Hello,
I’d like to advertise this paper “Perspectives on Crowdsourced Citizen Science and the Data Governance of AI Applications”, may I?
There are a few gold nuggets hiding in there, spot-on with the topic of the day.
2 Likes
I think this is the main takeaway from all this. iNat royally screwed this whole thing up by not bringing it up beforehand.
5 Likes
Yesterday on an inat india group this news broke, I recognised the potential for flamewars knowing that I too felt strongly, feel strongly. I read the comments here, the comments on the blog, the response by the inat team, the moderators, etc. Still feeling the need to say my bits so wrote a long note of how much inat means to me personally but also trying to share thoughts which many people have echoed better than I can and more eloquently but still i choose to say.
Link : Response to iNaturalist’s Google AI Grant and the discussion happening
13 Likes
why wouldn’t you discuss those now? what we’re talking about is defining and measuring success. you should definitely adjust these as necessary as you proceed, but if you don’t start a project with a fairly good idea of what success looks like and how to measure it, then to me, it seems like you’ve already lost your way (although it’s not impossible to recover).
i’m not against this, but if this would have been the first time you’ve introduced the general idea of gen AI to your audience, then i think you’re late here, too. the best organizations i’ve worked with and seen in action have master plans for work that go years, if not decades, into the future. granted, these plans can change as conditions change, and the fine details get worked out in real time, but if you’re thinking about, say, utilizing generative AI, that sort of thing doesn’t happen overnight, and you should be communicating with stakeholders months, if not years, in advance that you want generative AI as part of a master plan. if you don’t already have a sense of how your stakeholders feel about generative AI and educating, as needed, before you’ve even found the Google grant opportunity, it’s already too late.
does a master plan exist? it’s possible i’m just out of the loop or forgetful, but i don’t remember hearing of one, and i can’t find one just quickly looking on the website. i get that not all organizations can afford to put the effort into do this sort of thing, but it sure helps clarify all sorts of things if you have one.
the closest thing i’ve seen was a list of things that iNat wanted to accomplish back in the day, resulting from a staff retreat back in 2019, i think. a lot of those things have never been accomplished, even 5+ years later, and that’s fine. things change. but without updating and sharing even a basic roadmap periodically, you effectively never hold yourself accountable for anything and will always just be in reactive, not proactive, mode.
i hope that a small board like the one iNat has would have been aware of the grant proposal long before now. i hope that they discussed something about the potential negative public views of gen AI and in the end decided to move forward anyway.
to the extent that they might have underestimated the negative public sentiment, i don’t necessarily fault them for that (in this specific case, in the absence of better information), but if their next discussion is more about reacting to this situation rather than how to position themselves better for the future (to avoid similar situations), then i think they’ll have missed an opportunity.
…
all this said, i rhink a lot of the conversation about generative AI really needs to be happening outside the context of iNaturalist. if anyone here shared a lot of thoughts in this iNat forum and haven’t shared similar thoughts to your legislative representatives and companies involved in AI, maybe that’s something worth considering for everyone’s sake.
10 Likes
I should clarify that it can pull from and synthesize multiple data sources/comments, I was using one comment for simplicity and I apologized that caused confusion.
I personally hope it is similar to Facebook’s new group bot that comments things like:
“Hi! It looks like iNaturalist users have explained how to tell Species X apart from the very similar Species Y in these comments: hyperlink to comment 1, hyperlink to comment 2, hyperlink to comment 3,….”
Or even “Species X and Species Y can be told apart by the markings on the back. See this hyperlinked comment for a detailed explanation.”
7 Likes
I imagine for most people, they shared the sentiments here because iNaturalist seems like a place where it might actually be listened to, rather than the latter, which really seems ultimately like an absolute waste of time and energy. They know the situation, they do not care, they will not listen. (Companies involved in AI, legislative bodies, etc)
17 Likes
Part of the problem is they are contradicting themselves. They say they want it to explain CV’s decisions, but also that it would involve in some form, using existing identifier comments. These two things aren’t related. Why I identify something and why CV identifies something are entirely different processes, with at best trivial overlap
Either it will be presented disingenuously, saying “Here is why CV picked this ID” and then putting something generated from a model trained on old identifier comments, which may not even be true for the observation in question.
Or it will explain the criteria CV is using, something which probably isn’t technologically feasible and also has the potential to be very unhelpful, because CV doesn’t use actual diagnostic criteria and regularly identifies my planthoppers as turkey vultures because they are both the same shade of red.
Both of these proposed implementations are bad, would cost far, far more time money and effort than alternatives(like a wiki), and have the potential to spread large amounts of misinformation.
And to avoid spamming the thread I wanted to take an aside: Part of the problem here is iNaturalist’s absolutely terrible communication. The proposed technology is bad, but the way they went about it has destroyed a great deal of trust in them. They announced it initially on a platform that many iNaturalist users don’t use, with no indication, not even a link to the bsky post, on the actual website. We had to find out through this forum post. So clearly they aren’t going to be completely transparent and I assume where attempting to initially put the news somewhere they thought would have a more positive reaction(so they could point to it and say “see, the people want it!”). This brings into question any further communication. Like the upcoming qna, how can we trust them to be fair and transparent? What if they only field question from supportive commenters or even astroturf the suggestions? I wouldn’t have ever thought iNaturalist would behave like that, but I also didn’t think they would’ve tried to announce something this controversial through bsky and nowhere else
I don’t know what to think. I think they have absolutely botched the PR here and even if they manage to completely reassure people on this course of action I don’t know if they can recover the trust of the majority
17 Likes
I just heard about this on the news, and created a forum account to chime in.
I’m ok with using machine learning to help identify species from photos, but I do NOT want to participate in a project that embraces large language models and generative AI in writing descriptions, making comments, etc.
I’ll be watching the news carefully and letting my outdoors friends know about this.
Just like with Wikipedia, we want iNaturalist to stay trustworthy.
17 Likes
I hope this isn’t against the forum guidelines to say, but I’m not some outside agitator or rabble rouser; I’m someone who absolutely adores this site/app, who donates multiple times a year, who evangelizes about it constantly to other people (I’d gotten two new users on there last week alone), and who very sincerely believes in its mission and previously would have said I wholeheartedly believed in its values. I’ve made friends and connections in my area through its use, I’ve gotten more into nature photography thanks to the incentives it creates to engage with the community and help contribute to citizen science, and I’ve genuinely fostered a closer connection to the natural world around me after my time here. I LOVE this app, and the community that exists around it.
Watching iNaturalist burn through years of goodwill and trust in the course of the last 48 hours has been genuinely depressing.
Part of the backlash can undoubtedly just be chalked up to poor communication in the initial announcement, and confusion about what exactly this new partnership entails. I have read the updates, and some of my initial fears and hesitations were indeed assuaged by the clarifications outlined there. I do think a lot of the “Google bad!” hyperbole and misunderstandings that all user data would now being processed to train external models are objections that now have to be filtered out, to a degree, for irrelevance. I understand the myriad ways that iNat utilizes, and always has, machine learning already, and am not in any way befuddled by this distinction. But there remains a core signal in the noise of all these complains that really does still leave me aching, and that means I too will be deleting my contributions and my account if these changes move forward.
The clear and specific environmental harms of GenAI use are extremely tangible and known, all for potential benefits that are nebulous at best and nonexistent at worst. As with almost all use cases I’ve heard for GenAI, the rebuttal I am left with is still, always, “but we can already DO that, and it doesn’t cost us enormous cascading energy drain, pollution, and contamination risk.” There have been many thoughtful suggestions already about about user-sourced solutions that would address the same need the iNat team is hoping that a generative LLM partnership would be able to fill, and they’re good suggestions. But they don’t feel like they get at the heart of why this decision feels to so many people like a betrayal.
Moving forward with incorporating this technology in spite of known environmental harm factors seems deeply antithetical to the values of responsibile stewardship and preservation that the team has always purported to hold dear. It makes me question the commitment of the staff and developers to their supposedly core beliefs in a way that feels personally hurtful specifically because I granted iNaturalist a level of trust I never afford to corporations. And perhaps that’s on me, for being naive enough to think any board of directors is capable of restraint and measured action in the face of GenAI’s vague and expansive promises and 1.5mil of sweet sweet Google money… but here we are, regardless.
28 Likes
I will try to not repeat any of the points made above about the myriad of issues that GenAI/LLMs bring. This comment captures most of my concerns. I can’t square the environmental damage of GenAI with an app that helps us connect to nature. But the depressing take about LLMs somehow being more efficient than humans, even if true, misses the point about the current situation. The humans and the humanity at the center of this app.
I am relatively new here, but I would like to highlight this point from high above the thread:
-
In my opinion, and in light of the above, one thing to come out of this thread should be absolute transparency about when, how and why this decision was taken. How does this square with the current development philosophy and why was this grant prioritized over existing issues? What are the terms of the grant and what does iNat need to show after the grant is over?
-
iNat seems to be trying to solve a human problem (or at least using this as the motivation for accepting the grant). There’s been some talk about “efficiency”, “scaling”, “optimization”. Personally, the way I engage with iNat is deliberately “slow” and relaxed. This isn’t my job, this is something I want to enjoy. I don’t need to be efficient at it to do that. Good design, the wiki that has been asked over and over again, and good search seem like low hanging fruits that would help here. The wiki seems to be such an easy yes, that I’m surprised it hasn’t been accepted as a suggestion, at least to mitigate some of the backlash.
-
and with these users in mind, what are the expected changes for the ways we use the app? These changes aren’t magical and instantaneous. Users will change their behavior around the constraints a piece of software provides. So assuming for a second that the technology is a good solution for the problem you are trying to solve, how does this fit in the app? Are users and experts now babysitters for the LLM until it’s good enough for particular species? Does that increase or reduce the motivation of users to participate? Our intrinsic motivation is different if we’re correcting, grading or labelling the mistakes of an algorithm than if we’re engaging with the community. Like @jgw_atx said, it will also change how people interact in the app itself. Will that change how the community perceives itself (less human or friendly, etc) when we read each other? How are you measuring these effects in your evaluation of the tool? What other human aspects are you considering, if any?
I likely won’t continue with iNat when you go down this path, but for those that choose to stay, the answers to these questions matter.
20 Likes
I really appreciate your thoughtful post, and ability to put aside several layers of reasonable concern that many of us had stemming from the terrible way this project was announced (Are we training Google’s Gen AI, etc.?
I’ve advocated for a user-sourced ID assistance component for a while, and I’m frustrated that iNat staff has chosen to ignore many similar requests and forge ahead first with a Gen AI approach that seems much less likely to produce great results, for the reasons many have stated.
And like you, I’m also concerned about the environmental impact this initiative may have. I’m also aware that there’s no single answer to “What is the environmental impact of Gen AI?” and to get meaningful impact data requires analyzing a specific system configuration. Until that’s done, I won’t know whether the processing impact is ridiculously huge or reasonably small.
I do think one constraint will be that a non-profit-run organization whose biggest expense is already server resources (my assumption) is very unlikely to add functionality with major operational costs. The $1.5m from Google will cover initial development costs (inc. some salaries I assume), but I imagine an ID tips system that increased operating costs even by 50% would be unacceptable. I also assume that environmental harm will correlate with operating costs (power, cooling, server hardware).
iNat staff has started belatedly to address concerns about environmental impact in points 3 and 4 of last night’s announcement update.
- iNaturalist has successfully incorporated machine learning and computer vision (kinds of AI) since 2017, with a very small footprint (primarily three machines running in spaces we control). We understand that new technologies can have much larger environmental impacts, and we aim to quantify the environmental footprint of iNaturalist’s infrastructure for 2025 and beyond.
- This grant funds our team exploring new ways to surface and organize helpful identification comments — and if the demo we create is not helpful, compromises data quality, has outsized environmental impacts, or is overall too flawed, we will not keep it.
I would like to see much more detail on this. (Maybe the first disbursements from the $1.5m can be for an outside auditor to publish a framework for how the environmental and ethical aspects of the project will be evaluated.)
But, back to the concerns you raised. Are you open to evaluating the environmental harm as the project progresses, or do you feel that the harm is guaranteed to outweigh any benefit? I would hate to see more people deleting their accounts, observations and identifications based on assumptions about unverified harms from a project that we currently know little about.
13 Likes
Brought up excellent points and articulated ones that I didn’t as clearly! You should paste this into the official feedback form iNat opened up if you haven’t already
4 Likes
Based on the example in the blog post, the way I’m envisioning it working is that the CV suggests a species, and the CV suggestion is accompanied by a brief comment like:
iNat community conversations indicate that Red-bellied Woodpecker has a pale brown underside, a red crown, and a finely barred back.
(This comment is AI generated based on community conversations, [click here] to report a mistake.)
I can see this potentially being helpful in working against relying on AI, because it would encourage users to compare the CV suggestion and the comment with their photos to see if they match.
One potential source of issues is that space is very tight in the app, and there’s not a lot of room on the CV suggestion page. It’s much harder to write an accurate description of an organism if you also need to be concise, as the subtle differences in word meanings become especially important (e.g. “barred” vs. “striped” might be used interchangeably in iNat comments and be understandable in context, but use of the wrong one out of context could be misleading). The more difficult the ID is, the more words and nuance you need to be accurate.
Another potential issue with this use case I’m envisioning is if the CV suggestion is incorrect (wrong species, or more precise than it should be). How does the accompanying text behave in that scenario? I guess it just summarizes any community comments on the taxon, which might mean that it actually contradicts the CV. Which is fine, assuming that there are enough comments to train on and that they agree with each other. With taxa where there are a lot of CV issues, there are often already hundreds of identical copy-pasted ID comments explaining the issue…
Also, the information that iNat comments include will almost universally be tailored to comparison with specific alternative species and not comprehensive descriptions. You would mention the barred back of a Red-bellied Woodpecker if you’re comparing it with a Red-headed Woodpecker, but you probably won’t if you’re comparing it with a Golden-fronted Woodpecker. I don’t know enough about different kinds of AI to know if something in an application like this would have the general knowledge and intelligence to infer the connections between different species as necessary here.
Personally I’m curious to see how this experiment turns out and I’m happy for my comments to be incorporated in training. But these seem like big challenges to overcome so it will take some creativity to find solutions, if they’re possible.
10 Likes
Right there with you. I think quite a few people are doing the same thing, and not bothering to comment on this subject yet.
6 Likes
I personally feel like the harm is guaranteed to outweigh any benefit at least that we currently know about or can foresee (and feel like almost all GenAI seems to be banking on this promise of some eventual use case above human capability/efficiency that isn’t evident in any of it’s output yet, but could be soon!). More specifically, I just don’t find basically any environmental harm justifiable unless there’s a benefit that’s actually clear and obvious, which in my opinion is simply not the case here; especially when options for human solutions to the same problems are readily apparent, and being enthusiastically volunteered.
I also would hate to delete all my contributions, I’ve put years and many memories into this community and value it highly. I know everyone else does too — you don’t get this type of impassioned response from apathetic users haha.
11 Likes
Speaking just for myself, I’m not that worried about the potential environmental impact of this particular implementation of GenAI on iNaturalist. I (so far) trust the staff to evaluate this thoroughly and responsibly. And as far as I know, yes, there are ways how to implement LLMs with relatively low energy requirements. That’s not the point for me. The point is the greenwashing and the (very cheap) positive publicity this would bring to Google’s AI projects and the entire GenAI industry… and I’m sure there can be no doubts about the horrendous environmental and societal impacts these have as a whole. Every organization with positive social capital that adopts their technology is helping their PR. They can — and will — use us in their ads and business pitches: “Look, even the nature nerds are into our AI. It can’t be bad! Why don’t you try it too in your company?!” It being some monstrous planet-burning, worker-oppressing, propaganda-spilling behemoth. I just don’t want to be associated with this industry and to help it grow. That’s why I wouldn’t continue contributing to iNaturalist if it prominently incorporated GenAI — even if, by some miracle, it was actually useful for our purposes (which it won’t because of all the well documented limitations of this technology).
13 Likes
This is not what LLMs do unfortunately. They simply output probabilistic chains of words as they have seen in all the data. They can’t discern nuances or context, because they don’t understand meaning. Honestly, they aren’t even good at searching. Here are some reasons why some of the solutions to minimize hallucinations mentioned above (such as retrieval augmented generation) aren’t enough:
- The summary extruded from the LLM is still synthetic text, and likely to contain errors both in the form of extra word sequences motivated by the pre-trainining data for the LLM rather than the input texts AND in the form of ommission. It’s difficult to detect when the summary you are relying on is actually missing critical information.
- Even if the set up includes the links to the retrieved documents, the presence of the summary discourages users from actually drilling down and reading them.
- This is still a framing that says: Your question has an answer, and the computer can give it to you. This framing brings all the attendant problems, as outlined above and in the papers cited.
Now, even if we ignore the technical limitations, what is the “workflow” below replacing? And what do you gain by making such change?
ETA: I only copied a small relevant portion above, but the entire article is pertinent to search/resurfacing ID comments, and is highly recommended: Information Literacy and Chatbots as Search.
8 Likes
Ok, I’m super late to the party. Would anyone be willing to summarize/synthesize the most important points, link to comments which have done so, or generally get me up to speed (300+ comments is way too much to sort through)? So far, I’ve read about environmental impacts, “greenwashing” of corporations, AI not being able to explain why it came to it’s decision (though, simply citing the observations that it used would be enough for me), wiki creation as an alternative, and bad PR. Did I miss anything?
Given the staff response, I’m certainly willing to give them some slack. I currently think this has potential to be a really good addition if implemented responsibly and with strong community input.
By the way and just for fun, I tried to get ChatGPT to summarize this and it did a terrible job. :-)
9 Likes
The most probable chain of words is context-dependent; in complex topics humans are more likely to be verbose and nuanced and as a result LLMs are more likely to do so as well. Unless you speak bluntly to the LLM, and then it might copy humans and speak bluntly and tactlessly back (depending on the prior instructions it’s been given). iNat community ID discussions will generally have the level of complexity appropriate to whatever the species in question are; the level will vary from basic woodpecker ID to more intricate insect features. The developer team has also has a lot of opportunity to fine-tune the prompt to encourage the type of results needed. But as I said, I think in agreement with you, I expect it will be a big challenge to do it well:
Again, I appreciate all the thoughtful comments here. I can share some of my own perspectives and also try to clear up a few things.
The communication issues were entirely our fault. I completely understand skepticism about Google, but in this case it was on us. We anticipated the announcement from Google, but we didn’t have our ducks in a row when it came to communication and I think we’ll be apologizing for that for some time to come. We’re reviewing what did and didn’t happen and will come up with a better process for communicating updates and changes. We don’t want this to happen again. All we can do at this point is apologize, try to learn from our mistakes, and try to earn back trust, which I know won’t be easy and will take time.
I think a lot of people, with good reason, think that this potential feature will be like the AI summaries that Google forces on people in its search results, or something similar. Personally I don’t use Google search and changed my DuckDuckGo preferences to never show me AI summaries, so this is not something I use. Nor do I use ChatGPT or its ilk, the AI summaries on the Forum, Apple’s suggestions for my texts, or AI-generated text (outside of typo fixes and such) when I write. So I am not sold on it as a solution for everything.
That being said, myself and I think many of us here on the forum and elsewhere do use AI/machine learning tools that are really helpful, like machine translation, reverse image search, iNat’s CV and geomodel, or transcribing speech to subtitles. So I think it has its place and is at least worthy of exploration.
For example, based on some explorations I’ve seen, one potential implementation could be a page of sourced comments for each taxon in the model, similar to what the geomodel does. Each month, like we do for CV and geomodel training, a model could be used to find comments for X taxon, weed out irrelevant ones like “Sweet bird!” or “Is this a wild plant?” and show you a page of potentially relevant comments with links to each one, and a voting feature that allows the community to weigh in on their helpfulness/accuracy.
The CV suggestions could then come with a link that shows you this page, and there could also be a link on the taxon page that shows them. That could be one way of generating a page of helpful, accessible, clearly sourced comments that the community could use and weigh in on. Personally I’d love something like that if it worked well.
I’m not an AI expert by any means, but my understanding when it comes to energy use is that enormous LLMs like ChatGPT or Gemini take a lot of energy to train and to use. They train on datasets that are immensely more vast than just the comments on iNat, requiring a lot more energy than an iNat model would need for training. Also, they use a lot of energy when generating responses to queries. For the scenario I outlined above, the model would not be generating new responses for user queries (there would be no text-based query feature like with ChatGPT). Like the geomodel, it would generate pages each month that would be a stable resource people would access for each taxon, cutting out the energy required to generate a new response for each use.
Whether that’s what we eventually build for this demo (and again, it would just a be a demonstration), I want to clearly say I don’t know and can’t make any promises. But I wanted to share an idea that kind of breaks us out of thinking about implementations we’re familiar with and don’t like, because there are many possible approaches. We want to keep humans in the loop, encourage learning, and not generate slop. Whatever does get demoed will be evaluated and if it’s not a helpful solution, it won’t be used.
This past weekend I had the privilege of going to The Cedars here in California and shooting some footage of the recently described cicada Okanagana monochroma, along with some of the people involved in its discovery. It was first posted on iNaturalist and caught the eye of some cicada researchers who collaborated with observers and locals to collect and describe the insect. These kinds of interactions and connections are at the heart of iNat, and that’s not something we would want to trample on.
65 Likes