First of, a disclaimer. I have no proof of this, and I am not an expert on AI-generated content or copyright protections. AI art programs have recently become highly controversial and many artists who upload images online take steps to prevent their art from being used to generate outputs in programs such as Stable Diffusion, Midjourney etc.
As a quick experiment, I tried to see if Stable Diffusion would be able to generate a recognisable image of Leucomonia bethia, a hawk moth with only a few observations on iNaturalist, and just a handful of image search results on the wider web. I chose this taxa as it has a fairly simple colour scheme (light grey forewings, dark grey hindwings) and a limited number of images that exist of it online.
Simply using the prompt ‘leucomonia bethia’ on Stable Diffusion creates a bunch of plant-like images. I repeated this at least five times, and every result looks like some sort of plant.
a typical example:
However, using the prompt ‘iNaturalist leucomonia bethia observation’, images have about a 50% chance to have an insectile appearance, even though the colour is always green, and this is often vaguely of lepidoptera with large wings. In the following example below, there is even a mockup of a watermark at the bottom of the top right image.
On the surface, it seems that adding the term ‘iNaturalist’ to a prompt results in an output closer to the actual taxa. I understand that iNat trains its own recognition algorithms and am happy to contribute to that, but wonder if AI programs could be abusing it.
Edit: I have no idea why the program insists the output must be predominantly green in either case. There are only two Leucomonia bethia images with any green in them online, and they are iNat observations.
Now, as the wikipedia page on stable diffusion notes, there are serious ethical and legal questions about doing this. Obviously there is no problem if an observation is CC0, but it is hard to see how it does not violate a CC-BY-NC-ND license for example, because it arguably violates every single provision of the license separately.
I don’t know if this is what prompted the question in there first place, but there were several obviously AI-generated beetle observations uploaded to inat last night (they were flagged as copyright violations for the time being). Fortunately, they were pretty obvious AI generations. That is almost certain to become more of a problem in the future given the lengths some duress users will apparently go to to avoid doing the actual assignment.
I wonder. As I understood it, within existing law in UK/US(?) even full copyright is not violated if you are not reproducing a substantial part of the original image? - people usually talk about 10% similarity, etc.
If a single photo is just one of 100s used to train a neural network, in most cases it won’t produce as an output anything remotely close to this 10% rule in relation to a single user.
Things get more complicated if you specify a single artist and try and generate a new Wildskyflower photo or whatever…but with regard to licenses within iNat, I’m not sure this comes in to play.
I could theoretically create my own iNat image generator just using the existing observations to create new images of taxa, and I don’t think it would abuse any kind of individual licence. Either way, you would likely not be able to tell who’s photos were used if there was sufficient training data for a taxon.
I led a team of researchers that used iNat snake images to train a neural net (for identification purposes, not generative), and we came to the same conclusion regarding license violation, namely that our use of images to train an AI did not constitute an unfair use because:
we were a non-commercial research organization
we did not anticipate that our use will have an effect upon the potential market
we did not host or redistribute images and did not retain copies of all images indefinitely
In addition, we reasoned that most photos on iNat do not have an “individual character”. Almost all of them are taken by amateurs (i.e. not professional photographers who earn a living from their photography) who did not make particular choices about the lighting, angle, or timing of the photo, and, as such, do not constitute “works” under Swiss copyright law (where our research team was based). This is difficult to determine on an image-by-image basis but the nature of both iNat and many “ID-request” social media groups (from which we also used images) is such that nearly all photos are taken for documentation purposes only. On iNat, using CC0 vs other license types is probably a good proxy for this—in my experience, users who have some knowledge of copyright tend to opt into the more restrictive licenses. To make our use clear, we explicitly spelled out in our Data Management Plan and in our articles that we fully intended for all photographers to retain their copyright and Intellectual Property Rights.
I certainly agree that that uploads of AI-generated images could become a problem, and that other users of iNat images for AI training purposes might not benefit from some of the “fair use” claims that we were able to make. However, my overall conclusion from researching the DMP for the project was that existing legal instruments are far from clear regarding their application to these kinds of things, and probably will require future judicial or legislative action to clarify.
Probably just exploring new technology, maybe testing what the CV will do with it. If they generated the images themselves, there could be an aspect of gaming/trolling the system but I imagine the driving force is most likely just curiosity.
You wouldn’t think people would post images of humans either yet the number of such obs posted every month is in the thousands – Never underestimate our own species’ capacity for lame humour, technical ineptness, and general misunderstanding!
It seems likely to me that “Leucomonia bethia” creates a green plant-looking ouput, “inaturalist observation” creates mostly animals, so the output of the combined prompt are green or plant-like animals. This tracks with the results I’m getting:
It would be surprising if iNat or GBIF weren’t scraped, to be honest. Putting the AI genie back into its bottle copyright-wise seems unrealistic to me when Stable Diffusion has already associated crested grebes with photographer watermarks/signature, for example:
I mean I surely make ‘particular choices’ about the lighting and angle when I try to figure out how to bend over a flower get a picture of the underside of the inflorescence in good sun without getting my shadow or my phones shadow in it. And even on my iphone camera I have 3 different physical cameras to choose from so I am surely making a ‘particular choice’ when I choose which one to use for a given photo. Quality subjective and is not relevant for whether something is art.
That said people upload CC licensed pictures in part for use in research so especially if you aren’t using ‘all rights reserved’ pictures probably there isn’t much problem for your particular application. And anyway people will do this whether or not it is legal and no one will stop them because 100+ year old laws are not at all adapted to present technology, for example:
So stealing is fine, so long as you never steal less than 10 things at a time?
I agree, this is subjective and I never felt 100% comfortable with it. This was language I was advised to use by a member of the Swiss Digital Law Center. I suppose there is case law with examples of instances where a judge or jury had to decide whether or not something was art, which led to this “standard”.
While I agree that your usage of iNat images for neural net training likely falls under fair use in this case (as the usage is ephemeral, does not involve long-term storage or reproduction, and is for educational use), I’m very suspicious of the assertion that most iNat photos are not creative works protected by copyright in the first place.
Switzerland is party to the Berne Convention, which, in addition to specifying that each signatory state must afford works from other states equal protection as works from their own nationals, also specifies minimum standards for the protection of works (e.g. a minimum term of protection of 25 years for photographs and applied art).
It also defines what constitutes a “work” (emphasis mine):
(1) The expression “literary and artistic works” shall include every production in the literary, scientific and artistic domain, whatever may be the mode or form of its expression, such as books, pamphlets and other writings; lectures, addresses, sermons and other works of the same nature; dramatic or dramatico-musical works; choreographic works and entertainments in dumb show; musical compositions with or without words; cinematographic works to which are assimilated works expressed by a process analogous to cinematography; works of drawing, painting, architecture, sculpture, engraving and lithography; photographic works to which are assimilated works expressed by a process analogous to photography; works of applied art; illustrations, maps, plans, sketches and three-dimensional works relative to geography, topography, architecture or science.
— The Berne Convention, Article 2 Paragraph 1
The threshold of originality you’re likely referring to when talking about “individual character” and “particular choices about the lighting, angle, or timing” does indeed exist, but as far as I know has mostly been used to argue that photos produced by continuous or automated photography (e.g. CCTV cameras, trail cams etc) aren’t protected.
As far as I know, in every case where a human is responsible for pressing the shutter, the resulting image is recognised under the Berne Convention as a photographic work, and afforded all the protections which come along with that.
I am not a lawyer, and usually I would say trust the opinion of your contact at the Swiss Digital Law Center, but I’m extremely suspicious about their conclusion and would personally go looking for some second opinions if I thought the usage didn’t clearly fall under fair use!