I recently attended the Web Summit Vancouver and the atmosphere at the conference was that everyone is using AI… but are they really? Certainly, there is a lot of talk about how AI can be useful and I can see that there are industries using it more in their daily work. Environmental auditing, however, is a field of thoughtful, intelligent and cautious individuals, so certain drawbacks of using AI means that it has not yet become common, nor even desirable. Why?
Once you understand how AI works, it becomes clear that the technology needs to improve before it can be trusted with projects that have legal and serious environmental consequences if not done accurately.
So, how does AI work?
Well, in order to utilize this technology, you need to create an AI model, where you set up where the data will come from, what aspects will be processed and what will be analyzed.
It is not hard to see the potential this would have for conducting an audit, gathering information for an observation, running the analytics for a set of audit data or to generate audit tools (e.g., checklists). You could assess historical audit data, as well as other site monitoring information, to develop site risk profiles, which could then be used to prioritize future audits, schedules for next audits and what topics should be a focus, etc.
There are people who are building their audit protocols/checklists using this same data. But, how can you be sure that the information your AI tool gives is valid, or perhaps even the most update to date information (e.g., when asking it to tell you all the applicable requirements for audit use)? At this point in time, the answer is: you can’t.
Because of the way that AI works, there are numerous points of entry for errors. These errors are being called “hallucinations.” An AI hallucination is an incorrect or misleading result that AI models generate, which can be caused by insufficient training data, incorrect assumptions, as well as, biases in the data used to train the model. Who is to blame? Perhaps, the coder or the designer. Choices are made in the initial setup and what can happen is that these systems can, in the words of Gary Marcus (professor, NYU), “cluster piles of similar things together” sometimes just because the words are close together and the system doesn’t “sanity check” the work. So you end up with something grossly inaccurate.
Although it is difficult to find auditors who are using AI with their RegAuditor tool, there is some discussion out there in the field about having AI view an image of a part of the site or activity, then pulling previous audit findings from audit checklists and online regulation resources to come up with a list of requirements (regulatory or permit requirements, if the permit existed in the AI's library). It is also possible to use AI to deconstruct a permit, pulling out relevant requirements, but caution is required as the AI is not perfect. This way of working would give the user something to start with, but that something would need to be edited and refined. In other words, a “hybrid” use of AI, where AI works through the data, but a human interprets what is found and examines the viability of the findings.
Using AI for environmental auditing seems to be uncommon at the moment, but things are moving quickly in the tech world, so I would say, “stay tuned to this space.” There is certainly enough motivation to cut down on the paperwork review tasks, which means there are people out there working hard to improve how AI can be employed.
About the Author
Jane Dunne is a Senior Editor, who works on a diverse catalogue of environmental publications that are recognized as effective tools to ensure regulatory compliance with complex requirements.