Dr. Matthew Fielding, CABAH Policy Hub, School of Natural Sciences, University of Tasmania

Image: “[an artificial intelligence pen writing environmental policy” prompt. AI Creator Studio, QaamGo Web GmbH, 23 October 2023, https://www.img2go.com/ai-creator-studio

There’s no getting away from AI. Recent technological advancements and increased access to generative AI, such as ChatGPT and Bard, have caused AI to explode into the mainstream. The capacity for these algorithms to create novel text, audio and artwork has triggered a media frenzy and widespread digital panic. Over the next few years, we’re expected to see AI continue to creep into more aspects of our lives such as education and policymaking. One such area is environmental policy, with some heralding it as a potential saviour for the protection of our ecosystems.

Several obvious benefits of AI for environmental policy have been widely discussed. These include improved environmental monitoring or easier analysis and interpretation of large data sets. More and better information can support more effective policy decisions, especially when prompt action is required. Conversely, there are established risks. Providing AI with great swathes of data raises questions about data privacy, especially regarding sensitive information often relevant to policymaking. 

But how does generative AI perform in writing or refining environmental policies? Theoretically, this could save hours of human labour and lead to the quicker implementation of those policies – in some cases, that time could be the difference between losing or saving a threatened species.

Researchers recently tested AI’s potential to write company HR policies, finding that, in its current state, was better suited as a support agent, than a replacement. Was it a similar story for environmental policy?

To find out, I asked ChatGPT to help develop a policy document for the critically endangered, orange-bellied parrot (OBP), one of the few species of migratory parrot. The Australian Government’s current policy recognises degradation and loss of habitat, small population size and low genetic diversity, altered fire regimes in breeding habitat, and disease as key threats. They list priority actions such as:

• Habitat restoration.
• Provision of nest boxes in Tasmania for the wild population.
• Captive breeding and release – adapt methods and timing based on trial results.
• Managing genetics across both wild and captive populations.
• Disease management.

I asked Chat GPT to “use summary bullets to write the introduction to a policy document for the species recovery of the orange-bellied parrot”. At first glance, the AI-generated draft looked impressive. It introduced the species, listed key threats, and even mentioned a recovery strategy. But scratch below the surface and there was a lack of detail and the document was remarkably generic. 

I asked Chat GPT to elaborate on the “Comprehensive Recovery Strategy” it had mentioned in the summary bullets. Again, the result was generic and lacked creativity. However, it did manage to mention a few established recommendations for the management of the species, such as habitat restoration, including “controlled burns, to maintain suitable breeding grounds”, and managing genetic diversity. It failed to include nest boxes or disease management.

It also mentioned that invasive species could pose a threat to the parrot, opting to use “introduced predators like feral cats and European starlings” as examples. When I asked it to elaborate on these invasive species, it declared that foxes (a species that is absent from their breeding area) can lead to the loss of parrot nests. There were also several odd generalisations to the recovery strategy, such as “regulating pet trade”. While this is a common problem for many parrots across the world, OBPs, thankfully, aren’t facing such an issue. 

In this case, AI provided just a few of the known recommendations for the species and therefore would be useful for someone with expert prior knowledge. As with the HR policy documents, it may best serve as a writing co-pilot tool for planning and structuring environmental policies. Of course, over time, AI will become more intelligent and powerful, and these issues won’t be as obvious, however, there are still some issues that could remain.

Image: “environmental policy created by AI” prompt. AI Image Generator, AI Image Generator, 23 October 2023, https://www.aiimagegenerator.org

Firstly, generative AI is trained by past data. Given the state of Australia’s environment and the ongoing extinction rate, it’s clear that our current laws and policies simply aren’t cutting the cheese. If AI is just using past examples to form future policy documents, is it a step forward? Or is it just a meaningless sidestep to make our lives easier without having any benefit for the environment?

AI is unreliable when it comes to facts about the world.  AI regularly makes mistakes which means AI-generated documents still require a lot of human oversight. Additionally, the humans responsible for prompting AI and checking the relevant output will need adequate AI training as well as expertise in the subject matter. AI is not, and may never be, a substitute for in-house expertise.

Another key consideration is ensuring that we have adequate regulations and guidelines to govern the responsible use of AI for policymaking. Transparency is key, with AI-generated policies and guidelines being publicly identifiable. Of course, forming those guidelines around the responsible use of AI is one example of when we shouldn’t ask AI to help with writing a policy.

While AI still has some way to go, it will continue to have an increasingly crucial role in policymaking in the future. And while it might make our lives easier, it is integral that we don’t lose sight of what is paramount in forming environmental policy – the conservation and protection of our environment. If we adopt AI for the more monotonous tasks, we must ensure that we’re not only doing what is best for ourselves but what is best for the conservation of our environment.