I would see the new tool as useful for these obs. If we can free up taxon specialists to work on the interesting stuff instead.
Pattern-recognition software pointing to the wrong answer is not the same thing as AI “hallucination”, and you shouldn’t conflate the two. The CV saying “based on image pattern recognition, this photo most closely matches photos from this species” is not even a false statement, regardless of whether the species it points to is actually right.
And perhaps more critically, you’re not learning anything incorrect from the CV (or at least you shouldn’t be: iNat explicitly says not the use the CV as your sole basis for identifications). Contrast this with a genAI saying “this is Pinus strobus, because it has needles in bundles of three” ( hallucinated nonsense), that spreads into a cesspool of misinformation. This is why genAI is more dangerous than the CV: It makes actually false statements.
You want to learn ID characters? Great. GenAI is one the worst ways imaginable to do this. Because you’ll be learning false characters while thinking you’re reading correct ones. Now, instead of being curious and cautious, you’ll be confidently wrong.
Comments about identifications are usually under others’ observation. I don’t think there is an option to choose their licencing.
By the way, I would prefer not to feed AI with my work which it uses to replace my role in some way (why wait for ID from real user when AI makes it faster? and why listen to user saying the species ID can’t be made when AI says it can?) but I’m afraid the only option for that would be deleting the account altogether
I’d suggest that incorporating existing discussions and mixing that with natural language explanations of what can be seen would be a logical way for them to do this. There’s no reason to assume it would be done wrongly. Image Recognition already has massive limitations and we’re always finding errors in the suggested IDs but that hasn’t prevented people from using it and joining in. AI is no different in this respect and we should give them a chance to implement what they are thinking of doing - then comment.
I must be honest even if it hurts. My optimism is wanning. Im sure staff are in a difficult position, but they also got into this position themselves in some aspects if we are honest. One thing being, much more details should have been on that initial blog post. I can only hope they recover from this tenuous position as best as possible.
Regardless of whatever you think of the project yourself, you cant ignore the communities feelings. I can still say i personally dont want this in its current form.
Many parts of the community still oppose this, even vehemently. So staff really need to be a salesperson and sell us why this is a good idea / needed. They need to address our concerns. Even if you are pro this plan, if some people arent, sell the idea to them. Explain why this is needed and a good idea.
My optimism is wanning becuase i think this is going to go through even when seemingly a huge chunk of the userbase is very opposed and some parts even going nuclear.
This is a concerning sitiuation becuase it seems that no amount of community disagreement can create a course change. Becuase i dont know how you would realistically create a situation with more disagreement and outrage from the community.
I want to end this with. Even if i strongly agreed with the plan. I vehemently oppose pushing a change that so much of the community doesnt want. Even if i think that change is a positive. The communities voice should matter. If so many people are speaking out, theres cleary an issue and implementation of the plan should at a minimum be paused for review and community involvement.
I’m inclined to give iNat staff the benefit of the doubt here. Declining the grant might have consequences we’re not aware of. Running an organization like this is complicated, growing pains are inevitable, and overreach as a consequence of enthusiasm for growth is extremely common. I take them at their word that if this initiative craps out, it will be abandoned. And if we’re right about this tech, then no problem… it’s gonna crap out. If they fail to abandon it after a trial period, THEN we have problems. But for now, I’m inclined to see this as a forgivable screw-up. We’ve all screwed up, we’re just not all in a position to do it so publicly. Thank goodness. 
You can dramatically reduce hallucinations by training the LLM on a specific domain (e.g., iNaturalist comments) and then forcing it to use retrieval augmented generation. Basically, you use two models: the first model is more trained and is used to locate a particular comment or comments that are relevant to the prompt, the second model is less trained (and has less background knowledge to hallucinate about) and basically just repeats the source comment. Very easy to make it include a link to the relevant comment as well for verification. What almost everyone is missing is that the general generative LLMs that anyone can use online as a chatbot are not what iNaturalist is proposing to use. There are many of the use-case-specific LLMs such as PQAI (which I use very regularly for patent searches) that essentially never hallucinate.
I agree that deleting accounts is 100% the right of the user. I would just recommend people wait and see before making such a drastic choice that would require an inordinate amount of effort to reverse. As for predictive vs generative AI, they are certainly different, but I don’t think the difference is as great as people make it out to be. When iNat is talking about making their own model, I’m assuming it won’t be a generalist LLM like ChatGPT, but one much more specific and tuned to wildlife identification. I think this would limit the size (i.e. environmental impact) and hallucination potential, but again we have to wait and see. To that point, I don’t think the iNat team would roll out a feature if, after testing, it churned out misinformation.
Edit: welcome to the forum!
Predictive AI is very, very different from generative AI. iNaturalist staff themselves have in the past distanced themselves from using the term AI to refer to their CV model, insisting instead that users refer to it as CV, not AI.
Predictive AI is not attempting to create anything new. It does not try to transform users’ contributions into something new without their consent, which has many ethical and copyright implications. It does use users’ photos, but only to try to classify them- it does not create new photos.
Predictive AI is not prone to hallucinations in the same way as generative AI. The CV is very often wrong (or, we interpret it as wrong), but it does not make up information, nor does it make confident assertions about that information. It provides suggestions which are to be taken with a grain of salt. Generative AI will make up information, present it confidently, and often even defend the information, even providing sources that are often themselves inaccurate, made-up, or which don’t actually support the assertion. I believe this could even discredit users- if a model confidently argues that ‘this is species A because user B said it has characteristic C’ and user B never actually said that, others may still believe user B to be making false statements about ID characteristics.
Most importantly to me personally, generative AI is way more energy and data intensive and has a much higher carbon footprint than predictive AI. There have been many forum posts inquiring about the carbon cost of iNaturalist in which the environmental differences between predictive and generative AI have been outlined. This is important to me because I’m seeing deforestation to build data centers, and coal and nuclear power plants coming back online or being built solely to satisfy the energy demand of generative AI. I do not wish for my love of nature to contribute to the destruction of nature through the use of generative AI on this website.
It is very concerning to see the upset that has been caused, and I don’t doubt that the staff will be taking it all on board as they consider their next steps. Having said that I find myself really at a loss to understand it - my incomprehension doesn’t invalidate the feeling of others, of course - but I want to say that I, for one, do not quite understand the level of concern. I understand that GenAI has caused problems in many of its use-cases and there is a general opposition to it based on that, which makes sense. Some people find Google itself objectionable. OK, for those people I understand that aspect. But in terms of the actual proposition…
When I put a comment on iNat, I intend it to benefit the community. In fact, it only benefits those who look at that observation - very very few people. Why would I object to a system that enables the whole community to benefit from my comment? In fact, I think I would probably comment more if I thought the value of my comments were being amplified in this way. Even though there is a generative element, I don’t see the difference in principle between this being trained on the comments I freely contribute to iNat, and the CV being trained on the photos I freely contribute to iNat. I contribute these things to iNat freely, precisely to advance the purposes of the site - so that people can learn to appreciate nature in a new way while generating useful data. I want people to have them.
Now of course this could also be performed by the sort of wiki proposed by some above. If that existed I would certainly contribute to it: Anyone who has seen my journal knows that I would! But here’s the thing: that would be a lot of extra work for me, and the time I spend doing it could be spent identifying. This proposal basically just makes what I am already doing more available to the community I am doing it for.
I agree that an opt-in system might be useful. ‘Would you like your comments to contribute to iNat’s auto-tips function?’ (with a bit of info). Then those who are more confident in their knowledge can say ‘yup, I’d love to help with that’, and others not. This might help the system not to be trained on comments that are either unknowledgeable or speculative, and other people can relax a bit more not worried about if their comments might end up misleading the system, and just get on with using the site.
Rather than leading to an increase in misidentifications, I suspect it would rather do the opposite. Two sources of information (the CV and this new system) would both have to independently mislead. CV says it’s likely to be Eristalis pertinax ‘oh yeah, that looks right’, says the user: ‘let’s check the AI tip’: “Commenters say this species is identified by its yellow front feet” says the auto-tip. ‘Oh, mine has black front feet… maybe I should stay at genus…’
Ultimately these computing techniques are coming and, like all new technologies, to what extent these things are for good or ill depends on the uses to which people put them. Harvesting copyrighted artwork which then competes with the businesses of those same artists and puts them out of business is obviously terrible. But it is for good people to use them in good ways. This seems to me to be a limited, sensible, useful, and generally good proposal, from good people.
Except that there will literally not be enough human experts to check and verify every observation soon, when automated data collection starts flowing in. We need AI to learn and get better to take away the easy stuff and leave experts with the really interesting or difficult IDs to make … that’s good use of AI and it will enable big data to flow into scientific research.
What is not clear is why a LLM and GenAI are the solutions to this issue. There is currently no way for users to easily search IDs with comments. Even a link on a taxon page that brings you to a search of observations containing IDs with added text would address this issue. If these searches could be further refined by key terms, users, additional taxa, et cetera, they would be very useful.
GenAI is resource intensive in computing but also of users’ time in maintaining its outputs. @hawkparty made a post addressing scalability issues as well- these issues will fall on users to volunteer more time to maintain a LLM and its outputs rather than identify or write helpful journal posts/wikis of their own.
I just do not understand how this project is a solution any better than user-curated wiki, improvements to search functionality, and improvements to journal tagging and searchability.
I do not value whatever output GenAI puts out, I care about the community and literary body that it sourced from. The synthesized material will never have more value than its inputs. I want to access the inputs myself, not have it interpreted and spat out as an output that I must put extra work in contending with.
The blog post does not state clearly (to my eyes) what the GenAI is supposed to feed on. What will be the input(s)?
It states a problem to solve: “how to surface the most useful identification tips shared by these members of the community”.
I take it as: they would like to use as input, only those ID tips shared by members (…shared where? in personal journals? as comments left with their IDs? anything else?).
Then: “generative AI could provide a scalable way to synthesize and share useful information about how to identify the 100,000 different species included in our current modeling process”. (modeling process = CV, more or less? I suppose.)
I can interpret this last sentence in two ways:
- let’s forget about using members’ tips as input; to produce ID tips, we could use imagery (and only that) as input, asking some GenAI to write a human-friendly explanation of what the CV has been doing internally - peeking inside the black box, telling what it sees.
- let’s use GenAI to process members’ tips as input; then we’ll reattach the GenAI outputs (“ID tips, synthesized”) to the CV outputs (“suggested taxa”).
What would it be?
I have some additional thoughts.
Google has cleverly set this up so that it can’t lose.
They knew – I am 100% certain – that there would be drastic, radioactive fallout from a vague, third-party, buzzword-ey announcement. They chose not to warn the iNat team, nor prepare them. It doesn’t even look like they coordinated with the iNat team on when or how to make the announcement.
If reception had been neutral (even positive), Google could take the credit and add to the reputation of genAI – “even iNat supports it!” But when reception is bad, iNat takes ALL the blame and flame, as it is doing now.
similarly, looking forward, if this tool ends up working out, is developed with total transparency (which I imagine Google would prohibit), and is great – Google gets all the credit, and gets its hooks into yet another nonprofit. But if it’s a failure, again, iNat gets all the blame and its whole community suffers from burnout.
with their announcement alone, they’ve burned up a huge chunk of the goodwill and trust that iNat carries. a megacorporation only benefits from confusion, anger, and increased distrust of science. they paid no cost and can’t lose.
If this is how Google is treating you right out the gate, do you really trust them to treat you better down the line? Or will they keep using you, or manipulate you? Big red flags for me. Is there some sort of non-disclosure clause in that contract? Please, please tell me you had a lawyer review it before you signed.
I do have my fingers crossed that this can turn out useful, even work together with the community-first iNat ethos. But man, this is a costly way to start.
I am concerned that this will lead to an environment where people begin tailoring their comments to the AI that will be scraping them, instead of to the other users following the observation. With iNat’s current, very limited use of AI/ML, we already have a situation where identifiers have to consider the effects of their work on the computer algorithms - people are starting to hesitate to identify certain species, knowing that it will affect how “the computer” makes future ID suggestions. If users en masse start having to consider how “the computer” might (mis)interpret their comments, and potentially even explicitly formatting their comments to speak to the mysterious black box that is ingesting and analyzing everything, the element of human interaction is damaged. And that human-to-human interactivity is what makes iNat shine - far more than the data. Once it sinks in (even subtly/subconsciously) that your comments on iNat are not just a conversation about your favorite topic with another like-minded human, but feeding data into a machine, a major part of iNat’s unique allure may be damaged in a way that is very difficult to repair.
Like many others here, I work in tech and I like to think my apprehensions about this announcement are not coming from a place of uninformed paranoia. On the other hand, being that I work in tech I am seeing uncomfortable parallels to the current trend of executives and MBAs fantasizing about how they can eliminate the need for humans by employing the latest shiny black box. I am trying to share some others’ optimism about the small group of people behind iNat, knowing that they are good folks who care about what makes the site special. But I hope they are truly listening to all of the very very well reasoned alarm and skepticism both here and on the blog post. Oftentimes in business, once a decision is made or a contract is signed, the management must strain and contort themselves to convince everyone that it was the right decision - whether or not it really was.
I’m not any top expert, so I stand by what I said - my current ID work will be replaced by AI. And besides, I’m afraid that the cases left for the experts to ID will be not necessarily those which are most difficult but those for projects where accurate ID is important. And it might turn out that it’s not important for too many people/institutions.
And they’ve violated the licensing by doing so for many observations, like all my CC-BY-NC. iNat has permission to use my content and I don’t want to participate with genAI programs in any way. So, depending on how this goes, it’s possible that I will remove all my content from iNat or greatly restrict what goes in and how long it resides.
I’m not a huge fan of how CV is implemented. Having it as a choice would be tolerable but it is set as the default. Touch the species field and it kicks into gear. No option to turn it off and only call it when desired. I’ve developed a quick tap on a random letter solely for the purpose of stopping it.
Valid or not, rational or not, they’re still conclusions and iNat should evaluate it’s decisions carefully.
how do you suppose the proposed Q&A will actually proceed?
at one extreme, they could let every single person who wants to say something have x seconds to speak or submit a comment / question to be read, and then respond, one by one. that could potentially take many long hours and you run the risk of folks saying anything.
at the other extreme, you could have, say, a designated 1 hour session, fill the first half with a rehash of what was said in the blog, and then have a moderator “pick” only a handful of questions that align with ones that they are already intending to answer with prepared responses. that runs the risk of not delivering any new information, although i’m sure there are folks who will appreciate an alternative presentation of even the same information and the extra effort to do that.
then there are also alternative methods made possible by technology. for example, you could break out the attendees into smaller groups, have them discuss and vote for, say, the top 3 questions from each group. then transition back to the whole group and have everyone vote in a live poll for the top 10 questions to be addressed from the collection of the top 3 questions from each group. then answer the top 10 questions, plus additional ones as time permits. that theoretically cedes more control of the questions to be answered to the crowd rather than a friendly moderator but will take more time to achieve. theoretically you could have folks vote for their favorite questions to ask in advance of the actual Q&A itself, but then something like that is a little less transparent than doing it in real time.
after user testing, how do you suppose the determination will be made of whether the effort was successful / worthwhile and should continue?
one of the predecessors to this whole thing was that Natural Language Search demo. to me, it’s interesting but not very functional as an actual product. probably some of what was learned there was used to implement the recent feature where if you click to get more information on a taxon from computer vision suggestions, it will show you a taxon photo that most closely matches your own photo.
so if the thing they’re working on now ends up not being anywhere near an actual product, do you abandon it entirely or continue with the interesting things learned? who makes that decision?
Regarding Google and other companies…if you go to Account->Settings->Privacy->Learn More, iNaturalist says the following:
All iNaturalist servers are provided and hosted by Microsoft, all iNaturalist images are hosted by Amazon, and almost all iNaturalist maps are served by Google, which means every time you look at an observation on iNaturalist, you are connecting to services provided by these companies and exposing your IP address to them, at the very least (in the iPhone app we use Apple’s maps). iNaturalist probably would not exist without the services these companies provide, and while we can limit the amount and kinds of data you share with them while using iNat, we cannot stop the flow of data entirely.
That support is much of what allows iNaturalist to function the way it does. Understanding that iNaturalist was dependent on outside support, I have been a monthly supporter of iNaturalist since the fall of 2020.
Over the objections of the community, iNaturalist has now accepted a grant to experiment with GenerativeAI and they can use that money to do so. They won’t be using mine as I have suspended my monthly support.
I reiterate that I believe this new direction means a revision to the Terms of Use is needed, and quite frankly, should be a precursor to any implementation of the project.
Regarding a user-curated Wiki, I would be happy to participate (I have been editing on Wikimedia platforms since 2020), or to help explore other inputs (e.g. Wikispecies might be useful in tying users to primary sources on Biodiversity Heritage Library and elsewhere). That said, I believe iNaturalist would be better served by an in-house Wiki - I don’t think any of the Wikimedia platforms are what we need.
Between observations, annotations, identifications, comments, curations, forum posts, and direct messages, I have well over 500,000 interactions on the platform. I have the time and demonstrated the willingness to help others learn. I do not have the time to polish a turd. I’d rather just go out and pick up trash.
@catchang Unfortunately, I don’t see other Board Members on the Forum, but I certainly hope the implications of accepting this grant are being actively discussed at the board level, and if not, that you will bring this forward as a topic at the next board meeting.
I am pleased to see that iNaturalist has committed to quantify the environmental footprint of iNaturalist’s infrastructure. I look forward to seeing an annual sustainability report to clarify how the iNat infrastructure does or does not utilize renewable energy sources.
I agree with your general assessment. I do also want to say that I find it incredibly hard to believe that nobody on the iNat side saw this reaction coming. I do think iNat holds a lot of blame, because this is a PR/communications nightmare. Even if the tech is implemented and it is everything promised and more, with zero downsides… they still really, really messed up in communicating this with the community.
I’m less guarded than before about the specific proposed tool (though I’m still not exactly thrilled about it), but I’m still wildly disappointed in how iNat have handled this. They HAD to know there would be backlash from the “optics” alone (regardless of if it is warranted or not - it ultimately doesn’t matter). I’m glad that they’re open to more feedback and a live Q&A - I just hope they’re ready for what that is going to look like. I hope it will be productive. I really really do.