Not only am I not cautiously optimistic, I haven’t commented at all since the events being summarized. This hallucination reinforces my lack of cautious optimism.
I really do appreciate that the iNat team has continued to share their thoughts here. In particular, I’m still thinking about @tiwane’s great post earlier about using a model to help collect and filter “a page of sourced comments for each taxon in the model”. This application has some good use, seems feasible, and - most importantly - collects user-written ID remarks and shows them to you, instead of trying to make up its own sentences via a GenAI language generation model.
Yet, I have to reiterate the pressing need for iNat to say something official about its planned use (or not?) of generative AI. I get that not everything is decided yet. That’s fine - we don’t need or expect all the answers right now. Development takes time. But what we do need, and expect, is some clarity.
I’m still not sure what the answers are to the crucial questions:
Is the use of GenAI to write up its own natural language explanations for species IDs still being pursued?
Is the plan detailed in the blog post - the latest official news - still the latest plan for the use of the grant money?
I’d really like to know. Because while I genuinely appreciate the recent comments from the iNat team conveying their distaste for AI slop, I also need to see, in actions, that iNat is actually distancing itself from AI slop, by making it clear that their earlier plan of AI-generated text descriptions is no longer top of mind. If I sound like a broken record by now it’s because I think this is important. Most people won’t read this forum thread, or find all the staff comments that have given me some confidence and trust - those need to be shared much more broadly.
I wonder how much published research using iNat data, has been left in limbo by rage-quitters. That is sad for the working scientists.
I was afraid I would find a new pile of Needs ID among my bookmarked URLs. So far so good, iNatters continue to observe and ID on iNat. Toxic social media has (mostly) moved to a fresh target?
Hello again folks, and may this early afternoon find you at peace and in health.
I hear the frustration and continued confusion, and I absolutely understand where they are originating. There are definitely mixed signals, and a lack of consistent language between the blog, social media, and forums.
However, it’s the weekend now. It’s been a rough couple days. I suggest, with complete sincerity, that those of us who are upset take some time this weekend to rest. Spend time with loved ones, a hobby, or with nature.
It’s tough to exist when there’s intrusive fears spinning in one’s mind, but there are healthy methods of coping with them.
You don’t have to accept an apology, of course. Each of you has the right to walk away, or to be angry. But there’s also the option of waiting. It’s been a tense week in California - and elsewhere. I’ve got my fingers crossed that refreshed minds and flexibility can bring clarity next week.
I’d rather see something acknowledging sources in orderly fashion (the way sci papers do) also including context/caveat/limit reminders (re: geography, taxonomy, minority opinions etc.) - no filtering of dissent, no dumbing down. But hey, it’s still in its infancy, not even staff knows clearly what it’s all about, won’t spend braintime on pointless feature requests ^^
@DianaStuder If the “working scientists” don’t know how to properly handle, store, and archive the source materials and source data of their studies, the blame is on them. (Incidentally, I learned yesterday that DOIs returned by GBIF point to the live -not archived- results of queries, thus possibly inconsistent if containing data aggregated from iNat). Or else, let’s forbid correcting/casualizing/deleting (in essence: let’s freeze) the many obs that have been used in studies.
Thank you @tiwane for both a perspective from the inat team and also your own opinions
Well of course AI says you like it and that it is the best thing in the world and will prevail over us puny humans (OK not that bad but you get the point).
@loarie, is there going to be a new journal entry on the main website blog?
- @graysquirrel’s comment and your comment are buried within this 400+ forum topic.
- The existing blog entry has hundreds of comments as well and is locked down.
- The TLDR (too long, don’t read) crowd might get the wrong idea about where iNat is heading, even if they read @carrieseltzer’s update to the existing journal post.
- Not all regular website users are forum users.
Even better, if this information is provided on the dashboard / announcement banner.
For you, yes. You are willing to tolerate a disinformation machine spewing nonsense (“the enemy of the good”) in the hopes of someday achieving an impossible perfect (some sort of AI that can actually understand and synthesize information, which I don’t even agree would be a positive development anyway).
Making sure that your website isn’t lying about what users say might be necessary if you want people to continue using it, yeah. I think you’ll find that people’s tolerance for that behavior is quite low.
when i see other folks interacting with a lot of things that incorporate AI in non-critical applications, i actually find that their tolerance for occasional incorrect information is usually quite high because they value the incredible speed and convenience with which that information can be delivered.
i find that with experience, users tend to develop a general feel for any tool’s limits (AI or not ), and over time, users learn to work with that tool to get better results. (ex. many users of iNat’s existing computer vision system know that they often can more tightly crop their intended subject to get more useful suggestions.)
that process of familiarization also is not so different from talking to a particular human and figuring out over time what you can absolutely trust that human to know or be able to figure out, what that human might not know or be able to figure out, what biases and blind spots they might have, and how to ask the right questions and otherwise interact with them to get the best answers.
there’s a difference between disinformation and misinformation. if iNat’s proposed system showed a pattern of magnifying the most harmful information or showed some pattern of presenting things in a consistently misleading way, that’s different from it just being wrong sometimes.
again, we know it’s likely that some level of misinformation will be produced. what’s a tolerable level before it might be interesting? what’s a tolerable level before it should be unleashed to the masses? iNat staff have promised not to release “slop”. so here’s your chance to define what “slop” means to you. if your line in the sand is still that it must be 100% correct, then that’s fine.
I have printed field guides where I have found factual errors and misinterpretations of evidence that I’m familiar with. Nothing in this world is 100% perfect or correct.
I know. But, my point is simply that this has the potential to hurt the iNaturalist community (and possibly the scientific community) more than the staff depending on who leaves.
There are several observers on here who, if they left, significant biological information would simply be gone. We can hope that they would take their observations elsewhere, but I haven’t found a better place for sharing biological information. Take the observations of Euphorbia duckei as an example. It is represented by only three occurrences on GBIF (and I know of no others to this point outside of GBIF as it was known only from the type at the time of description, and I don’t think anyone has published on the species since). One of them is the type specimen which has vague collection information and gives little information about the habitat. The other two are iNat observations from a single observer that have photos showing habit and habitat and have coordinates showing exactly where the plants grow. From those two observations, we get a lot of important biological information about the species including an understanding including that it appears to be semi-aquatic, making it the only semi-aquatic member of sect. Anisophyllum. If that observer were among those who deleted their accounts, we’d lose 2/3rds of the occurrences on this species.
If the only Euphorbia species the observers who left ever posted were E. maculata from almost anywhere in the US or Europe, then I agree with you (on their Euphorbia observations). They are a dime a dozen at this point. But I don’t want to assume that.
So you ( and all those making broad comments against AI and Google) are against developments like Alphafold then too?
i went out to the Internet Archive’s Wayback Machine to see if those pages were archived. one already was, and i just saved the other one. it wouldn’t be a terrible idea for someone to create a third-party site where folks could submit significant iNaturalist records (and potentially records from other places, too), with explanations of why they are significant, and after review, it could save those either in that third-party site (if licenses allow) or in the Wayback Machine or both. it would be nice be able to see a list of significant observations like that, with explanations of why they are significant.
but that’s beyond the scope of this topic.
If not already done, I’d advise the authors of observations of prime scientific value to try and have such data published somewhere (no, I don’t necessarily mean Nature, could be the Quarterly Letter of British Euphorbia Lovers). Then archive it, together with pics and relevant metadata, in various repositories - at least one beyond a single online website operated by a 501 org - e.g. through national academia, or gifting offprints to remote libraries. To me it seems a bit over-optimistic for scientists to assume that iNat (or GBIF, or…) will still exist and deliver bit-for-bit their datasets 50 years from now.
(I hope iNat and its hosts have some disaster recovery plans already in place, but…)
/off-topic
Please do not fall into the trap of treating AI as if it’s a monolith.
Those of us who oppose the use of AI-generated text descriptions on iNat are opposed to the use of AI-generated text descriptions on iNat.
We have never been opposed to the use of image recognition AI (iNat’s CV, used since 2017), nor have we ever stated any opposition to machine learning tools in general. Such tools, carefully applied to technical applications such as protein folding (AlphaFold, used since 2018) and the classification of astronomical objects, are sensible and useful applications.
Note that none of the tools listed above involve any AI-generated text, images, or content.
None of the tools listed above involve the use of generative AI (GenAI) to write natural-language sentences that try to explain things to the user.
If you read the blog post, you will see that the use of GenAI to generate text descriptions explaining the reasoning behind a species suggestion is, despite some vagueness, exactly what it proposes.
It is this use of GenAI that we continue to oppose.
Recent forum posts by @loarie and @tiwane seem to suggest that iNat staff, at least on a personal level, may also be opposed to this use of GenAI, due to the myriad technical issues, practical issues, usability issues, drawbacks, and risks many have already stated.
However, until an official announcement from iNat is made that indicates this, I must continue to oppose the implementation of a GenAI language generation model as specified in the blog post.
Sorry, that’s simply not true.
Alphafold generates 3D models of proteins.
It very much generates visual content.
You talk about others seeing AI as a monolith.
Ironically, it seems to me that it’s those speaking so vehemently against this plan who are doing exactly that - flattening AI tools into one undifferentiated category.
As others have said above me, this response is even less reassuring than the original post. As far as I can tell, nothing about the original plan has changed here, all the responses seem to indicate that yes, the intent remains to incorporate LLM generative AI into this project except… I guess we’re just not calling it generative AI because people got upset? Frankly a very disappointing response to reduce many hundreds of users valid concerns to some unwarranted premature panic and unjust hatred of “AI slop”.
The “interaction with a large language model” is the entire problem we were expressing concerns about. LLMs, in case it wasn’t clear, are generative AI, no matter what you prefer to call them.
I’m not an A.I. expert, but I’m pretty sure AlphaFold only predicts protein shapes as numeric data, and standard 3D modelling software is used to convert that into rotating, coloured video content, not to mention superimposing the “ground truth” measurements of the real-world proteins.
Strictly speaking, AlphaFold can’t even be called generative A.I., since it only predicts a single outcome for each possible input.
Just now I was writing an ID comment about how Rosa multiflora is evil, treacherous, and creepy.
…that’s going to potentially influence an official description? 