Amazing work team! I love the feature and LLM summarization has great promise to collate all the human knowledge federated across hundreds of threads and posts. Love it
4 Likes
@juliereid It is really beautiful to me that you brought up how you feel about this. Thank you. As in all things AI, that’s the first thing that goes. I have had such lovely (if brief) connections on iNat with people who have either shared my enthusiasm, offered personal encouragement and connection, or (occasionally) engaged in really excellent discussions about natural history. None of these are available in any meaningful way through AI.
My background is in psychology. A key factor in happiness is human engagement, even for extreme introverts. This system does not provide that cheering mental health benefit. AI can’t tell me they too live with a parrot or that they share my enthusiasm for the Baraboo Bluffs. Quotes can’t share someone’s life-transforming experience at a Hawk Watch. I am super fortunate to have a lovely friendship with several fellow iNaturalists. We don’t keep up closely, but what a delight it is when we do connect. Even brief comments or seeing someone’s humor or enthusiasm for learning is such a positive.
An important issue here is that the tone is academic. I’m all but defense on a PhD I walked away from, so I have a high tolerance. But jargon limits engagement. While it could be argued it is more accurate to use scientific names and discuss taxonomy, that decision shuts out some. A human can match tone in responding and engaging. This is what outstanding teachers do. One thing I find delightful is that people in the country around here are excellent birders, but have entirely different names for things.
We already have constantly improving AI IDs for many IDable organisms, here or on our smart phones. Listing key field marks is helpful, but has to be tailored to the image or recording provided. Only humans can do that. And you can’t tag AI to ask for help or share aspects of the experience. Having said that, I would use this tool if you could link to it and do what the Cornell Lab’s excellent All About Birds pages do: Tell me what species are similar and (ideally in non-technical language, although I get that’s sometimes important) how they differ, and all with illustrative photos. I don’t think the quotes help unless you’re just trying to feature and thank your super IDers. It’s redundant and harder to polish the lesson, and it is never going to be a satisfying human interaction.
And wow. I find the winner-take-most message here off-putting. I put in as much spare time as I can manage IDing. I honestly can’t imagine how people put in this much more time IDing, but it’s demotivating if good to know that my contribution is negligible. Thank you for helping me set better priorities. But oof. The time I’ve wasted here is pretty staggering. Onward!
And thank you to whoever these people are. I hope that part of your excellent work includes stopping to respond to comments or questions, encourage new observers, or acknowledge when people have clearly done a lot of work to present beautiful photos. That’s work that I don’t see AI replacing.That kind of work is what sticks with us. We are social creatures in so many ways. Even the shyest among us.
14 Likes
I do not support this effort. My background is as an energy engineer, environmental activist, and sustainable technologist. I’ve been using iNat for over 5 years.
-
Like others have said, it would be more meaningful for me to read the actual comments instead of an AI summary. I find computer vision less useful to me than user comments. I am constantly meeting iNat users do don’t understand how computer vision works, and were shocked that AI generated images could be posted because they assumed CV could stop that.
-
I do not want to provide feedback on the accuracy of the summary, that is labor I do not wish upon myself or others. I think it lacks empathy and consent, and makes it harder for us to hold individuals accountable for their contributions to the community.
-
There are more meaningful ways of understanding how expert contributors might teach others about how to ID things if you are going to be reaching out to them anyway, that are actually designed with the audience in mind with the considerations for the positionality of a contributor who has had the privilege to both the safety and education to participate to the extent they have so far
-
I think a more rigorous energy assessment is needed to understand energy use at scale, with examples that aren’t a 100 W light bulb: car emissions, use of streaming services, energy used to ride a bike, sending emails
Vulnerable populations are underrepresented even if they have the skills to be strong contributors because they worry about their own safety or the safety of their subjects. I am talking about indigenous people who are trying to protect their native lands, I am talking about low-income folx who haven’t had the chance at a formal degree in ecology, I am talking about displaced disapora trying to reconnect with nature after losing their homelands and ancestors.
Multi-marginalized people have been exploited in the infrastructure development for AI, are at the most risk of the increased socialized cost of this technology, and deal with extremely poor labor conditions in training AI. I do not consider a small scale LLM to be independent of this history given the research and funding incentives that are driving this initiative. There are plenty of threads that explain why vulnerable communities don’t want to participate in iNat and AI has not been a critical factor in resolving those concerns. There are people who do contribute to iNat and regret it. Clicking delete on ones own post isn’t a permanent way of taking something truly down from the internet.
Here is what is more useful to me: I am a victim of stalking. I cannot provide the same photo on iNat on any of my anonymous profiles on social media without exposing my current location or location history. As a nature educator, I can never recommend iNat in good faith by only talking about its positive research impacts without also talking about the risks and ethics in the context of social, racial, and climate justice.
Banning is not enough and if iNat cannot enforce more support for user safety in a preventative manner, this needs to be as clear as possible to a new user: before a teacher asks their students to use iNat in class. We have a lot of ways to protect endangered species but this is not covered when most people first learn about iNat. They should know about this beforehand so that they can be prepared if/when safety measures come in after the fact.
I would like an improved onboarding experience that is transparent to users, no matter their age or background, about cybersecurity risks before posting an image. I would love if there were mechanisms for communities of people to collectively share iNaturalist contributions under one rep so that the data can be obscured from the individual without being able to tell that it’s actually 2-3 people just sharing an account.
Designing from the margins will be more abundant than we may fear and doesn’t need to be mutually exclusive from our open data mindset. If not facilitated with due diligence, it furthers colonial ideas of what it means to ask vulnerable populations to give institutions data for free under western scientific criteria for what “good data” is, with little consideration for what is actually useful to those populations or providing any form of reparations.
Also @juliereid : thank you for your comment. i saw in another thread that u are also from the LGBT community. I came out in 2015. It means a lot to me that you wrote your comment. Please know that your existence makes the world feel safer for those who come after you. I didn’t know if I should speak out but the more I thought about your comment the more I felt empowered to do so.
16 Likes
I agree that wiki is a good idea, and definitely a much simpler one. AI summary (referencing existing comments) is interesting but it has more caveats to it… But it definitely could be useful as a start for a wiki page.
I hope iNat staff listens to the community and finds the best middle ground.
I would definitely love to edit a wiki page rather than leave comments with tips, I would put a lot more effort into it! (meanwhile my ID tips can be very inconsistent… and probably dependent on many factors such as location, photo quality etc)
19 Likes
I’m happy to see this released so we’re aware of the status of the project and a preliminary idea of what it looks like. I like the general format including the inclusion of source comments and multiple sections for different aspects of the species. A couple questions: How are those sections selected? And are comments within IDs included or just isolated comments?
It’s clearly stated that this is an early mockup with limited training data, so I feel like some people are overreacting about some of the obvious concerns like redundancy, missing important context (e.g. range, life stage), and advanced vocabulary. Many of these were brought up in the initial thread (along with general AI concerns, suggestions of a wiki instead, etc. etc.) in speculation about what the demo might look like, and seeing them in early stages of this experiment seems expected; I wouldn’t expect them to be all fixed at this point. Presumably what happens with the results of the experiment will depend on how successfully these can be addressed with improved prompting or whatever goes on behind the scenes of something like this. Maybe it will stay like the Vision Language Demo and never leave the stage of being an interesting experimental demo page. Or some parts will be used and other parts won’t.
9 Likes
I would strongly prefer the ability to write identification guides that are easily accessible during the identification process. I’m imagining an embedded system that allows users to create locally relevant dichotomous keys, or, as others have suggested, a community wiki. This AI system is probably useful, but what I personally enjoy and find useful is learning from well thought out, comprehensive identification guides, rather than a smattering of comments that may be of limited use to the situation at hand.
I also think this will limit my enjoyment of the site from an aesthetic point of view. If I am constantly exposed to a flood of AI summaries of others opinions, it will make it more difficult to make my own unbiased identification of the observation at hand. Not to mention it adds additional clutter and noise to the website. I understand that this is not currently the case but it does seem to be the general trend. I guess if this is implemented, I’d like to see it either limited in scope or easily disabled.
20 Likes
I have questions. The first one that comes to mind is: if users vote that a summary is untrue or not clear, will this lead to revisioms? If so, how? If not, why?
I guess I will just have to sign up for the Q&A session, and stay tuned…
6 Likes
could be similar to the projects list on obs. More. Or less. As we prefer.
Perhaps a default setting - do not show me this. But I would prefer to see it for a while - then decide if I would prefer to ask if there is ID info for this obs.
2 Likes
It seems like a lot of people are wanting wikis. I wonder if there’s a world were we can get the best of both worlds and use the AI tools to help craft wikis. One of the reasons I’m looking forward to the AI tool is that I no longer have the time to creating wikis and guides and don’t foresee having that time for some time to come. As such, if someone wants a wiki for the Euphorbias I’ve commented on in the past, it will be up to someone else to create them from my old comments and existing resources.
Also, I wonder if a tool that directly compares two different species at a time would be more practical. I’m thinking with two prompts where you could add a species to each one. Something like this could easily be facilitated using the AI tool. Perhaps prompts could even be linked to withdrawn disagreeing identifications?
21 Likes
This is the only way I’d want my identification tips to be used by a LLM, personally. I like the idea of using the LLM results to find observations to refer to and cite when writing the wikis. But I want to be clear, I’d prefer that they be used while writing a wiki, not for generating a “first draft” or “default” to be voted on or edited later. I think that this would ultimately save everyone a lot of work revising/fixing/voting (and stressing!), and prevent folks from being mislead in the meantime before corrections can be made for each taxon- more human work up front, but ultimately less overall IMO. And that way we could still make use of the LLM results- just in a more hands-on/nuanced way.
Also, for the record, there are numerous guides already written and available on iNat (like the one I made that lives as a project journal article that I know is hard to find). It would be great if there was a way for those to be found and used more directly than to be fed into a LLM and summarized to the point where they lose context and become less helpful or misleading.
18 Likes
To put it briefly: I understand that the development of this feature/demo is probably a deliverable for a grant and it will likely be implemented regardless of community feedback. However, I want to express that I have no confidence in the accuracy of an LLM when it comes to summarizing identifying features, particularly for groups that are reported and IDed less frequently than flashy, easily observable charismatic taxa.
I’m concerned about the process of rectifying inaccurate information as well. If a summary is not useful or contains completely wrong info, will it be possible to fix that in any timely fashion? (I already have to deal with bad CV IDs for spiders… a problem that appears to be intractable, so I’m not eager to add more computer-powered ways to mislead observers.) Staff are not taxon experts- to expect them to manage corrections on a wide variety of topics that they themselves are not experienced with seems totally unreasonable to me. The upkeep that this feature might require in order to produce useful results could be so extensive that the “convenience” of using the LLM in the first place could be completely negated.
A plainly written summary of ID tips would genuinely be a very useful thing to have on taxon pages, but I’m really not convinced that using LLM output is actually more convenient than sourcing help from the userbase.
16 Likes
Wouldn’t it be better to just have an editable “community notes” type of section? Like one we can just write ourselves. I have no faith that this ai program will be able to accurately summarize identifying features.
15 Likes
I expect it would not only be better but also easier. I’ve worked in and around software for over 12 years; I’d be astonished if making community-editable ID notes via a wiki on the taxon page is a harder thing to build and test than a whole AI summarization system with its own back-end systems and a user interface that needs to be inserted into the existing platform.
13 Likes
Yes, everything in the “User Comment References” section is verbatim and displays the full original comment/remark.
This demo won’t have dynamically updating/regenerating content (this spreadsheet has some more detail about how many prompts and such we anticipate going into it).
In the future we might update the demo content periodically (at intervals we haven’t yet determined) and incorporate the voting feedback directly at that point. Until then, people interacting with the demo will be able to see the vote counts on the page. Much of this system was inspired by the current annotation voting and DQA systems on iNat observation pages. We’re also still in the process of exploring options for edge cases of needing to entirely remove content.
Thanks for the questions and suggestions! I’m hoping this response addresses the question below, too.
8 Likes
@tiwane Since the iNat community is iNat (the two cannot be separated, and much harm can come for displeasing even a small percentage of the community), how will the decision be made whether to implement one of several potential versions of this project? Will there be some kind of a formal vote? How will iNat staff know what proportion of the community supports the project in the end? Will that be taken into account in some quantifiable way?
iNat has never done anything this controversial, and many people feel that their voices are not being taken into account. I think much of the gnashing of teeth could be avoided by ensuring the community that quantifying user buy-in will be a priority.
Thanks!
12 Likes
Just really want to thank everyone for sharing their thoughts in a constructive way in this discussion.
I think making the comments visible like they are in the demo could actually facilitate more discussion and interaction, since you’ll potentially be able to find individuals whose comments you could refer to directly. For example, if I’m not sure about a butterfly ID, I could say “@sambiology does this butterfly have the Viceroy’s black band you’ve referred to?” Which I think could be pretty cool.
To be clear, this is only using comments from a handful of users. As I discussed in a previous post, this does have clear limitations, but again this is just an experiment with a known small set of inputs and variables.
I’m not sure if there’s a specific skill level we’re aiming for. I think mostly we’re just trying to summarize what people have said in a clear and accurate manner. If people use a lot of technical terms, those would probably show up in a summary.
I definitely also find the wiki idea intriguing, and as @nathantaylor said, this could be a tool to help get wikis started in some fashion. One concern I have about a wiki are moderation issues that would arise*, and also that it would be just extra work on top of making IDs. I like the idea that if I craft solid helpful ID comments when identifying an observation and communicating directly to an observer, that comment could also be used to inform these tips, so I can just focus on observations and helping people directly.
*(As the main staff point of contact for moderation issues, I’ll note that I have a skewed view because I get called in the few times when things are bad and not all the times things work out fine.)
27 Likes
Thank you iNaturalist staff for sharing a progress update! I look forward to more details and discussion.
I’m intrigued as to how often the guides created by this tool will be updated and reviewed, and if there are any plans or ideas for how many species will be rolled out per iteration. Since it seems fairly clear there will need to be some level of human review, it doesn’t seem likely that releasing thousands or hundreds (or even dozens!) of new species per month would be feasible.
Relatedly, I’d be curious as to how species are being prioritized as needing a guide. I would assume a lot of the most commonly observed species probably don’t really need one, or certainly not as much as more obscure taxa. But if we’re really focused on all levels of expertise, then maybe that’s a bad assumption.
I’m also interested in the wiki idea, though I share the moderation concerns there as well. A wiki-style guide, depending on how it’s presented, could be perceived as even more authoritative than something that clearly references comments from individual users, shows votes in agreement, etc.
- As others have mentioned, iNat staff are not necessarily taxon experts, so who gets the “final say” on what appears in a guide for a species? Who decides who gets that final say? Would there be some sort of credentialing that needs to happen? How do we prevent gatekeeping? Would the guides themselves need some sort of “Research Grade” status, voted on by the community?
- In my experience on these forums, even taxon experts frequently don’t agree with each other on what the most relevant details are. In my experience on Wikipedia, the moderation discussion for some entries is orders of magnitude longer than the entry itself. Is that a good use of everyone’s time, or a good use of iNaturalist’s resources?
In general, I don’t think leaving whether or not this or any particular feature will be implemented up to “quantifiable user buy-in” is a good precedent to set. How many people supported Computer Vision when it was launched? If iNaturalist somehow got everyone who’s ever posted an observation to vote, and 99% of the community supports any given proposal, that leaves tens of thousands of users who might still feel like their voices are not being taken into account. What then?
8 Likes
So, I’m pretty excited about this. I know that this mock-up is in early stages, and I recognize that it will change over time. I especially appreciate that the staff emphasize the technology principles and are quite transparent about this as the demo/program/initiative rolls out.
My excitement for this is also quite personal – one of the iNatters that originally nourished me through his identifications was Greg Lasley. Greg would painstakingly put in information on the ID’s with comments of why it was this species or that species. Now, Greg is no longer alive, but his comments still speak on those ID’s. I foresee that this LLM could take his comments and project them to others that are learning the differences among damselflies. Rather than having to dig through loads and loads of ID’s, hoping for an explanation, I imagine that this project would amplify these explanations, making them easier for the general user to see/read and learn more about the organisms. Thousands of these comments and explanations exist on iNat already by thousands of users – even those not with us anymore, like Greg – I would imagine that this tool make those easier to access.
*edited to recognize the “opt in” nature of this… just an example of how someone’s comments on an observation could be more impactful than just on that single obs
Looking forward to the next steps!
16 Likes
There’s a few other thoughts I have at the moment,
- It seems that including our comments in the data is opt-in. That is good, it is at least better than having it be opt-out, or worse, no option. In theory, I would like this, but… if I don’t want to provide my comments, then my ID tips will not be included, and there may be more editing needed to add that missing information. It seems tedious, and gives me an incentive to opt-in, though I really do not want to. Yet, if I do provide them, I do not trust that the LLM will interpret them correctly, and people will still have to fix that as well.
Having to correct an AI feels tedious. I love IDing and informing people on how to ID, but an LLM does not truly understand what you say, so it feels more like I would be fighting it, not teaching it. I would rather spend my time on iNat normally IDing.
2. You make a very, very good point.
Many identification arguments occur under observations. To put ID tips in a seemingly definitive guide would probably cause even more contention. So, perhaps, a typical wiki format which coalesces everyone’s opinions into one document is not a reliable idea, similar to the problems with the LLM summary. I can just imagine the edit warring. And it’s hard to verify who is and isn’t “qualified” to speak about a species, and may silence most people.
So, maybe forums for taxa instead? People can put their own ID tips, and others can vote and comment on them. It can still represent the community’s knowledge without preventing anyone from speaking. In theory… I don’t know for sure. But that’s what we’re here to discuss it for :)
7 Likes
I want to start off my comment by mentioning that I was pretty optimistic about the possibilities of the LLM/AI application to summarizing the disparate and varied comments across observations. However, after seeing the demo version, I have to say that I am unimpressed and actually a bit concerned about the value of this application as currently envisioned as being confusing at best, and a waste of identifier effort at worst.
I have to agree. It’s not particularly useful to a non-expert, or at least an amateur, which I thought was the whole point of the LLM summaries.
Could it at least translate some of the jargon into plain English? As a plant person primarily, I didn’t understand most of the terminology e.g. what’s a “trailing edge” what are “hindwings”, then I could see some value in these types of summaries.
Agreed. I think quality control is going to be a major issue if the demo version is any indication. What would be the consequences of marking the summary as incorrect or not useful even be? Would they be revised by re-running the model? Changing the training set? Human revision?
I was skeptical seeing these types of comments at first, but I agree that in comparing the AI summaries to the original comments, I don’t really see the value in the summaries. Others have mentioned it does repeat itself in the demo, and I could imagine it making mistakes that would only lead to more confusion.
Wasn’t there an earlier version of this type of LLM application being considered to actually look at good models and use them to explain what the computer was seeing to determine the CV suggestions?
I think this would be a lot more useful to the layperson and expert in related taxa alike.
A possible idea to not throw the baby out with the bathwater here, maybe experts with a particular comment count and “correct” percentage could be used to train bots of themselves? For instance an @CarrieSeltzerBot or @tiwanebot? It sounds a little dystopian, but we’re already really doing the same thing with the proposed model, it’s just an aggregation of individual users under the guise of a “combined” expertise.
Another possibility for improvement, since I think this is a long way from being useful for solving the problem as currently defined (too many observations, not enough identifiers and experts to guide the IDs): Maybe the LLMs could be cross-validated against Google’s AI summaries of how to identify various species, as well as in comparison to similar species. Though I see this would take a lot more power, if future iterations follow what we’re seeing here, it’d be a waste of energy (both machine and human) to continue unless some major improvements are made.
I still see the possibility for this technology and approach to prove very useful, but again, I’m surprised at the lack of utility of the comment summaries displayed of the demo itself. Improvements can be made, but again the mock-up underwhelmed me to a degree that I’m now cautiously skeptical of the whole endeavor from what I’m seeing.
7 Likes