Well, as I said: personally, I cannot see any justification for going back on my agreement to everything stated in point 2 of the terms and conditions. It’s all or nothing: I either agree, or I don’t.
Exactly. If I want to continue using the site, I have to abide by all of the terms I originally agreed to. Cherry-picking is not an option. It’s all or nothing, where “nothing” means exactly that (i.e. account deletion).
Well, once again, IANAL, but that is what appears to be implied by point 2 of the terms of use. It’s not reasonable to expect iNaturalist to list all possible future uses of your content. The wording is deliberately broad and quite open-ended, for precisely that reason.
Nor do I - but I don’t personally think that justifies selectively withdrawing permissions that I have already granted. The extended search capability (which also requires an LLM) is much, much more useful than the summaries (which are little more than window dressing). I’m not going to throw out the baby with the bath water because of something as superficial as these AI-generated baubles.
TL;DR - Scientists discover that we don’t live in a perfect world. More at eleven.
Yes - as I already said in my post. However, deleting content needs to be done with great care. Blind deletion of content en mass cannot be reasonably justified, even if the site allows it. Doing so would affect many other users who referenced your content in their own and would often destroy the flow of the conversation. You’re not only hurting yourself by deleting all that content you painstakingly created.
In case I am not the only one who didn’t understand it, and to save you having a long list of irrelevant searches suggested by the web, IANAL means I am not a lawyer.
Though I hold no position of authority in this… wonderful forum, might I ask why one would feel the need to make an ad-hominem adjacent comment that accomplishes nothing? Please simply voice your opposition to the independent elimination (api link for the elimination of involuntary dissemination of information, free for anyone to use at any time) for the involuntary dissemination of sometimes highly personalized, and valuable information, it would really help to keep the forum an open space for friendly discussions, where we are all on an equal level, and without any sorts of snide microaggressions.
First, AI isn’t better. It relies on us training it. Second, if we can train the AI to help us whittle down the 500K+ unknown observations on this site, including ones of city trees or just absolutely nothing identifiable, then great, because it’s a slog going through unknowns.
So if we are stuck with AI, how do we humans manipulate the AI to give better answers. How can we improve our answers so that the AI product is palatable to even the fussiest person?
i’m guessing you won’t hear any additional follow-up on this from staff, unless it’s to show a new version of the demo or something else entirely. there’s definitely been a lot of effort put into this, and i bet there’s been a lot learned, but my sense based on their presentation at the Gen AI Accelerator demo day is that the demo we saw will not make it to production. when they get around to showing the product, there’s one slide that shows a just screenshot of the product, and then the next slide shows a vision of a future product that is quite far in the future from what was developed here. so that suggests to me that there’s not any intent to give up on pursuing things that could utilize AI in the future, but that it may be a while before we see an actual product.
I finally had the time to watch the webinar and then look at the ID Summaries Demo, yesterday. The first species discussed in that webinar, Hippodamia convergens, struck me as one that I might add to my identification repertoire. I have some personal experience with observing this species and the distinct, converging lines on the pronotum was a clear morphological feature that I could see in observation photos, without having to slog through the more confusing (and less visible) characters in a published key. And there were a good number of observations still at Needs ID.
I began reviewing these observations (limiting myself to the USA and Canada) and adding my IDs where these lines were clearly present. Then I came across an observation where there was a discussion in the comments about the possibility of that being H. sinuata. Although the thumbnail photo for that species is quite different in the color and markings on its elytra, that species apparently also has wide variation in those markings. While it was decided in that particular observation that the geographic range for that variation excluded H. sinuata, it left me without being able to distinguish these two species when they both have the convergent lines. Looking more closely at the other North American species, this problem also exists with H. glacialis, which may also have those convergent lines.
The comments that were the input for this demo were discussing differences between H. convergens and various species of Coccinella. These differences may be true and useful. The comments apparently did not go far enough into distinguishing all of the other species of Hippodamia.
I’ll note that when comparing the observations of these species, H. convergens is much more common than the other two, by more than two orders of magnitude. Taking that into account, and also considering the cases where I’m confirming the ID of the top identifier of the species, it’s statistically unlikely I made any mistakes last night. On the other hand, these similarities may be contributing to false identifications for the more common species. Are there other visible morphological differences I can use to be sure of my IDs? I’m an amateur, trying to take this responsibility seriously.
These issues have already been raised, in the abstract, in this discussion. Take this personal experience as a real-life case study.
Your experience in trying to use the output from the ID Summaries demo to identify Hippodamia convergens was a really interesting case study that I think illustrates one of the limitations I also found in the demo—the ID tips offered and the way they’re presented tend to be less comprehensive than one would ideally like to see.
Lots of other human-generated works have tried to address similar ID challenges and the starting point for anything like this is to set an appropriate taxonomic and geographic scope:
The scope can be as broad or narrow as you like. The point is that in a human-generated guide, the descriptions, key characters, tips, photos, etc. are all written in the context of that scope, with the goal of distinguishing each taxon from any others that it might be confused with. The writer has a mental picture of a reader trying to identify one of the subject taxa and the possible lookalikes that need to be distinguished.
So far, the ID Summaries demo seem to be quite limited in that respect and I can see three possible reasons:
The very limited nature of the source texts (comments from less than 40 iNat users) is limiting the variety of identification advice. With a broader range of sources, ID tips may be more comprehensive.
The way that the LLM prompts were written for this initial demo did not sufficiently stress the importance of distinguishing each taxon from any it might be confused with. Choosing better prompts might generate more comprehensive ID tips.
The innate nature of iNat comments is often limited to comparing a small set of possible identifications. This is a narrower task than providing comprehensive ID tips and this therefore sets an upper limit on the value that the ID summaries can provide.
Problem 1 can be addressed with more data, and Problem 2 with better training, but it seems possible that Problem 3 exposes the gap between LLM predictions and the kind of knowledge synthesis performed by a human field guide writer.
Can an LLM distill a bunch of isolated pieces of ID advice into a comprehensive ID guide for a particular taxon? I guess we’ll see.
For me, if a later version of the ID Summaries demo can generate accurate advice in a format like the following then we would be getting close.
In many cases, Hippodamia convergens can be identified by the converging lines on the pronotum, which reliably separate it from various species of Coccinella. However, other Hippodamia species may show quite similar converging lines on the pronotum, and in areas where multiple species occur, identifiers need to take care to distinguish among them. In H. sinuata, the lines on the pronotum…