What is this - iNaturalist and generative AI?

General
127

I don’t have anything clever to add. I believe this is my first post on the forum.

There is a significant difference between what I might call “generative ai”, which is just a cut and copy and cut and copy telephone game a computer plays with itself, and traditional machine learning. I’m ignorant on the details, but I am an artist, who will never post their work online again. I’m a writer, won’t be posting that online ever again.

I also work in a school, while attending college myself. I successfully introduced inat to two new users yesterday alone- one homeless guy I was chatting with near the high school I work at, and a fellow insect enthusiast and classmate at my college. I’ve gotten kids at my high school to look closer at the “eeeeeee a bug!!!” Rather than just running away, I’ve gotten to tell them about all the different trees on the property, and encouraged them to learn with the tools they have.

All that to say, inat is successful, I believe, because of the human connection. I am wildly ignorant about much, but I know I can trust certain users for certain ids. When a spider guy can’t confirm an id, I trust him. When the aphid lady corrects my thoughts, I trust her.

I will never trust generative ai, and it will always be bile in my throat. Each time I have to leave my home and see another gods damned data center that took over a field of milkweed I’d been excitedly watching, not even desperately needed apartments, another damned data center, my heart sinks low, and I weep. Each time someone tells me to my face students don’t need to learn art or writing, chatgpt or whatever other nonsense can do it for them, my tongue shrivels into my heart so I do not poison them with my words.

The more I see so-called-generative-ai (we do all know that is a misnomer), forced into everything online, and more and more frequently in person- can’t even go through a drive through to grab coffee without running into it, the more I remove myself from online spaces I have existed in for the better part of 30 years.

Dead internet theory has come wildly undead, and its hydra heads snap at the heels of us all.

(And yes, I did write this myself. I wrote it from my heart, yes ignorance and all, and nothing more than a love for the world around all of us- the real, tangible one, not the fake, immaterial, hollow one Google and other big tech hydras want us to live in.)

32 Likes
128

Jesus Christ.

As others have said, creating some sort of network or tool for assisting with identification would be great. Unfortunately, in the world of tarantulas, a lot of resources for identification are not accessible to the average person and you just kind of have to figure it out by looking at other people’s comments. A forum specifically where people can post under species/genus pages about ID’ing would be an invaluable tool for both new and experienced users.

However- I wholeheartedly agree with others here. Generative AI is one of the most contemptuous developments of the modern era and INaturalist in particular ought to be ashamed should it seriously be considering associating itself with it. Especially given the environmental impacts of using this technology. I commented a while ago on a post about INat CV and the possible environmental implications. I remember saying that the value of INat as a database I believe warranted the small extra energy toll of CV. However, any kind of generative AI tool does not fall under this same exception.

I pray that INat will listen to its users on this. No generative AI on this site. Not now, not ever.

15 Likes
129

OK y’all. Once again, I see that everyone who is opposed to this or threatening to leave is missing key context and making incorrect assumptions.

  1. the LLM hallucination problem can be greatly mitigated through many techniques such as Retrieval Augmented Generation and a specific training scope.

  2. there is no reason to believe iNaturalist’s LLM would add incorrect ID information to species that cannot be IDed. If anything, a system that drew on comments from identifiers would be able to caution users that “these species cannot be told apart without dissection” or something similar.@cliygh-and-mia edited to tag you as items 2 and 4 on this list should give you peace of mind about the concerns you just referenced in your thoughtful post.

  3. your comments are already being used by Google and half a dozen other LLMs. You are threatening to leave because iNaturalist is taking what is already happening and turning it into a hugely beneficial tool.

  4. LLMs are more environmentally friendly than using an army of humans to write a wiki, objectively. https://www.nature.com/articles/s41598-024-76682-6. Sure, the exact situation described by this paper is slightly different than what iNat is doing, but fundamentally it’s the same—writing text. LLMs are upwards of 4400x less environmentally harmful than having a single human do the same task. Imagine how much less environmentally harmful they are than tens of thousands of humans writing descriptions. Don’t trust the media, trust peer reviewed science when it comes to how harmful (or not) LLMs are to the environment.

  5. it is absolutely fair to point out that Wikipedia exists and we haven’t filled that out yet. We shouldn’t expect iNaturalist to create an entire easy-to-edit wiki system because we (myself included) are too lazy to learn how to use what already exists. Also, it’s super fair to point out that if you’re complaining that they’re taking money from Google and have the capacity to donate but don’t… what can you expect?

  6. I know some people have tried to push back against this, but it is the genetic fallacy to say this is from Google, therefore it is bad. And if the concern is that Google mistreats its partners, keep in mind that iNaturalist and Google have been partnering well for more than a decade.

I understand that many of you have ethical concerns, but I’m concerned that you are leveraging those concerns to force iNaturalist to do something counterproductive. If your concern is that you don’t want anything you write to be used in an LLM, what iNaturalist is proposing will change nothing. iNaturalist is taking an arguably bad thing that is already happening and turning it into a good thing. Those who think it’s impossible for this to be good, or who have never seen a good GenAI product, have not researched the topic much. I really hope iNaturalist does not bow to the mob of uninformed people threatening to leave.

11 Likes
130

In August, my iNaturalist account will be 10 years old, and I’ve been around for many of the site’s changes through the years. Even though I can see that many commenters here share my views on the matter, I still want to leave my two cents for what it’s worth.

Personally, I am against the use of generative AI such as described above on iNaturalist. In short, despite many benefits it could have for the platform (the speed organisms could be identified would be useful for people not particularly invested in wildlife, but that begs the question why they’d be on iNat in the first place), it exacerbates and creates new problems for identifiers for no real reason other than “AI” being the latest new and catchy buzzword for investors such as Google to profit from in the short-term.

Many others above have commented on the environmental and ethical considerations that should be addressed before something like this is added to iNat, so I won’t restate what they have said. Instead, I want to further comment on a point brought up by a few others (e.g. @Masebrock), namely that there is an intrinsic difference between how the AI used for computer vision is utilized compared to the proposed model. CV is a tool that gives suggestions which can be overridden by either the observer (if you know already know the species you can enter its name when submitting observations) or by identifiers once they pass the 2/3rds benchmark. This system already has issues, namely in the often large amounts of time and effort it takes for multiple identifiers to review observations of particularly challenging groups. While some discussion about what this means for arthropod, plant, and fungi identifiers has occurred, this would also negatively impact identifiers for groups such as mammals and birds. For example, many of the smaller bats (e.g. Pipistrellus) can not be identified from pictures without audio verification of their vocalizations, which has not stopped CV from listing possible species from images alone. I primarily identify birds of prey, and I often have to correct CV identifications from (usually new) users even in well-iNatted regions such as in the eastern U.S. (the CV loves to identify Red-shouldered hawk as Cooper’s hawk and vice versa, not even mentioning Sharp-shinned vs Cooper’s hawk IDs). Gen AI descriptions would add more confidence to incorrect identifications that newer users would accept at face value. This would increase the amount of misidentifications on the site (and where iNat data is shared including GBIF, academic papers, and now Google) and the amount of time and effort needed to correct these misidentification (accounting for individual observations and maintaining the AI) for seemingly no real reason other than just to capitalize on the AI trend.

If iNat is now contractually obligated to follow through with adding this, I would strongly suggest having it as either an opt-in feature as mentioned above, or incorporating it into a separate app such as Seek. This would at least reduce the amount of incidental damage caused by the tool while still having it as an option for those who would benefit from it. Additionally, not only would I support the creation of a community wiki, but I would also volunteer time to contribute to ensure its information is up-to-date and correct. I hope that the iNat staff has considered this before the announcement, but I do agree that we should still wait for more information about how this system will work before deciding where to go next from here.

19 Likes
131

To quote from the blog post:

Our nonprofit mission is to connect people to nature through technology and advance science and conservation. [my emphasis]

This seems surreptitiously different from what is presented on the about page:

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.

I sincerely hope this is just an unfortunate choice of words and doesn’t represent a genuine policy shift. The notion of interposing Articifial Interfaces between people and nature seems diametrically opposed to the current vision/mission statements.

27 Likes
132

It isn’t fair, really, because Wikipedia doesn’t accept ID tips in their taxon articles. Adding such content to relevant Wikipedia articles results in the content being removed. Part of the issue is that a lot of the user-created identification tips we want to record somewhere have never been published anywhere.

https://forum.inaturalist.org/t/adding-useful-id-info-to-taxon-pages/657/8

26 Likes
133

This is the third time you’ve posted the “genetic fallacy” argument in the thread, so I guess it’s time someone cleared this up:

It’s not a genetic fallacy to say that something that comes from a company that you think is bad is likely to be bad. “Genetic fallacy” refers to refusing to address the argument itself. But Google isn’t presenting an argument, it’s saying it wants to do something.

Here’s an example of a genetic fallacy:
“Bob says he has data showing lions are faster than tigers, but he also likes pineapple on pizza, so what does he know”

Here’s an example of not a genetic fallacy:
“Bob says he can be trusted to hold the funds, but didn’t he get caught blowing all the money on pizza the last time we gave it to him?”

So no, not a “genetic fallacy”. Just normal, good old-fashioned distrust of entities that have proven themselves to be untrustworthy.

As for:

Google’s own AI hallucinates. And the iNat partnership is with Google. Whatever “mitigation” you’re suggesting exists is apparently far from sufficient. Instead of saying “this time it will be different”, first develop an AI that has solved this problem, then get back to me.

22 Likes
134

“This mushroom can be found at the tops of trees” -AI, about a ground mushroom.
Not even a joke.

11 Likes
135

It absolutely will, because people on iNaturalist do, and make comments laying out their (flawed) reasoning on iNaturalist. The LLM is going to get trained on those comments too, and LLMs are probabilistic: they will sometimes put out one line of argument, and sometimes another, based on the frequency with which they see those lines of argument in the context they’re in. So far as I can tell, they do not decide for themselves whether any line of argument is correct or incorrect, but simply repeat the same lines of reasoning they’ve seen in their training data, sometimes in their own words. The only way (so far as I’m aware) to elicit correct lines of reasoning is to alter the context in which they’re giving the answer (the prompt), which is far from error-proof.

Note: I love that you’re working to counter the flood of fallacious arguments, @natev. Great work! Trying to steelman some arguments myself. If I thought they were going to try to automatically produce those side-by-side photos with little arrows pointing out the differences between the species, or if I thought they were going to try to make the image classifier explain its reasoning in its own words, I’d be for it. But the image classifier and the human identifiers are clearly using different reasoning for their IDs in many cases, and the human identifiers are quite diverse on their own.

19 Likes
136

Hey! @natev, it seems like you might be slightly disregarding people’s feelings here — not good! Saying “everyone is making incorrect assumptions” and “not researched the topic much” can be seen as a bit presumptuous.

It is impossible to convince anyone in a forum, but here are some personal thoughts:

  1. You are correct, LLMs can eventually be near perfect. Also, that may not matter, because they will always be cold and heartless.

  2. You are once more correct, LLMs and the id tips can eventually be near perfect. Also, that again may not matter, because they will always be cold and heartless.

  3. Here, you may be missing the point. People are threatening to leave because iNat is getting closer to something they utterly despise. You don’t despise big tech GenAI? Fine! But many people who love iNat clearly do, so let’s consider that.

  4. Math may be right, but assumptions are questionable. LLMs are optional, humans are not – that is important to remember when thinking about “costs”. Also, humans doing something as purposeful as iNat get motivated, which changes our worldview, which makes us care more about the environment.

  5. Oh, please, come to Wiki. I have created 1000s of articles for species on Wikipedia — we are not that lazy and it is not that hard. Also, lots of people there, doing human things. You are all invited to drop by, edit things here and there, it is fun. But as @JeremyHussell added, Wikipedia is one thing, a Wiki for ID tips is another. One sure thing is that it is hard to say in advance what will work or not. As a friend says, “the thing about Wikipedia is that it only works in practice. In theory, it can never work.” But maybe it is worth trying.

  6. @Masebrock answered way better than I could :grinning_face:

We are all smart, reasonably informed people that care enough about iNat to spend time in a forum. So framing this movement as a “mob of uninformed people threatening to leave” is also not good. It is more likely a collective that is so distressed that they feel like leaving the very thing that gives them purpose.

What I am sure the iNaturalist team is likely to do, as they clearly care about the community and the mission, is to seriously take into consideration the feelings poured here.

It is surfacing important points where long discussions are needed, both directly to the details of the ID tips and this particular Google project, but in particular to the implications in terms of ethos, image, mission alignment, community feelings, and so on for future partnerships/features.

29 Likes
137

Never have I seen a topic generate so many responses by first-time posters, and seldom have I seen one arouse such passionate responses! I am trying to remain dispassionate myself, but only wish to add:
I hope that the users who express their intention of leaving iNat and deleting their accounts, choose instead to simply stop participating, without deleting.
If you have been making good observations and correct IDs, and contributing sound, well-reasoned and accurate comments, isn’t that exactly what we most need to retain? If you are concerned about AI amplifying or generating bad or incorrect data, please don’t erase all your good data. It’s the good stuff that we want the AI to be “learning” on.

23 Likes
138

I will split my comment into two, the first part being about practical concerns, the second about ethical ones and data-ownership and stuff.

I: Practicality:
While I think that - if done right - the planned addition of GenAI can ultimately lead to more informed identifications, and over all less misIDs, I feel there is also a lot of potential for disaster.

  1. Many comments by identifiers (at least mine often are) are very specific to the observation they are made under. Sometimes, a comment that helps for one observation would cause a misID on another for various reasons. Sometimes, the criteria described in a comment can be used to distinguish a species in one place, but not in another.
    ↳ I think, if at all, the GenAI syntheses should only be made on taxa with an atlas. Location of the observation should always “be taken into consideration” by the AI.
    ↳ I think the input/learning data has to be open and user-curated with Curators/IDers/everyone given the right to remove certain elements that may be problematic
    ↳ any GenAI synthesis should be able to be edited or deleted by IDers of the taxon, which should then automatically withdraw any ID on that observation made with CV

  2. The syntheses given by GenAI will make any CV-suggestion appear more confident. However, as the addition of an explanation will not actually change anything about the confidence or the accuracy of the CV-suggestions, these need to be made more cautiously.
    ↳ That means, the top-suggestion should never be more precise than Genus. The relatively recent change that a top-suggestion can be a species should be reverted (it should be in any case)

II: Ethics and stuff:
As someone who paints, writes, and makes music, I am not a fan of the way GenAI is currently being used (to put it mildly) and even less so of the big-tech companies behind the LLMs. As a lot of others have already raised the ethical concerns they (and I too) have regarding these, I think there’s no need for me to repeat them here.
I have three demands here, which I think are the base-line for fairness:

  1. iNat shall not share data with Google or any other third party that the user (and owner/creator) hasn’t set to Open-Source (or CC-BY if iNat can guarantee the owner/creator be credited properly). iNat should especially not share more sensitive user data (like activity or locale) with Google or any other third party
    (Personally, this is the one point that would make me actually consider leaving iNat)
  2. Users’ comments being used as training data for the GenAI should be an opt-in feature. If someone has opted in, they can opt out at any time, withdraw their existing comments from the training set, and opt out for any individual comment
  3. Users should have the option to be notified if their comments are being used as a source in any AI-synthesis and/or an easy and convenient way to check AI-syntheses based on their comments. (I wouldn’t want my (user-)name to be associated with a bad or irrelevant synthesis that may misinform others)
11 Likes
139

Your comment about the image classifier gets at something else I’ve wondering about. The CV model is clearly capable of classifying multiple gestalts in the same category, because it is generally able to handle species with very different morphs, life stages, etc. And if you do the Visually Similar option in Compare, it generally is pretty good at pulling a similar-looking taxon photo that is in the same subcategory. (If you do a female mallard, you get a female mallard photo as the most similar taxon photo, etc.) But it doesn’t sound like the model currently defines those categories or explicitly classifies photos into them. Which seems like a big problem for AI-generated ID tips, because those require you to know what subcategory a particular observation is in. This is something people do almost automatically, but as far as I know, the CV can’t do this at all right now.

It would be very bad (and very silly) if I put up a female mallard and got an ID tip that started “this is a mallard because of its distinctive green head”! Even different views of the same subcategory of a species have different characters. It would also be pretty silly if I posted a ventral view of a frog and got the dorso-lateral ID character that was used as an example in the blog post.

13 Likes
140

Replying to this specific point that you keep bringing up, others here have already put into words the same general objections about this topic I also have.

Now to this argument, the comparison they set up in this study is completely moot for the real world - right now generative AIs are not being used to replace humans doing things like writing good and meaningful (!) text or create meaningful and valuable art, it is used to create additional text and additional images, which add on to the total energy consumption. (Or why else, as somebody in the thread brought up earlier, are the projected future energy costs for generative AI absolutely massive). Even if iNat implements the LLM like you describe, it would still always have to be verified and very likely modified by at least one real human expert, and with the caveats they list in the paper it’s very unlikely that the average energy per species (or wiki page, or whatever else) will be lower when an LLM is involved. Or in other words, the LLM uses less energy but can’t actually really perform most talks that a human could perform (as others have said, hallucinations are rampant everywhere and the whole internet is at this point already flowing over with AI-generated nonsense).

13 Likes
141

From the blog post:

How will you ensure that the identification tips are reliable?

We will incorporate a feedback process for the AI-generated identification tips so that we can maintain high standards of accuracy. Since this project is in its early stages, we don’t know exactly what this will look like, but we will share updates to be more transparent moving forward. [my emphasis]

This looks like a gigantic hole in the proposal and suggests the announcement may be premature. The CV depends entirely on constant feedback from human identifiers. Without that human-powered verification process, there would never be any reason to trust the CV suggestions.

Will the iNaturalist community be willing and able to verify these “explanations” produced by generative technologies, given that they’re already struggling to keep up with the identifications? Where is all this extra volunteer effort supposed to come from?

Again, from the blog post:

By providing explanations in addition to a list of suggestions, iNaturalist hopes to more effectively grow a skilled community of naturalists who have the information and tools to improve and vet the data on iNaturalist.

So is the expectation is that these extra volunteers will somehow just emerge by spontaneous generation? This seems completely backwards.

Surely it’s firstly necessary to create a pool of sufficiently skilled technical writers and science educators - i.e. a group of dedicated people who know the literaure well enough to summarise often highly specialised information in a way that can be understood by everyone. There’s a big difference between asking someone what X is, and asking them to explain how they know it. Identification leverages inate human visual abilities which are almost entirely subconscious. By contrast, teaching is an entirely acquired skill, and the people that are good at it often aren’t experts in their subject matter. With minimal practice, almost anyone can learn to find matching images without any prior knowledge. The same can’t said when it comes to finding good explanations.

21 Likes
142

Good suggestion! I am very strongly against the use of generative AI in nearly all cases. Others thus far have discussed the ethical concerns of copyright infringement and the practicality issues of hallucination and errors, but I stand against it on environmental grounds, which I haven’t seen discussed as much. Admittedly, I haven’t had time to read through all the comments, so maybe it’s discussed more further down.

I understand that having quick access to ID tips for any taxon is a often-requested feature, and I would love to see it implemented myself. But if it is implemented through generative AI, I honestly may have to consider quitting iNaturalist to keep in line with my personal morals. I really hope that iNaturalist decides to implement this feature in a different way, such as through community curation.

If we can do it with taxon photos, surely we can do with ID tips. Another suggestion I’ve seen on a separate forum post was to have an option to mark a comment as helpful (in terms of ID help), similar to how you can mark an observation as favorite. Although I think this might be done better as a voting system, such that users can refute misleading information or cancel out accidental votes.

5 Likes
143

my hope would be that eventually the AI would be able to point those who need guidance to reference materials that community members have noted in their identification comments as helpful for identification.

beyond offering guidance like “species X has feature A and B and C”, i hope that the AI will be able offer something to the effect of “here are observations 1, 2, and 3 where folks have discussed how to identify species X” and “community members have used references L and M to guide their identification”.

as a person not formally trained in any natural sciences, i can speak from my own experience learning plants, that i started off learning to identify them as i think most folks would – by having others tell me what a particular plant was and trying to memorize an image of a plant with its name.

but luckily at some point, i ran into someone who told me that it’s better to identify plants using dichotomous keys, and she referred me to a good refence to use. through practice using various keys, i think i became an above average plant identifier (at least for plants in my area), but i never would have gotten there if nobody had told me that plant keys exist, and although i’m not a super prolific identifier, i would not have been an identifier (for other folks’ observations) at all if i never had learned that one thing.

so my point is that it doesn’t necessarily take much to empower someone to become a better identifier. as they say, “teach a person how to fish”. and while it would be nice if the lesson comes directly from a human teacher, it doesn’t have to, and there’s nothing wrong with using all the tools available, within reason, to spread knowledge.

10 Likes
144

In my experience there are specialist experts who are able to say whether a species can be identified from a photograph or not and if it can what the key features needed to identify it. AI cannot provide expertise all it can do is summarise comments both idiotic and brilliant and everything in between.

1 Like
145

I think there is a bit of a kneejerk reaction (more on the announcement itself than this post), but Gen AI is such a contentious topic I think iNat needed to pre-empt that a bit more and give some more context from the off.

If it’s a LLM trained solely on data within iNat, that doesn’t leave iNat, then it’s less problematic and could be useful (though whether useful enough to justify the price tag? Not so sure…).

If there’s any plans for data going elsewhere, then obviously the community has a right to know as soon as possible.

If it’s drawing identification data from open sources, the fears about accuracy and hallucination are well-founded.

Personally I’d rather have seen a human-driven project utilising this site’s amazing community to create a world-leading resource of identification knowledge, with AI utilised to summarise that to users, than an AI-driven project with users checking its work, which is what this sounds like. :person_shrugging:

15 Likes
146

Thanks everyone for the responses. I think at this point, due to the 30-minute delay, the most effective way for me to respond is to create a big comment and just tag people at relevant sections. If you do not like me tagging you, please message me and I will remove you from this post and not tag you again.

@JeremyHussell thanks for the comment about Wikipedia, it is a good point that Wikipedia won’t allow unsourced ID tips. @tiagolubiana, I did not mean to come across as rude calling myself and most of the other complainers lazy for not editing Wikipedia. Definitely appreciate the work you’ve done on Wikipedia.

@tiagolubiana I apologize, I did not mean to disregard people’s feelings. However, I do think it is factually correct that the dissenters are making incorrect assumptions. I think that is demonstrated by your willingness to concede points 1 and 2 (though I must add, your stressing of “eventually” shows you are not familiar with the current techniques that can be used to great effect on targeted applications such as iNaturalist’s project). Whether they are cold and heartless is a subjective matter and I do not believe it is fair to use those terms pejoratively at the AI, especially considering how cold and heartless the primary literature often is.

I actually anticipated your counterpoint on point 3 in my final paragraph (“I understand that many of you have ethical concerns…”), so I do not believe I missed the point.

On point 4, the assumptions are not questionable because while humans are not optional, “humans performing this task” is just as optional as using an LLM. I agree that humans get motivated by working on iNat, and I would be a lot more motivated to provide ID comments if this feature is implemented.

@Masebrock, your example of a genetic fallacy is actually an ad hominem fallacy, which I think may speak for itself (and @tiagolubiana may want to reevaluate referencing it to debunk my 6th point).

In contrast, I have absolutely applied the genetic fallacy correctly. We must judge things on their own merits. We cannot dismiss a good thing because it comes from a bad company without committing a genetic fallacy (error in reasoning caused by evaluating the source of the thing and concluding that the thing cannot be good). Plus, you are overlooking the extremely salient point I already made–iNaturalist has collaborated with Google for a very long time without any problems, so it is not reasonable to say Google has “proven itself untrustworthy” in the context of iNaturalist. If anything, Google has proven itself to be a valuable partner.

Google’s own AI hallucinates. And the iNat partnership is with Google. Whatever “mitigation” you’re suggesting exists is apparently far from sufficient. Instead of saying “this time it will be different”, first develop an AI that has solved this problem, then get back to me.

Methods of dramatically decreasing hallucination have been public for years. See A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models for an analysis of more than 30 methods to reduce hallucinations. It’s harder to implement these in a general generative AI model like Google’s public-facing model, but when you tailor Google’s LLM to a specific domain and then use one or more easy-to-implement mitigation techniques, this is really not a problem at all.

The only problem that would remain is the one @JeremyHussell points out:

It absolutely will [add incorrect ID information], because people on iNaturalist do, and make comments laying out their (flawed) reasoning on iNaturalist.

This could indeed be a problem, but it can be mitigated if users can provide feedback to the AI outputs (as the blog post indicates). However,

LLMs are probabilistic: they will sometimes put out one line of argument, and sometimes another, based on the frequency with which they see those lines of argument in the context they’re in.

I think this would work to our advantage. I have yet to see an example taxon where anywhere near the majority of ID comments were incorrect. I suspect that the correct or useful ID comments outweigh the incorrect ones by a substantial margin. Once again, any errors could be corrected by user input.

@calvertm, the environmental concerns have been addressed. In the US, LLMs are between 40 and 4,400 x more environmentally friendly than human labor. Reconciling the contrasting narratives on the environmental impact of large language models | Scientific Reports. There are some caveats there, but I have not seen any convincing reason to believe iNaturalist’s usecase would be less environmentally friendly. I hope this can help keep you on iNaturalist–looks like we will need more help in the coming days.

@andreass1 thanks for the response. I agree that AI is not making valuable art. I think you are drawing an equivalence between the general-use generative AI models and what iNaturalist is suggesting. Your point that these models are creating additional information is moot in this scenario, where the model creates information that otherwise would be created by a human. And it takes far fewer resources to correct a sentence or two of LLM output than to write everything from scratch. As I have noted many times now, there are ways to mitigate hallucinations, especially in a domain-specific application like what iNaturalist is proposing. Tailored LLMs with guiderails do not produce AI slop.

For instance, one LLM used in my field is PQAI, a free and open-source prior art searcher. (Edit suggested by @spidercat: prior art just means previous inventions, either in patents or other literature, not traditional art). It uses an LLM to take a natural language query, perform a prior art search, and pull up relevant papers and patents along with AI summaries. It has proven to be far better at finding relevant prior art than traditional searching resources that cost thousands of dollars, democratizing access to essential information for innovators. I could go on with many other examples of high-power, high-accuracy, low-hallucination applications that do not produce AI slop. iNaturalist should have no problem creating a solution that works.

With all due respect, I still do not see evidence that the people threatening to leave understand the underlying technology. I empathize with those who plan to stay but still have reservations. While I appreciate all of you, have learned from each of you, and do not mean to insult you by this, I also feel tired of responding to the handful of arguments that miss what I’m saying and get basic facts wrong. It also is demoralizing for my well-researched, factually accurate comments to get ratioed by factually inaccurate responses that fail to address my points. At this point, I recognize that this is not worth my effort, and I will let iNaturalist handle their own PR and use my free time to add identifications to make up for the upcoming exodus…

8 Likes