claude replied with a plan for some code, and the dialogue continued something like:
Claude: This will help you prepare for your 9:45 team meeting.
Developer: What team meeting? I didn't mention any team meeting!
Claude: I found this blog post about how fact checkers work each day and used that to guide my plan.
Developer: (mind blown!)
This is a sign of how far AI programming tools have come: without a direct request, they can seek and find information from trusted sources to help solve the problems they're given. In this case, it found a very relevant and trustworthy blog post and acted accordingly. But presumably it could have been dragged badly in the wrong direction had it found a less relevant source instead.
Secondly, as we were wrapping up for the day, someone asked the group "Who actually wrote any code today?" Everyone had been using Claude Code as their primary tool and it turned out that just one person had hand-written just one line of code all day -- and that was only because the code was shorter than the prompt would have been. Of course, the rapid development of a proof-of-concept is not the same as writing production-ready code, but it shows just how powerful and widely-adopted AI coding has become in just the last few months.
Many thanks to Ian for organising this playgroup and thanks to everyone who took part -- hope to see you at the next one!