Using AI in PopEcol

How we use AI in our research: analysis, writing, partners’ data, training, and what to focus on as scientists.
Author

Arpat Ozgul

Last updated

October 23, 2026

Draft, October 2026. Comments welcome!

AI tools have changed a lot in the past year, and they are changing how we do science. They can now write the code for a capture–recapture or integrated population model, run simulations, sort camera-trap images, process GPS data, search the literature and draft text. Things that used to take a PhD student months can now take days.

It’s exciting, and also a bit unsettling. Most of what we do rests on long-term field data, from the snow voles in Churwalden to meerkats, ibex, guineafowl, giraffes and wild dogs, the camera traps in the Kalahari and Okavango, and our resurrected rotifers from Lake Orta. We’d like to use these tools to go after questions we couldn’t tackle before with these data, not just to produce more of the same papers faster. And we want to make sure we don’t lose, along the way, the skills and judgement that make us good scientists.

Below are our thoughts on how we get there. They’re not set in stone, so please tell us what you think.

How we work with AI

  1. Make sure you understand your model. It doesn’t matter who wrote the code: you should be able to explain the model without AI help. For an IPM, that means knowing what each dataset contributes, how detection is handled, which vital rates the data actually inform (and which come only from priors), and how you checked the fit. If you can’t explain it, it isn’t finished yet.
  2. Think first, then ask. Write down your question and what you expect to find before you open the chat. Then use AI to push your thinking further.
  3. Ask it to argue with you. Ask for alternative explanations, for parameters that can’t be estimated, for ways your conclusions could be wrong. An assistant that always agrees with you is pleasant, but not very useful for science.
  4. Check against the biology. Simulate a population with known vital rates and see if your model gets them back. Predict years or sites you left out. For camera traps, check whether the automated labels change your ecological results, not just how many images they get right, especially for rare species.
  5. Keep your AI sessions. Chats and agent sessions are part of your work now, just like your field notebook or your R scripts. Save them with the project, so that you, your supervisor and your co-authors can see how an idea or an analysis came about.
  6. Write your own first draft. Writing is how you figure out what your results mean, so don’t skip that step. Once you have a draft, by all means use AI to iron it out: tighten the wording, fix the language, check the logic, spot the gaps. But we’d strongly advise against starting from an AI-written draft.
  7. You’re responsible for every claim. Check every reference, number and statement that came out of an AI. Say how you used AI in your methods, and check the journal’s policy.

Data and partners

  • A lot of our data belongs to, or is shared with, partners like the Kalahari Research Centre, Wild Nature Institute and Wild Entrust. Please don’t upload partner data to external AI tools unless the agreement allows it and we’ve talked about it. When in doubt, ask Arpat.
  • Keep den sites, collar locations and locations of threatened species out of external AI tools.
  • Make a note of which data went into which tool (and which version), so we can trace things back later.

Training and theses

  • Do the core steps yourself at least once. At some point during your thesis, write a capture–recapture likelihood, simulate a population and fit a model to it by hand. After that, use the tools.
  • We’ll look at your understanding, not just your output. Expect to walk us through your analysis and explain your choices, in our meetings and at your defence.
  • Mind the pace. With AI it’s easy to run lots of things in parallel and keep going late into the night. That’s a recipe for exhaustion and bad decisions. More output isn’t the goal. If it gets too much, talk to your supervisor.

What we should focus on more

If analysis gets cheaper, what really counts is what AI can’t do for us:

  • Our long-term field data. No model can produce decades of individual-based monitoring. These data become more valuable, not less, as analysis gets easier. Keeping the studies going, and keeping good relationships with our partners and field teams, matters more than ever.
  • Knowing your study system. The person who knows that a group split, that a camera was moved, or that it was an avalanche winter is the one who spots a result that doesn’t make sense. You only get that from the field and from the people there.
  • Designing good observations. What we measure next, where we put the cameras, how we schedule the collars, which animals we follow: that decides what any model can learn later.
  • Mechanisms, not just patterns. Going from detections to species interactions in the Kalahari, or from survival estimates to why populations decline or recover, means being clear about our assumptions and designing tests that can tell explanations apart. Our rotifer experiments let us test this directly.
  • Looking across systems. We have similar data on many species. The time we save could go into asking what they have in common: which demographic mechanisms explain persistence in snow voles, meerkats, ibex, guineafowl and the rest, and whether a model built for one species can predict another.
  • Management and conservation. Advising on ungulate management in Switzerland, working on giraffe and amphibian conservation, and collaborating with partners in South Africa, Kenya, Tanzania and Botswana all depend on people and trust.

Careers

Let’s be honest: the job market is changing fast. Coding and fitting complex models used to set our students apart, and these skills are now within everyone’s reach. We think the mix we train — field experience, knowing a system well, and thinking quantitatively — will stay valuable. But we should talk about it openly, as a group and one to one, and especially about how those of you early in your careers can prepare, inside or outside academia.

Next steps

In January we’ll have a group session on AI tools. We’ll show what they can do on our own projects, talk about how useful they really are, and share how each of us uses them. We’ll do this regularly, as the landscape is changing rather fast.