I attended the Ambassador program talk as I could for certain easily be an iNaturalist Ambassador, but I still have looming questions and concerns surrounding the addition of GenAI to iNaturalist.
The thing I found most interesting is that a core iNaturalist’s Mission and Value in the Code of Conduct is
Emphasize that iNaturalist data should be accurate and relevant since it is used for science
We have already shown in this thread how GenAI gets ID information wrong, how it falsely attributes statements to users who did not say what it says they said, and more. Using GenAI to create ID tips will create falsehoods, and thus, goes against the supposed iNaturalist Code of Conduct.
In addition, OpenAI just released their own research which confirms our statements that these “Hallucinations” are actually a structural property inherent to these machines.
Thus, according to iNaturalist’s own stated ethics in their own Code of Conduct, as a future ambassador looking to maintain the scientific integrity of iNaturalist in alliance with their Code of Conduct, I am hereby requesting that iNaturalist immediately cease and desist all use of GenAI and do not now or ever deploy it on any of their products, programs, sites, aps, or architecture.
I look forward to your formal and publicly announced and posted reply as to the status of your GenAI project.
Humans also “get ID information wrong, falsely attribute statements to users”, etc. If sometimes making mistakes and misattributions renders one unfit to contribute to iNat, then we should all be kicked off. The question isn’t whether AI generated ID tips will be infallible; the question is whether their inclusion will ultimately lead to users making more correct IDs or more incorrect IDs, compared to the system currently in place where no ID pointers are present. Would the benefit of ID pointers that are mostly correct outweigh the potential harm of some incorrect ones showing up? I’ve made 362,371 IDs, but only 356,094 of them are still active. That’s a 1.7% change-my-mind rate on my part! I “create falsehoods, and thus, go against the supposed iNaturalist Code of Conduct”. Plus as an American I have a huge carbon footprint. I might be worse than AI…
If iNat staff is still investigating the pros and cons of various versions of implementation of these “ID pointers”, I definitely don’t want them coming out and saying they will “not now or ever deploy [AI] on any of their products, programs, sites, aps, or architecture”. Nor do I want them rushing in and deploying it without careful consideration. I want them to spend all the time they need comparing all possible avenues. If it’s going to cause more harm than good, then scrap it.
And from an environmental energy use standpoint- I’ve seen students use iNat to try to identify things. Guess what they do when iNat won’t give them a solid answer: they ask Gemini or ChatGTP. If anything, using an LLM one time “in house” to provide the info right here might ultimately lead to less AI use overall, since users don’t have to go ask their favorite AI for help every time they’re stuck between two species. How much energy is used every year by hundreds of people typing “Northern vs. Southern Catalpa” into Google and regenerating an automated AI summary each time because the information isn’t available right here on iNat? How accurate are the Google results that people get when they make these searches?
There are so many factors at play here, and science is a long, slow process. My current take is that the relevant data to make an informed decision on this issue does not yet exist. Thus my only opinion is that I’m against the “we want answers now!!” take, which unfortunately seems to be a popular one.
A user who is posting hallucinations and is influencing thereby a handful of other users can be taught to improve. Even if he/she does not, the damage is limited, and ultimately, this person could be “kicked off.”
A system with an inherent AI engine that does the same cannot be easily influenced. It influences ALL observers and IDers. And it cannot be “kicked off,” as it is inherently part of the process and system.
About the risk, take the risk attributed to a single hallucinating human user and multiply it by the number of active observers and IDers in INat.
That’s why people go crazy about the idea having AI as a super-influencer.
Why should somebody care about peoples’ hallucinations in future if the system is accepting AI hallucinations?
Thank you for not giving up acting against this decay!
I also thought about the Ambassador program. Being now abstinent since some weeks, I decided that for me it’s either too early or too late, pending the outcome of this story.
To be honest I still don’t really understand how an iNat AI experiment with funding from Google is practically different in any way from the iNat AI experiment with funding from Microsoft which was done in early 2024 and made barely a ripple in the iNat community. The proposed experiment I guess will be moderately more generative but it’s still presumably going to only exist in a similar restricted experimental environment.
That really was an “experiment”. That functionality ended up not really being used in the main site and is just an as is interesting thing you can play with. It still is a demo.
We’re now half way between when this was announced and the end of the year. Could you give any updates on the general timeline?
I’m also interested in an update or any additional info you can provide on the Q&A session. When it was announced, it seemed to be like it was meant to address the questions around the grant. I may have misunderstood its purpose. When is it supposed to take place? Is it relative to the release of this experiment/feature? What is it going to cover (assuming it isn’t about the grant as I had thought)?
Following up on this, have you found anything else? We’ve made it to December without any further information, as far as I am aware. I haven’t used iNat since this announcement aside from checking to see if any new information has been posted, and I would like to know whether they intend to move forward with integrating generative AI so that I may make decisions accordingly.
I watched all the webinar and tried for about an hour in the demo.
I do support much of what is said in the webinar and agree with the good intentions.
But for an objective statement, one should also involve people with more foundational information technology background. Overall, the many opposing and factual arguments of AI sceptics are not represented, except for the environmental impact, which I personally believe is largely exagerated and can be relativized by a bit of simple math (as it is shown) and as long as cities are still bombed and forrests still burn.
My view on the demo: Pro:
+It shows what a good WIKI summary could do in terms of educating users and giving them a quick start.
I took a look at some common geraniums, ladybugs and other species that I know from my area, and for those, the summary was helpful.
+The option to rate the entries is also a good idea
+Links to external references can help users to grow further into the topics Con:
+It is intransparent how much content is generated by LLM and how much has been curated/was withdrawn to come to this state
+The “source of truth” for the LLM is from within INat and is not distinguishing experts from non-experts. Therefore, a lot of curation is needed to keep the information quality of such a dynamic LLM generated WIKI at level. Need more experts ?
+The “Logic” of the WIKI and the rating option does not indicate clearly from a mathematical perspective, which individual features are necessary to document to distinguish from similar species and which combination of species features are “sufficient” to justify a true determination. Instead of preventing arguments, I am not sure if it may therefore not even provoke more disputes in the end.
I am missing an independent “flag” at the top, telling “cannot be determined from photo” or “needs DNA sequencing to determine to species level”. Sometimes, setting focus by exclusion is the best thing one can do !
The biggest part of challenge lies for me in the question of who is qualifying the information, a field that already today is an acknowledged weakness. If the information is getting dynamic (“living”) through an LLM process, this will get even harder because more experts will be required to permanently “curate” the auto-generated WIKI. If they don’t do, wrong or misleading information is the consequence, resulting in the opposite of what was intended.