I am very much against all AI ‘assistance’ in writing and similar tasks. People need to use their brains and actually learn subjects rather than just copy-pasting whatever and LLM spit out.
Honestly, I don’t see it as bing much different from plagiarism as it’s people claiming the ‘work’ of a different entity as theirs.
Yes. I think asking someone using an LLM for sources can lead to an incredibly frustrating process I’ve seen many times:
“Could you provide a source for ____?”
“Sure. It’s in (out of print 500 page book)”
“Could you give me a page/chapter number?”
“Yes, it’s in [non-existent/irrelevant chapter]”
<…>
“Oh sorry, it seems I was mistaken. Try (out of print 600 page book)"
and so on
I’ve hidden a post that was too general. Please keep discussion focused on AI and iNat. See https://forum.inaturalist.org/t/park-field-guide-examples-using-inaturalist-data-with-ai-assistance/77023/13
Obviously AI is a huge and controversial topic and some broader thoughts about it and how it relates to the specific question maybe needed, but it should be there to provide context to the focus of the topic.
Yes - to be included. Words are important. We are all learning a new language for iNat - mine is mostly Botanese, but insects are also interesting.
For newbies I sometimes see a wall of text in the notes. I don’t check whether it is scraped from Wiki or AI. If, if only, the iNatters would indicate the source if not my own text. A small warning flag (to pair with Opted out of CID) so identifiers are forewarned - Shall I ID or Next ??
imminent ? Coming soon ? (will delete this later)
If I write a macro, set up a bot or other software to open observations and add comments, running these would create “machine generated content”
When I add a comment and press Save, this is not machine generated content, even if the comment was copied from somewhere else.
Unfortunately people don’t know that LLM generated responses are an unpredictabe mix of correct, incorrect and made up stuff - called AI slop.
A real person has to enlighten them.
I agree some guidelines are needed discouraging posting AI generated responses, asking users to disclose the source as a minimum.
It is tricky to handle people who just want one particular answer (one ID for whatever reason) when they have no clue about it. The less you know, the more likely you are going to be confident in an AI nonsense because none of it, including the counter arguments, make sense to you.
So if iNaturalist has a solution to that, I’m very interested as I deal with the same issue at work! 
I think the main place where AI-generated content is concerning is in observer notes. I see this a lot from student accounts, presumably because they’re required to add observer notes for their class. For example, I just prompted Gemini “write 30 words to explain why this caterpillar is a monarch” and it said:
“This caterpillar is a monarch because of its vibrant, iconic pattern of yellow, black, and white bands. It exclusively feeds on milkweed, which is essential for its life cycle.”
But I never showed it a picture of the caterpillar in question. So it creates the impression that I saw something with black, yellow, and white bands on a milkweed plant, while I may have seen something entirely different, misidentified it, and then copied nonsense into the notes section. I see this sort of observer note fairly frequently, especially when going through Unknowns, and I could see it misleading an identifier as to what someone is claiming to have observed.
I agree. However, don’t despair completely. My 16-year-old is against AI 100%. They won’t use it and scold me if I accidentally read the AI summary of a search instead of scrolling down to the original sources. They were taught Media Literacy in elementary school, which included verifying sources, questioning why something is being posted, detecting scams and general BS. They recognize AI generated content faster than I do. I’m sure that they’re in the minority, but there are definitely communities of teens (paradoxically, on-line communities) that are aware of the dangers of AI.
Thanks for explaining that. A failure by the educator, not the student then.
Never underestimate the laziness of an unmotivated student 
I think AI is going to be a thorn in our (everyone … not just iNat) sides for a long time to come. It’s part culture war, part technical innovation, part shortcutting/laziness, part knowledge replacement, etc. There are so many facets to this that it’s hard to know where to even start.
I think the key issue is transparency. Did I use AI to generate any of the information in my observation and/or in any ensuing discussion? On a basic level, we all use AI in our observations … CV is an AI image interpretation tool. Losing CV would mean all observers and identifiers would need to resort to identification books, which many people don’t have and may not know where to get (and a casual observer probably wouldn’t pay for them anyway). Happily, CV is a much better tool than most. The “answers” I get from Google Lens can be quite comical and I’ve considered sharing some of the more absurd for pure humour value.
I don’t understand the use of AI in an ID discussion. The person using the AI learns nothing about how to identify their sightings, and the identifier is arguing with a chatbot. An epic waste of time for all involved. This is where AI should be controlled … How we would do that is unknown. I’m not sure I could recognise an AI response if I saw one.
New technologies, processes, techniques, etc. take time to settle. We’re not there yet with AI and I don’t personally think we will be for another decade or more. I share some of the concerns in this thread about the “AI for everything” mindset. AI will have a use case, but it will take time to determine exactly what the parameters are and how to best use the capability. Until then, I think this conversation will repeat itself ad infinitum.
I have continued to mull over this in the past few days and I think I should make a confession. I have, out of the desire to be helpful, done almost exactly this. This is my scenario: I am aware that there are very few experts relative to the number of identifications posted each day. I skim through images that are awaiting ID. Usually, I stick with things that I feel very confident about, in an effort to reduce the load on true experts in a given field. But on a few occasions, I have become interested in one and thought- I know what this is! But I’m not entirely confident, so I do a google reverse search. And lo! I was wrong! But apparently so is the other identifier. So then, in an effort to be helpful, I say- I think this is what it is, and this is why, though deferentially, because I’m out of my zone. I do then, go back and remove my ID if it turns out I was completely incorrect, or a little bit incorrect. And I feel a little embarrassed, and I learn something. My reason for confessing is that, most people use AI out of a desire for good. So, if we can isolate the places where AI is problematic on a systemic level in the iNat user interface (UI) then I think some of these problems would be significantly reduced. Here are some observations I have that may improve the UI, if technically feasible and user acceptable:
Create a separate section for multiple species in a single “observation” and require that the species to be identified be delineated in some way at time of upload. Example: simple yes/no: is this one species or multiple? if answer is multiple, create observations for each species automatically.
Create separate and intuitive tagging for pathogens/stressors/galls/bores etc. the number of “something??” and “unknowns” are quite high- this leaves (pun intended) it to the volunteer identifier to potentially misinterpret what the observer’s intention is, and by extension what the requested ID is. I have had the experience of being chastened over ID’ing a plant, when they actually wanted to know what type of gall may or may not be present. For those of us who wish to be helpful, this creates an awkward and perhaps exclusionary situation which can be discouraging to both sides. This could also be improved on a simpler, back end level, by allowing the ID to be “organism causing leaf color change” or something of that nature.
2a) In parallel, make a specific annotation easily accessible in a UI sequence to mark as track, scat, skin, alive, dead, etc.
tag photos as “reviewed and undecided” if an expert has looked at, doesn’t know, and passes it over. put those in a separate cue. “experts aren’t sure” is a group of observations that shouldn’t be released to the wild unknown of potential identifiers.
to accomplish 3, users would have to be tagged by the system in some way for their expertise level and area. this is complicated, and I would defer to those who live and breathe iNat data to think about how that would actually work.
This is a continuation of UI optimization. Standardize basic inputs. For example, the “annotations” sections is very difficult to find/access to the casual user. Make it easy to tag- Example: “I think this is a flower/animal/etc.” Allow the greater detailed questions to be skipped. Already this greatly improves your data quality and helps channel id’s to the correct experts with a vested interest.
I have seen users discussing the use of the comments section to give general feedback- make this easy with a “like” button, or “share,” or ask a question. Engagement between observer/identifer is a highly valuable opportunity that is already available. Normalize this through standard common questions. Examples: 1) Situational clues needed- anything in the area like other tracks, scat, etc. Elevation? 2) How did you take this photo? 3) Do you have more photos that would help with this ID? 4) Preload comments to help the user for next time. Example- this is hard to identify in this photo, when observing (example: a flower) try to take photos that include (pre-agreed upon set).
Timely feedback: through AI or algorithm or combo, if an observation is lacking, trigger a warning or guidance at the time of upload. Example: This appears to be a (mushroom). Your current photo allows a 60% certainty. Upload (gills, stem, etc) to increase your identification score. This gives the observer the opportunity to take or upload additional photos to improve the observation in real time. And then people learn, and with increased skills comes increased engagement, better data.
Consider gamification for the most identifiable observations- which could increase engagement for students as well. I know I like positive metrics- so incentivize it. For all of the potential drawbacks of using iNaturalist in classrooms and for required projects, there are many opportunities to guide any user toward better observations, and by extension better IDs.
Thanks for all of your input and the work that you do.
Meghan
I agree but I believe there is an additional layer of this. Avid AI-users are missing out on critical opportunities for growth. Identifications mean little if you are unable to attain them without AI or google lens. There are lots of fantastic books out there to assist and solidify identification skills (for more common species). Why choose ease over quantifiable knowledge? Where do they believe that the information is sourced from? Experts and knowledgeable people in the field..
But even the knowledgeable people trip over homonyms without noticing. Or even noticing that their homonym has kicked the CID to something silly like Life due to Kingdom Disagreement 25K (but not all homonyms)
For newbies even knowing where to start to name That thing, is daunting.
A lot of your good points were intended to be addressed with observations fields. But considering the mess it is, it’s not much more than a tag nowadays.
Regarding the gamification, you’ll find a lot of discussion about it. And I believe the wisest move was to steer away from it. We want students to be interested in nature but if the teacher or the students want to play with it then they don’t need iNaturalist for that. An interactive checklist, a competition, a quiz or what not, it’s not bad at all; but it doesn’t need to be done by iNat. And since iNat is not monetised they can’t have the rules moderation than any type of game would require.
Going back to the main point, it’s normal to use all tools at our disposal, we just need to understand that tools have their limits and some tools aren’t a good fit for some issues. With AI there’s so much hype and communication about how awesome it is (and there’s a vested interest in maintaining that way if we want to avoid a global financial crisis for a bit longer) that we can easily forget how it works and how unreliable it is. People seem to understand relatively fast that the CV isn’t perfect, but they have full faith in a LLM. As the technology is less wow and new generations become more used to it, the issue will also be less prominent.
The temptation to get a shortcut to knowledge isn’t new, isn’t going to disappear and it’s something that will always be the distinction between learning and giving it a go.
That is not my experience.
Or at least, having CV suggestions corrected dozens of times for particular taxa does not stop some people from continuing to use the suggestions for their future observations of those taxa. I don’t know whether this is because they still trust it, or whether they realize it is not always right but use the suggestions anyway because they assume IDers will correct it and they think they are more likely to get a response with a wrong specific ID than a correct but broader one. Or whether they simply see it as the fastest and easiest way to put initial IDs on their observations, or whether they don’t know how to enter an ID that is not one of the suggestions. The outcome is the same regardless.
I think this is the correct take.
This has not been my experience. I think everyone is well aware that AI is wrong a lot, but the stuff they get from AI is “good enough” to check whatever box they need to check. I work with high schoolers every day, and they absolutely copy and paste answers from LLMs to cheat on their work on a daily basis. But none of them are under the illusion that it’s always correct. They just don’t particularly care if it’s right or not, and they figure it’s close enough to an answer that they’ve checked the “answer this question” box, so to speak. I think CV over-users are in the same boat. They know they’re supposed to add an ID. They don’t want to spend the time right now to work out an ID. So they use the AI response as “filler” and figure now it’s someone else’s problem. In my classroom, the same teenagers who copy a ChatGTP answer onto their homework to avoid doing it will call their friends over to laugh at a stupid incorrect thing that ChatGTP just said. They know it’s frequently wrong, it’s just that in some situations they don’t care. I’ve never met anyone who actually thinks they’re getting 100% correct information from a LLM, or anyone who thinks the CV is always right. People who over-use these tools to avoid thinking are doing so more out of apathy than out of full trust.
I’ more worried about adults than kids in their lack of understanding how LLM works.
I recently received “legal advice” from colleagues that was wrong, or “due diligence” that was also inaccurate as the local documentation was hallucinated. And if you’re not experienced in the field it would’ve have been harder to spot the inconsistencies, luckily if you know what you’re talking about then you can realise that the topic is not so binary and simple.
Hopefully iNat’s use is different.
Many people are touching on some aspect of AI being used or misused in a way that fundamentally shifts the way that some people are engaging on the platform, some of the time. This is certainly frustrating and is happening in many sectors. General lamentation is unlikely to improve the situation efficiently.
In bringing up common mistakes that can be made by anyone, @DianaStuder very astutely points out that the standardization of inputs and guidance to improve observations, whether in the observations/comments/ID fields, decreases mistakes for new users as well as experts. In other words, a high tide lifts all ships.
I am noticing that participation, encouragement, and safe spaces for learning are an underlying thread in this dialogue. How do we make meaningful and positive interactions happen more naturally on a platform that is built on a community nonprofit, mission driven foundation?
Vision: iNaturalist’s vision is a world where everyone can understand and sustain biodiversity through the practice of observing wild organisms and sharing information about them.
Mission: iNaturalist’s mission is to connect people to nature and advance biodiversity science and conservation.
There are always bad players. But you never know when someone who seems to be a newb/new user/outsider/disinterested student/AI enthusiast is just one kind word away from becoming an avid naturalist. The way other humans respond can have a major impact.
I said this on a thread about Merlin identifications several years ago, and I think my opinion on it still stands: “I don’t see the problem here. People are free to make identifications using any method they wish. If you think the ID is incorrect, just use your own ID to counteract it.”
If you disagree with the ID, just use your identification to disagree with it. Unless this issue is way more widespread than I realize, then this seems like the best solution. I’m not sure how you would even know if an ID was informed by AI or not. And I also doubt that all AI identifications will be wrong.