AI generated image passed the Computer Vision test

Why would someone target inat like this tho?

I want to develop that wetland.
You say there are leopard toads there.
I will do - what I can, what I have to - so I can force my development thru!
But the toads won this round.

https://groundup.org.za/article/good-news-for-leopard-toads-anonymous-donor-buys-crucial-habitat/

This is exactly what I was thinking as well… The CV doesn’t have a possible output for “fake! this is nothing!” If you ask a program “what organism does this look like?” and its only outputs are limited to suggesting which organism it looks like, then it’s going to suggest a name. It could be an impressionist painting of a cheesecake covered in wire nuts, and it’s still going to throw out a name, because that’s all it’s programmed to do. If anything, I’d be more surprised if it didn’t suggest a mint when you give it a fake image meant to resemble a mint. What else would one expect it to do?

I think flagging the photo and marking the observation as No Evidence of Organism is the best choice.

Very true.
Wouldn’t it be a good start if CV was offering

as output for pictures that don’t show an interpretable organism? If CV sees no evidence, it should be allowed so speak open and make people think twice to what they may or may not consent. If there is no obvious evidence, it should be actively stated!
Maybe even go a step further and train CV to identify obvious characteristica of AI generated pictures (like for any other species, maybe based on mathematical pixel pattern, artificial choice of color distribution etc) and propose then a species “organismus falsus” ?
It would make me feel better if the robots could also take their share in data protection. If we need them, then use them at least in what they can do best: Pattern recognition and data screening.

This is not a good idea. It would require humans to first learn how to reliably identify AI generated images, since that is the only way to train the CV. This is a waste of valuable identifier time and effort, and in the long run could actually be counter-productive. The danger is that a feedback loop might be created whereby AI generators start leveraging human-created training sets to improve their output (like anti-biotics breeding super-resistant strains of bacteria).

Because some very powerfull people believe right now that literally only the sky is the limit and nobody knows what foolishness may come next.
They rather want to settle on mars than preserve habitat and society on earth.
And if they don’t get as they want (and the chances are high), they simply may want to destroy everything they believe was keeping them apart from fulfilling their weird desire.
Call it malevolence. Call it bad education. Call it greed to just demonstrate power for its own. Call it madness.
Unfortunately, people are rarely logic.
Unfortunately, machines, however intelligent, will never be people as they will never physically participate in the ecosystem like a living thing does.
However, preserving things effectively must alway start small, not by reaching out to the sky.

And the thing is, all of the fake plants shown in that blog are identifiable to genus (all Hosta except for one Iris). That’s part of what makes them plausible to some people.

I understand your concern.
In consequence one should then stop wasting time in developing better antibiotics.
And stop producing them and let nature follow its own business…
Why do we need health insurance if everything it does is setting bias to natural selection? Such a waste of effort…
The real joke here is that this “bacteria” is fully man-made and not even carbon based.
And it is dangerous only but everywhere where it interfers in our social interaction because there is no immune system.
So easy to prevent.
But so high temptation to (ab)use.
How to manage with so many human intelligences at the rifle trigger?

I don’t think this is how the CV works. It can’t recognize an image without an organism or a faked image because it doesn’t actually know anything about the content of the images. It only looks for visual resemblance. No matter what image you present it with, there will always be some photo in its training set where there is some visual similarity, even if only very distant. Images are only patterns of color, after all.

It already notes when its confidence level is not high and there are no good matches. Perhaps it could be set to only make broad suggestions or no suggestions at all in such cases (or better yet, give a message with some common reasons for this – e.g. suggest cropping the photo and uploading it again).

However, I imagine training it on observations that have been marked as “no evidence of an organism” so that it could suggest that there is no organism in the photo would create an awful lot of false negatives, because a photo of a landscape where the bird has escaped is going to be very visually similar to a photo of a landscape where the bird is a small spot of just a few pixels in the sky.

Training it to flag photos that don’t match anything in the dataset as possible AI-generated also seems likely to backfire. The CV is often able to suggest fairly good broad IDs when presented with photos of species it has not learned because of visual similarity to closely related species that it has learned. So “I don’t recognize this” may have various legitimate causes, the most important one being that the CV’s training set has major gaps.

The problem is not with the antibiotics, but with their over-use. If we didn’t keep over-using them, we could afford to spend far less on trying to improve them. It’s no coincidence that our species suffers from an unusually broad range of virulent pathogens: we create environments for ourselves which provide ideal conditions in which such pathogens can evolve and thrive. Up until very recently, this has made us victims our own success. However, modern medicine has now swung the game of life decisively in our favour, as any graph of human population growth will testify. Needless to say, winning this particular battle does not mean that Mother Nature won’t still win the war…

We don’t need anything like an immune system on iNaturalist. AI-generators do not pose any real threat. As has already been pointed out, the entire premise of this thread is a red herring. The only real threat comes from our own users. The Terms of Use and Guidance make it quite clear what sort of content is acceptable:

Photos or sounds attached to observations should include evidence of the actual organism at the time of the observation, observed by the user who is uploading the observation. Media used in your iNaturalist observations should represent your own experiences, not just examples of something similar to what you saw. Please do not upload photos you found elsewhere

The only way unacceptable content like AI-generated images can get into the system is if users deliberately upload them. In the unlikely event that this became a significant problem, we should focus our efforts on educating users about their responsibilities and making sure they understand the potential consequences of their actions (e.g. suspension of their account). Reactionary measures like attempting to train the CV to identify unacceptable content are counter-productive and don’t address the real source of the problem.

I agree.
There are already today lots of users in here, posting and waiving through minced meat pictures repeatedly with no consequence while at the same time they dare to review other people’s input.
And you say that this

is affordable even in a situation where automatition may act as a multiplier.
I am pretty sure the identifiers will rather sooner than later develop their own oppinion.

I know nothing of plants. Or at least I thought I did. I took one glance and thought, that’s a stinging nettle, isn’t it? Quite pleased ;)

Valid concern! Sounds like penetration testing in cyber security.

Not yet, but makes you think it won’t be able to in the next few years? I assume that as other AIs train on iNat images they will get better, and know what characteristics need to be included. AI will have access to every scientific paper written about biota. I think the potential is there for it to produce better identifications than humans.

I must strongly disagree. That’s much too surfacial for me.
From a perspective of individual identification correctnes, future AI’s may be statistically more accurate than humans. Same as chess- or go-computers, which win world champioships. Who cares, when playing chess?
In the core of the act of a good identification there is an intention. Our human purpose is knowledge and preservation. What’s the intention of identification by AI?
Our control/corrective action is peer-group review. What’s the control/corrective action mechanism of AI?
Besides some old foundational AI developers, who meanwhile sent the whole sh… to devil because of their severe concerns, those industrials who are still active in that field stay quiet while proactively undermining all legal regulation by political measures and bribery.
Sorry, its only better to the logic of a chess computer.

AIs know nothing about the images that they are imitating or suggesting IDs for. They analyze images as patterns of color. They don’t know what a leaf is, or a leg. iNat’s CV consistently makes certain types of mistakes that are the result of the fact that it doesn’t have concepts for anything and it doesn’t “understand” what it is seeing in anything resembling the way a human would. They can’t know “what characteristics need to be included” because they don’t know what objects are in the first place.

For example, there’s a recent article that explores how an AI interprets images compared to humans by giving them a classification task (asking them to make decisions about which images are part of a set – i.e., making semantic judgments). The differences are suggestive: https://www.nature.com/articles/s42256-025-01041-7

For an observer: to get a correct identification in a timely manner.

That’s maybe your intention using such tool.
And we will have to use such tools SOMEHOW, no doubt.
But the AI thing itself has no more intention than a tabletop calculator.
What I say is the following: People give it the visibility and influence of a 24/7 superuser. People may want BELIEVE its creating knowledge because its so convenient. But its inherent corrective routines, if any, are unclear and it has the responsibility of a stone. Who will handle the collateral consequences but humans?
If handled like introducing AI, i can only say “good night!”. Do human users want to carry the consequences for decisions they haben’t been involved in for the sake of convenience of decision makers? I do not want letting me force into a new workflow for which nobody today is answering my very valid questions. And that to the risk of the entire project.

Wait, is that why Inat keeps suggesting fishes and mammal ids for my insect observations?