Large language models: looming large in terms of threats for the environment?
Do we need a lobby for environmental topics when it comes to LLMs?
We recently published a paper summarizing some initial thoughts about the problems and opportunities for the environment associated with the more widespread use of large language models (LLMs), like those used in the chatbots ChatGPT by OpenAI or BingChat by Microsoft (and many others coming up).
Briefly, in this paper, we see benefits for environmental education and in terms of enhanced access to tailored information, or for streamlining the workflow of environmental scientists, thus enabling faster research progress. Risks are energy consumption (a direct environmental impact), enhancing digital divide effects, and the possibility of using LLMs to very efficiently generate and spread biased information with a simulated air of authority.
Were we too optimistic?
LLMs are becoming incorporated fast into other apps, perhaps even faster than expected. Pretty soon, you will also see them pop up in your productivity apps such as Word or Powerpoint, or you can already see them in note taking apps, and the list of such integrations will be quite long soon. BingChat is probably already at your fingertips when doing web searchers. So this use of LLMs will become - I think it is fair to say - truly ubiquitous. A couple of months ago I think only few people saw that one coming. It is one thing to use a dedicated user interface such as ChatGPT, but it is next-level to be surrounded by such model interfaces and outputs in all kinds of different contexts.
It is also becoming clear that there will be many of these models, and there is no general oversight, so - as we have already hinted in the paper - there will likely be some LLMs with the safeties off (or less effective), and they can also be trained on specific datasets. This is already fairly easily possible now.
Taken together, the unexpectedly rapid spread and integration of LLMs into all kinds of applications, and the potentially wild growth of LLMs with limited supervision - could this constitute a greater than anticipated threat for the environment…?
I believe the threat is mostly from bad actors, organizations (or individuals) with an interest in downplaying environmental topics or concerns.
It would be great to have a certification process in place that gives LLMs some sort of a ‘seal of impartiality’ for different topics, like the environment, - but who will be responsible for this? and what criteria will be used, what are the benchmarks? and will environmental topics (such as biodiversity, global change, pollution) even be on the radar when it comes to such regulations, as the opportunity for bias is multi-dimensional and real? do we need an environmental LLM lobby?
Why even worry about this? LLM output comes with an air of (simulated) expertise that seems super-human, and thus it will likely be perceived as an ‘impartial’ expert opinion by the public, even the term ‘artificial intelligence’ carries this connotation. When it becomes routine that we access information via LLMs, it will really matter a lot how exactly these models produce their results. And it will be extremely important to widely train people (from an early age) to become knowledgeable about the limitations of LLMs.
What should the environmental LLM lobby look like? How can we best capitalize on the positive opportunities afforded by LLMs for the environment and for our science?
So far we have only thought about text. What about other domains, such as images or video? Is there something we don’t yet have on our radar?
What do you think?