Recently, there has been a massive amount of discussion over the question of iNaturalist using AI to access expert identification information (https://forum.inaturalist.org/t/what-is-this-inaturalist-and-generative-ai/66140). So much so, that it’s become very difficult to keep up with the information, perspectives, and opinions on the matter. For this thread, I want to focus on solutions.
What specific features would make you either willing to tolerate or perhaps even comfortable with AI providing access to identification information and comments?
This should preferably be in list format (see first comment below). Feel free to copy and paste from the previous post. In time, perhaps this could form the basis for a description of what we want the end product to look like while also clarifying our “lines in the sand”. This is not a place for elaborate discussion. Answers should be simple and direct. Ideally, I would limit it to one comment per person, but I don’t want to complete eliminate dialog on the utility and limitations of particular solutions.
Suggestions as of 1 Jul 2025 (before comment 48)
The suggestions seem to roughly fall into 11 categories:
Attribution
Dislclaimers
Source accessibility
Correction/user input
User choice
AI quality
User notifications
Overall system function
Transparency, communication, and community involvement
Assessment
Other
1. Attribution and copywrite (3 comments)
2. Disclaimer (2 comments)
3. Source accessibility (1 comment)
4. Correction/user input (7 comments)
5. User choice (5 comments)
6. AI quality (2 comments)
7. User notifications (1 comment)
8. Overall system function (9 comments)
9. Transparency, communication, and community involvement (3 comments)
10. Assessment (1 comments)
11. Other
Generally a strong desire for more tools to enable user-created identification pages, guides, or tools. These is perhaps a little beyond the scope of this post (more like a feature request for something independent of the project that the iNat team wants to work on), but relevant to the discussion. For example,
An option to opt-out, so that a username and their contributions can in no way be fed, processed, or linked to the above-mentioned AI. (As a bonus, a way to hide users according to one’s own taste – including those users who opt-in this AI).
And roughly similar to your ‘2.’, a big, clear statement/disclaimer that the machine-generated content must not be considered authoritative nor sound. Self-evident maybe, but better repeat it. “Warning: iNat and its members accept no liability for your incorrect use of this machine-generated text, for damage to persons and property including but not limited to bodily harm, failed exams, job loss, toxic omelettes,…”
@nathantaylor Just to clarify, can we assume that you are referring specifically to generative AI, since other forms of AI already exist and interact on the platform?
I’m absolutely against any kind of generative AI. There are other technologies lumped under the umbrella of what people call “AI” that are fine, but genAI is completely a no-go as far as I’m concerned.
I agree with the sentiment of the other posters within this thread. Referring to the new proposed additions as simply “AI” misrepresents the differences and acceptability of Generative AI vs other forms of AI. I agree that AI is a very useful tool and has been used to build the backbone of the iNaturalist identification software, however I do not believe that generative AI will be constructive for holding up the standards of iNaturalist.
To answer your question, the only way that I would tolerate iNaturalist using Generative AI is by creating a separate identification app to implement the software while refusing to incorporate those observations into the main database or at least preventing them from becoming research grade. The process of manually reviewing every observation to ensure the AI got it right is unwieldy, and there would be very little preventing “experts” from using similar AI tools to confirm the false identifications made by the AI.
Multiple advanced anti-hallucination techniques should be implemented.
Curators or even users should have the option to hide LLM responses until further review, not just flag them.
Some way for users to opt out of their comments being used this way, or an opt-in system (the latter would probably cripple the whole project though TBH)
Nice to have: some way to “promote” comments that the AI can then draw from (though at that point, a non-AI solution may be simpler).
Nice to have: if the system is opt-out, iNaturalist could send a notification to any users referenced by the LLM with a link to the relevant page.
If generative AI is used to produce identification information (compiling data from comments, etc.), it should be clearly labeled. “Information compiled by AI”?
As far as identification itself goes, the CV we use to produce suggestions is enough. AI shouldn’t make ID’s.
I hadn’t considered this, but the more I think about it, the more I REALLY like this idea. That way we could see how our contributions are being used by the AI, for better and for worse.
What exactly do you think the proposal in the grant consists of? As far as I know, there are absolutely no plans to use generative AI for automatic identification or even to use generative AI to make ID suggestions. Given that iNat’s current guidelines explicitly forbid machine generated content and use of the CV to make IDs without any human input. It would be a radical break with iNat’s philosophy up until now to suddenly put complete faith in the output of a generative AI algorithm.
The proposal, as far as I understand it, is to explore ways to search existing comments made by identifiers and collect them into something like a natural-language summary of the content. In other words, it would be generating text with identification information that would be attached to a particular taxon. It would not be connected in any way with the mechanisms used by the existing CV to analyze photos and suggest IDs.
By the way, one of the problems you are concerned about already exists: we already have to manually review and correct AI-based identifications – in the form of CV suggestions – all the time. For those of us who ID taxa that the CV is really bad at, this is indeed a tedious process and requires constant vigilance to avoid situations where the CV suggestions become self-reinforcing (e.g., inexpert users using the CV suggestions to confirm observations).
When I imagine how AI could aid identification on the platform in a way that doesn’t make me wince, I’m imagining something like…
an AI that looks for comments on relevant observations. for example, the computer vision suggests both a common raven and an american crow. it looks for observations where both of these taxa are mentioned or suggested as IDs, along with having comments that the AI considers helpful (not just “nice photo!”). then,
it compiles a list of links to these observations, presenting them as “for discussions on how to differentiate the common raven and american crow, see these observations…”
What it doesn’t do is try to take the comments found in the first step and remix them into its own list of identification tips.
This means:
people’s words are not being taken out of context, remixed, or plagiarized
the information presented actually has a human source that can be referenced and spoken with (or tagged to come help with your own observation!)
the wealth of helpful comments and discussions on the site are being uplifted and made easier to find, instead of replaced or hidden behind a wall of computer generated text.
if the content delivered is a list of relevant observations, if people want the information they will have to click through and see it for themself. as opposed to genAI text with a link to these “sources” at the bottom, I feel people will not often follow the sources to check if the AI has represented the information accurately or made things up that were not part of the sources at all.
This ends up being more like a fancy search engine than genAI integration. But I think that would be a lot more helpful and intellectually honest than genAI.
If iNat could prompt a Physalia obs to consider the journal post I linked to in my comment - then I wouldn’t have to remember - (split to 4 new sp) was a … journal post by …
Or if that journal post was in the menu at About on the taxon page. I would prefer that to going to a Wiki.
Artificial intelligence is an amazing invention of humanity and should be used to help in strictly controlled activities. This tool should not be used to fool teachers, write scientific papers or solve expert problems. It is supposed to be a tool under the control of specialists and not an independent creation that will replace humans in many matters. People in iNat should be creative and cope in a classic way and in cooperation with other users. For me, introducing artificial intelligence wherever possible is a very unwise way of acting…
We already use AI to suggest identifications. Given that such identifications are drawn from a list, and then manually selected/curated by people, I think this is pretty reasonable.
I’m not sure I understand what the newly proposed use case is, but if it’s just scraping all comments for “oh can we use this information to train an AI to ID stuff”, that feels a little unethical.
However, I do think that it might be better to put an opt-in box (i.e. defaulting to “off”) for people to say “hey, use this info to help the AI learn to ID this taxon”. For example, people who just say “nice photo” would likely just ignore it, but if someone says stuff like “small lizard with translucent skin found indoors at night in south Florida = probably a Hemidactylus gecko”, they can voluntarily check the box to choose to say “hey, you can use what I said to teach the AI to understand this”.
And once you have that built up, you can also potentially have the AI provide reasoning in natural language for a suggested ID – for example, the AI might suggest Hemidactylus with a reason stated like: “[AI-generated reasoning for this suggestion] location is south Florida, observation time is night, lizard has yellow-pink skin”.
None. No generative AI. Usage of LLM is theft. There are many many solutions that would both work better and not damage the environment by using irresponsible amounts of energy. This is, in my view, the worst possible solution to the problem of specific ID details being inaccessible. Just adding a comment section/forum under species and genus pages would both be more reliable and also not a heinous waste of resources. Genuinely there is nothing that would warrant this. Just make an iNat Wiki or something and stop wasting our time, please.
It appears that the grant will go towards adding explanatory text to autosuggestions. Is that correct? Are there any other features that will be added using grant money?
If so, I am more positive than negative about the proposal BUT here are the things that would alleviate my discomfort.
IF POSSIBLE, the text should include bibliographical information - which textbooks / field guides were ingested, at least. Even better would be specific references to which texts were used.
The algorithm should include feedback from human identifiers. It needs to notice when it’s getting voted down.
In connection to this, I would suggest an “upvote/downvote” button that is specific to the text. It is entirely possible that the AI model could get the right ID but for the wrong reasons - in which case the text would mislead.
If I’m understanding correctly, your proposal is not changing the way that IDs are made, but the way that they are explained. THAT NEEDS TO BE MADE CLEAR to the iNat community.
Thank you. I appreciate this step in the direction of “AI that explains itself.”
Oh yeah, I would like to emphasize this point as well. AI is known to be power-hungry, so we should not be making use of it lightly.
The stuff I said in my earlier post, despite my being able to think of some new uses for AI on the iNaturalist platform, I am not convinced about it being useful enough to be worth the cost.