Election 2025: Use BallotBot to find your candidates
The Quarry has created this tool to help you find and save the candidates that best matches your priorities.


Initially based on what candidates have said in the manifesto booklet, some hustings and podcasts, this is being updated as more material about each person becomes available through hustings and interviews, so please keep checking back.
We're working on how it handles questions! For now, please try and follow the format of the preloaded examples, or to dive into what individuals say, try "what does xxxxx say about housing?".
For more on how it was developed, and to track updates, head to the bottom of the page.
Building the candidate finder - and recognising its limitations
One of the first conversations we had when the idea began to propagate of what has now become The Quarry was over politics and driving engagement.
Solving the problem of the overwhelming nature of the candidate booklet in the first island-wide election, where and how people would get information to make informed choices, was at the heart of that.
This was in the early days of ChatGPT. At the time the public version couldn't read a PDF, so we extracted the text from the 2020 booklet, attached it as a document to an assistant and set about asking it some questions. The result at the time was OK, enough to show some promise. But also begging the question of whether something that simple would be of any value - after all, anyone could do it.
The months ticked by and AI developed and our understanding of how best to use it did too. Other media outlets, particularly in Europe, were looking at election coverage and use of AI tools and chatbots too.
This project sat quietly in the background as other parts of the business took priority. It was late in April and we had to make a decision on whether to push on with this or not. Could we make it sophisticated and reliable enough to be a useful and valid tool? Remember, we are journalists first, not tech boffins.
The only way to find out was to really dive into the 2020 information. The first iteration didn't use AI in the functionality at all, although it had become much better at dealing with images and text and extracting those elements so it helped extract the data. In the simplest terms, we built a spreadsheat with categories and put the candidates views into the relevant box. Really, this still remains the heart of things.
It is important to remember this is one tool to help people make a choice, to signpost them towards more detailed research if they want. It is not giving you everything, everyone said about everything (that remains the dream). There are limits to how much data we can send to be interpreted at any one time, so we have built a cache on the key topics to serve answers quickly.
In the longer term we will develop this to help track what deputies say and how that aligns to their voting patterns over the term while also building in a policy tracking element to help monitor the decision making process. All will be available and searchable for the public to use.
We came at this from a journalism first approach. No doubt tech experts will have other solutions and hopefully can build on what we have done here, we'd welcome any input. And if you do spot any rogue responses, please let us know!
Update 1: We've added a basic search for when candidates have been silent on a major topic in their manifesto. If you ask "Which candidates don’t talk much about [insert topic]" it searches for how many times that word is mentioned, and if it is zero or one, will list their name. This comes with a MAJOR health warning, it is only working on general topic areas and is limited, sensitive to wording. Always try a more detailed individual search and link through to their pages.
Update 2: Clearer response messages and fallback on more specific queries (still work to be done here, including looking at pensions as a more detailed topic)
Update 3: Pensions is now covered in more detail and education returned to the topic cache so it should no longer error. Added some logging updates to better track which routes are being used when people ask questions.
Comments ()