This demo will mainly be to just give a quick overview of what I’ve built – a demo of how Agentic AI is now at the point where it can reliably pass the Red List Assessor Exam. But then it’ll be great to discuss what this means for the IUCN and the Red List bigger picture.

But before getting into that, I thought I’d briefly give a bit of context on how I started working on this.

I arrived here in Cambridge to start my PhD in Cambridge, eager to do impactful research in AI for conservation. But, leaning on my startup experience, I first wanted to look at what the actual big problems in conservation are, before just force-fitting AI in as a panacea solution. And so it felt like very serendipitous timing that the start of my PhD coincided with the IUCN’s World Conservation Congress, which only happens every four years! I thus registered and attended the conference remotely.

Attending the Opening Ceremony of the IUCN’s World Conservation Congress from a cafe at Jesus College Cambridge

Here, the Red List repeatedly came up as a crucial resource for conservation, but facing significant challenges with scaling assessments and re-assessments. This ChatGPT thread spells it out pretty well, explaining why the Red List is important, what it’s big challenges are, and how AI could help.

Why the Red List is important

What the Red List’s challenges are

How AI could help assessors

This is all good and well in theory, and I know the IUCN is starting to think about this too. Eg we see in the Red List’s 60 Years of Success document:

But next I thought okay at the very least, to ground myself more, I should learn more about the actual Red List workflow – see what it takes to become a certified Red List assessor and see from my own experience where AI could be helpful and where it wouldn’t.

So I found the online training course, and did the 12-hour course over the course of a couple weeks:

At the end I did the exam and got 80%.

But then I naturally wanted to see how well AI could do! Turns out pretty good!!

Here you can see it averages 86%, better than I can do. At the end I can demo it answering the questions in real-time if you want.

But stepping back, this is just a small example. In reality, the question is: can Agentic AI assist assessors? I see 3 ways: