I am not sure if this is discussed elsewhere here, but I have already seen hundreds of observations in last few months in India where the observer is passing their image to some AI and simply pasting back those long paragraphs from AI as is back in notes.
I am not trying to judge quality of each observer notes and their ways, but this blatant copying is inturn gonna spoil the value of biodiversity data (all notes are synced to GBIF) and obvious biases of such AI generated notes (very often wrong)
I sometimes remind users that Notes is not for that, but most often they ignore it or even deny that itâs not AI(probably assignments). I really think iNat should also consider adding DQA fields for these AI generated notes just like we got for images AI manipulation now. Maybe such notes flag will simply hide/remove notes and leave observation intact (but again since text is different from image, differences in ways to handle that? - a tampered image is tampered as whole and rejected, while a single AI paragraph in whole dozen paragraphs user notes can be incised more effectively)
people have been adding notes about how they identified something or how to identify something even before AIs became commonly available. how would you be able to tell that notes were generated by AI? are the notes themselves the problematic thing in your opinion? or is it that someone used AI for their workflow? or is it that folks are copying and pasting chunks of text, regardless of where that text came from?
One can tell things upon seeing enough samples of AI texts (I am in that field and I can DM you some sample notes I have seen and you can sense it too, donât want to call them on forum for their privacy) and that isnât primary question of how one can tell or prove - itâs actually same question for implemented current image manipulation flag. Idk how that is handled or supposed to be handled currently.
Yes notes which are copy pastes from AI in particular and yes some people may have probably copypasted blogs or wiki or something before AI came but this one seems different problem where copy paste becomes very wrong - it sometimes states things of species from another continent.
Well either way AI is making it easier for those generative notes and they are going to increase so. whether some of that just falls as wrong information just like any other ID or comment flow on iNat or whether it should be clubbed under manipulation flags is not I can fully form opinion of, and hence this thread
I think this is a good question. Iâve only seen this type of posting (seemingly AI generated text blocks in notes) once, but given the question, it seems likely that it may become more prevalent.
My own take it that this shouldnât be allowed/discouraged, though enforcing that would be challenging, due to the difficulty of being 100% sure that any text is AI-generated (though I agree, itâs possible to be reasonably sure).
One reason I think this should be discouraged is because Notes text does go to GBIF as noted above. Users of GBIF will be assuming that Notes are created by the user, not AI, and so may not make good decisions based on that content.
Additionally, and analogously to problems with posting pics that arenât their own, most users claim some type of license over their observations when posting on iNat due to default licenses. There have always been questions about what observation (not media) licenses might mean in practice, since facts arenât generally copyrightable. One thing that probably is copyrightable is text in the Notes. However, in the US at least, AI-generated text is not copyrightable unless there is substantial human involvement (ie, rearranging/editing the text, not just writing the prompt to the AI to generate the text). So claiming a license for AI-generated text shouldnât be done. This is sort of an obscure problem, but, if thereâs an easy solution, it could be worth considering.
I think the lowest hanging fruit would just be to have a note somewhere in iNatâs guidelines that Notes text shouldnât be AI-generated or that prohibitions on AI-generated content also include the notes field. That way, users can at least be pointed to that. Curators would then also have grounds to address the issue for users that might be doing this at a large scale.
I agree, itâs frustrating when folks copy blocks of text from another source into the Notes/Description field (regardless of the source). Itâs also frustrating when they enter goofy messages with multiple emojis, or long winded discourses that are devoid of useful content. As far as âincorrectâ content goes, folks have been doing that without the help of AI. Folks frequently duplicated their ID guess in the notes (along with related discourse) and donât revise it after their ID has been corrected. While making it possible to flag the content of the Notes could help keep âgarbageâ out of some un-curated databases like GBIF, this might also be subject to abuse (users flagging Notes that are simply inane).
I manage a database that aggregates observation data from a number of sources including iNat. Iâve managed to automated many of the tasks associated with cleaning up the data and weeding out the âgarbageâ from the iNat data, but I still have to curate the âNotesâ manually. Itâs probably one of the most tedious and time consuming tasks in the whole process. If such a flag was available, I might be tempted to âmisuseâ it to flag Notes that create extra work for me.
Perhaps I should enlist AI to perform this tedious task for meâŚ
Yes. A guideline or rule at minimum helps everyone to point. Maybe combined with even a blog post mentioning Good ways of writing notes or highlighting existing some well written notes with some notes that helped a lot in identifying species - inat can find all those someway.
Yes, a model to detect AI notes can be made by iNat and when a flag is raised by another inat user on note, it can baseline check if itâs AI generated - and if the detector thinks say higher than 80% probability of AI note it can then keep flag for more humans to decide or reject it but then keep track of count and if more users raise same flag it can then validly open that flag then with these posterior higher probability. Ofc itâs all possible and very implementable if the stance of community and inat on notes is on that side. And even though a model to detect AI text itself maybe biased at start it can be improved just like inat CV is with more flags and their decisions of those flags, and such model only exists to control both abuse of flag and limit curator load and not a decision maker while also having something in place to not let things flywheel. Although the problem of whether a human is able to truly detect all AI texts and pass judgements with group unlike current IDs which can be easily referenced to literature is some caveat.
i did see one observation for which the observer accepted the CV suggestion (correct), but then changed it to a different species in the same family based on info from chatGPT or some other bs-machine. i left a comment dissecting the slop and pointing out the numerous times it was confidently providing information that was opposite to the truth.
i find it amazing and depressing how quickly people accept the âanswerâ given to them from a chatbot on a topic they have no knowledge in. ask it some questions you know the answers to, and see how good they are.
I hope this isnât too much of a digression, but I didnât realize notes were uploaded to GBIF, and now Iâm paranoid about what Iâve been using notes for. I found a forum discussion ( https://forum.inaturalist.org/t/tags-and-notes-what-and-why/66369 ) that indicates Iâm generally following the intended use (e.g. this one), but Iâve also sometimes included things that, for instance, I intended as an apology to the IDers, but would be useless for GBIF (e.g. this one).
In the interests of GBIFers, should I be more careful with my notes?
If you give a hammer to a 3-year-old, it will most likely be used to destroy.
That same hammer, in the hands of an experienced craftsman, can create great beauty.
So there is nothing inherently wrong with, e.g. using AI to write clearer notes if you are already a taxon expert and English is not your first language.
The quality (or the lack of quality) of the AI notes relies on the expertise of the observer and the identifier.
Using a tool for a use for which it was not intended (e.g., using a chefâs knife to pry the lid off a jar) is not the fault of the tool.
but that is not how these AI notes come across. For âmyâ students - either teachers have (wrongly) encouraged them to do this. Or the kids havenât been taught how to evaluate / discard AI text.
On other obs the notes ask - What is this used for ? Can I eat this ? Perhaps they add the AI answer as âseriousâ notes.
Iâm in the same boat! I had no idea notes were synced with GBIFâI thought they were for personal comments like âreally windy and rainy that dayâ or âsaw another but couldnât get a photo.â If these arenât intended to be used for personal comments (including jokes and whatever emojis the observer feels fit the situation), iNat needs to make that clearer as part of the onboarding process. Iâve been here since 2019 and Iâve never heard about this before⌠and now I feel a bit goofy.
If I see above issue description plus the tons of other unresolved systemic problems and conflicts around any LLM type AI generated content in INaturalist current and future, discussed in the forum since June 2025, it is absolutely beyond my imagination how this shall work in future based on current intentions and controverse oppinions.
If this platform will not systematically shield its content from any such LLM content soonest, it will not allow that - we or anybody else can do worthwile science - with the documented observations in it.
Scientists, private or from academia, will observe that GBIF data from INaturalist are no more relevant or trustworthy than information from any other hobbyist forum site on lizards or orchids. Maybe even less.
And nobody with clear mind will consider those for scientific analysis without further in-depth proof.
Maybe create an AI based filter to screen for any sort of LLM generated content, thereby flag it and make reviewers sensitive to questionable content? I wish it would be as easy as I write it.
YES
And for this exact reason I do not repair my computer with a hammer.
And I do not use LLM generated content to defend my scientific positions at work and in INaturalist.
The answer to the question for what content from LLM tools is good will likely be given in five or fifty years from now. We will learn by experience because logic and science are no longer âen vogueâ.
And we should not expect an objective answer becaus the winners will dictate the text.
I would say itâs up to you. Consider why you are posting observations. Do you think of yourself as doing âCitizen Scienceâ ? If so, you may want to compose your Notes accordingly. That doesnât mean you have to take yourself overly seriously, but just keep in mind that what youâre including in your observation may go beyond iNat.
I also wish it were easier to remove AI generated notes. What I detect more often is information about the plantâs structure or ecology that must have been copied from Wikipedia or somewhere, usually to comply with the rules for the assignment. Theyâre so obviously not related to the individual observation that they donât bother me. I mean, itâs obviously that the plant photoâd did not lead to the data.
I donât see any problem with any kind of notes being sent to GBIF (I mean, assuming theyâre your own words). For many (most?) cases, the researcher wonât see the notes or wonât use them if he does. If he chooses to read notes or comments, he chooses to read all kind of things we might post. Unless youâre posting things youâd be really embarrassed for others to see (which you shouldnât be doing anyway), surely this is OK.
Exactly ofc. To be clear, we are not judging a random AI model accuracy scores in this thread nor a single user copying those notes and their expertise on relying there.
The discussion is more towards position of behavioural digressions occuring already (as in question) and how those should be handled eventually. If a teacher or project manager failing to check these assignments observations properly, itâs largely part of their fault, but the broader theme of ease and such generative notes as a problem exists whatever the cause of that observation and the methods of how we respond to those will be partly responsible on how the behaviour changes for better, as in your case making every tool user more educational than mere copier from lack of any feedback to them.
I tend to completely ignore AI notes and not factor it at all in my IDs. Especially when its a huge wall of text with emojis and other AI signs. Although i will say, notes in general are rarely helpful for what I identify anyways.
