I oppose and condemn iNat’s project to “improve” species suggestions using generative AI. This project - funded by, advised by, and open to being infrastructurally supported by Google - is a mistake.
For those familiar with species identification and AI tools, it should be obvious that trying to use an LLM to explain the reasoning behind a species ID is a blatantly silly idea. Without constant verification by expert IDers, it will be untrustworthy, misleading, and unaccountable. For it to be usable, it would take so much AI-babysitting work from expert IDers that it would have been much easier, simpler, more reliable, and more community-focused for those expert IDers to have built a user-edited wiki in the first place.
While the iNat computer vision is an amazing tool for science, generative AI (a very different technology) is, as iNat plans to implement it, decidedly not. In fact, it is anti-science: it stands to make community science worse.
If I want to learn why my observation is what it is, I want to hear from a person. A person has reasons, and can articulate those reasons. A person can explain why. I want to hear a taxonomist’s nuance, an amateur’s passion, an expert’s tip. Most importantly, a person can cite their sources. A person can admit when they aren’t sure, when it’s a “most likely this” kind of situation, when it’s more vibes and gestalt than professionally keyed out.
genAI can do none of this, simply because genAI does not have reasons and cannot articulate reasons. Forcibly training a glorified autocomplete, until it appears like it can sort of articulate reasons, most of the time, is an extremely inefficient and alienating use of your community of IDers.
iNat is precious to me. I love it. It has woven webs of trust, balanced accessibility with expertise, and built a fun, empowering ecosystem of learning.
It is this sense of community that is iNat’s greatest asset. It is the very thing that makes it unique. And it is exactly this that is threatened by its genAI project.
We didn’t ask for this. We don’t want it. Please get rid of it.
Many have voiced concerns with genAI that I consider broader and more structural in nature. Some of these include data privacy concerns, intellectual property / labor concerns, environmental concerns, and concerns related to Google, including its well-known business model of allowing small orgs to use their compute power while eventually making them dependent on Google computing services.
I believe many of these wider, systemic concerns are, to varying degrees, valid reasons to condemn this use of GenAI.
My main point is that even if none of these concerns existed, I still wouldn’t support this use of GenAI. Even if training the LLM had zero carbon costs and was funded by another company, I would still condemn it, simply because it can do nothing but make community science worse.
Even if you haven’t seen what the AI tech bubble has been doing to our world, you should still oppose this project, for the utterly pragmatic reason that a fancy autocomplete does not and cannot reason about why a particular observation was identified to its particular taxon, nor can it articulate such a reason.
P.S. I’d happily contribute my time and expertise to a user-edited wiki or similar community-driven project to improve iNat species suggestions.
P.P.S. the one use case I can envision for LLMs on iNat is as a functionality of an iNat-internal search engine that helps users easily search iNat obs and pages to find comments, notes, ID explanations, etc. that are relevant to their search keywords. It is in this kind of search engine role - as a circumscribed library-research-type tool - that a glorified autocomplete may not be completely useless. Of course, the AI slop vision laid out by the iNat team in their blog post has little to do with this idea. This hypothetical use case for genAI on iNat is also not immune to the broader structural concerns mentioned earlier.
