It is not the same thing at all.
In a sketch every detail in the image was intentionally put there by you. If you ask a generative AI model to draw something for you, it will invariably add all sorts of details of its own which you did not explicitly ask for and which you might not have seen, but which might influence IDers.
See the discussion in this thread: https://forum.inaturalist.org/t/use-of-ai-upscaling/52724/
I agree that it would be helpful to clarify that
Iāve tried to find this in official iNat documentation before and was unable to (though I might have missed it; a search of iNatās Help just now for āsketchā only returned this) when dealing with situations where observers had made sketches under various procedures including:
- Creating a sketch days/weeks after making an observation and
- Googling a description that they wrote of what they saw, and then sketching from the images Google returned.
These sketches ended up being for rare/unexpected species (not necessarily surprisingly) where having a correct ID is quite important.
I disagree that
Thereās no reason the rule/policy canāt be documented more clearly/citably so that users can act in good faith in accordance with it or that IDers canāt ask the observer/sketcher about their process to see if it aligns with the policy. In the cases above, I just asked the observers whether they made the sketches from life, and they described their method. Assuming good faith in this situation means working from a starting assumption that the person isnāt directly lying about how they made their sketch if they say that they made it from life. More broadly, plenty of users could submit a sketch made from memory in good faith if they donāt know about the policy but could also act in good faith regarding their sketching once they become aware of the policy (eg, making observations with sketches not made from life Casual, making future sketches from life, etc.).
True, valid point. Those are indeed very different.
The other conversation with AI upscale also illustrated the risk with this approach to get involuntarily influenced by the generated content, even as the observer.
On a side note, I have a question. Is starting a ācontinuing the discussion fromā¦ā considered a request to reopen the thread in question? Because once I borrowed a quote from an old thread to start a new thread, and was very upset and frustrated when it was later merged with that old thread. I hadnāt intended to ropen the old thread.
I donāt either, but realistically, the rules have been tightened. Try submitting a manuscript to a peer reviewed journal with graphs drawn using drafting tools instead of computer generated.
I mostly agree with you regarding good faith. (Out of curiosity, how would you make obvious drawings from memory casual? None of the DQAs really seems to fit)
And I do think that it is very hard to remember sufficient detail for a species ID for most taxa and that thereās risk of not misremembering things. However, I think that the drawing that can be made from memory are often times useful. For example, I saw a kingfisher fly over the local river recently. Basically just a deep blue bird-shaped blob. But in Germany, that would be sufficient for a species ID as there arenāt any other bright blue bird species.
Additionally, I see a lot of people (irl and online) describing certain features that if put in a drawing could be IDed to species.
Itās a shame, IMO. I like making stuff the traditional way. I think especially for maps, computer generated ones just lack soul and any sort of sense for aesthetics. Donāt think it helps science much in overcoming the ācold and sterileā stereotypes
I would downvote āEvidence of presenceā as drawings from memory arenāt an accepted form of evidence. āDate is accurateā might also be acceptable as the date of creation of evidence wouldnāt match the observation time. Or I could just delete the picture if it were my own observation.
I guess in your example of
I donāt see any difference in you drawing a kingfisher based on this memory versus just writing the memory in words here. I would consider the drawing the equivalent of an observation with a description saying āI saw a deep blue bird-shaped blob.ā I can evaluate both to the same degree - basically, I need to trust you that that is what you saw. The evidence is hearsay, not an independent document that I can evaluate. To me, the drawing would add nothing in terms of evidence, and Iād also recommend the observer submit to eBird, as they take this kind of evidence.
To make an analogy, if I were accepting evidence, would I consider a transcript of a conversation written down during the conversation the equivalent of someoneās memories of the conversation that they wrote down a week later? I would not - the transcript is much stronger evidence than summary written a week later. Thatās not to say that the summary someone wrote a week later is necessarily wrong or that the drawing is either. I can tell you perfectly accurately that I saw an eastern gray squirrel in my front yard yesterday. I could draw you a reasonable picture of the squirrel from memory. But the evidence for the squirrelās presence doesnāt meet iNatās standards for a Verifiable observation. If I want one of those Iāll need a pic, field sketch from life, or maybe an audio recording of the squirrel yelling at me if Iām feeling ambitious.
I feel this misframes my original post. I did not āask a generative AI model to draw somethingā there. That was not what that particular thread was about. That was about AI upscalers. My initial intention there was just to take an existing actual observation and to increase resolution/sharpen.
This thread would be a more relevant thread to link to wrt to your comment and @adrien_ferrandās original comment.
Text prompting a generative AI vs using an AI upscaler are very different kettles of fish in terms of user intent and in terms of input ( original image, not text ). Letās not conflate the two.
I also disagree with your point though fwiw. I have drawn stuff from memory and unintentionally hallucinated minor detail in much the same way a generative AI would :
e.g. this drawing of a glowworm larvae from Greece I uploaded vs the photo of one I found later :
And, following your logic re:intentionality - if I saw something in an AI generated image that I knew was incorrect, I could then in theory just prompt the tool to remove said artefact.
I attempted this out of curiosity, using the text description from my above drawing upload as a prompt :
AI drawing of glowworm larva by chatGPT after 3 prompts:
In theory, I could carry on prompting to get it more in line with the description I have on my obs. I am not sure my drawing is much better than the AI drawing.
Worse, arguably.
Personally, I donāt really think drawings should count as hard evidence though.
( and I donāt think wholly AI generated images of things people saw based on text descriptions should be hard evidence either)
Upscaling an existing photograph as I did in the other thread referenced, is more complex - as there will always be a spectrum of how much you upscale / use AI to affect the original photoā¦and doubtless different tools use AI for upscaling in very different ways ( Topaz.ai ā Krea.ai ).
EDIT - perhaps in your comment @spiphany you are only talking about drawing in situ, as @tiwane states is a dividing line here ā¦ iād missed @tiwaneās comment originally and didnāt realise that was a consideration when it comes to drawing uploads on iNat.
Still, one could in theory use an AI to generate an image in situ, with multiple prompts to refineā¦
I linked to that thread because there was subsequent discussion there about gen AI adding additional details not found in the original image. This is the problem that I wanted to point out. It applies regardless of whether one is using generative AI which adds details in the process of improving an image (as clearly happened in the bird example discussed in the thread) or using generative AI to create an image. The thread also included a fair amount of discussion about sketches and whether or not these are comparable to creating images using generative AI. Both aspects of the discussion are relevant to the post I was responding to.
Of course it is possible that humans creating a drawing may add details that they thought they saw but were not actually present. This is a known phenomenon e.g. in witness accounts. And it is something that IDers need to take into account when assessing the reliability of drawings on iNat.
But there is nonetheless a significant difference between humans intentionally adding details they thought they saw (but may not actually have seen) and generative AI adding details that nobody told them to add. An AI image is always going to include a lot more detail than that provided by the prompt. It is virtually impossible for the asker to control every single aspect of the final photo, and it is also unlikely that they will consciously notice and examine every detail and weigh it against what they remember seeing and approve or reject it. They are more likely to focus on the gestalt and on particular prominent details that they noticed. But often it is less obvious details that may be crucial for ID. And if these details were added by the AI, they are, essentially, evidence constructed out of thin air, because they do not even have a basis in what the observer remembers having seen.
I still donāt agree that that is reason enough to discard all drawings from memory as being insufficiently strong evidence. They are weaker evidence, yes, but in a lot of cases they are sufficient.
In any case, I donāt think this should really be discussed here (either moved to a new thread or just agree to disagree) as the thread here is about images without any evidence and is marked as solved.
But in practice, there will always be details that you didnāt account for. In the larva, did you prompt the AI to add segmented legs and antennae with exactly that many segments? Did you tell it to add the two claws on each leg? What about the texture?
And if you say āno, but thatās how it looked likeā: How can you be sure? Or are you just agreeing because these features are plausible? No field notes will reach the same amount of detail that an AI-generated image will give you. In practice, an additive method of creating an image will always be a more accurate depiction of what you remember than a subtractive one.
I was taught that for scientific data, the results should never be more detailed than what it is based on. Seems to apply here too, IMO.
Similarly, this should probably be moved to a new thread (or one of the existing AI-image threads).
Sure. But nevertheless, I feel your original comment misframed the thread, as you premised the link with :
Which I found problematicā¦but nvm.
Upscaling remains a more complex issue, as upscaling is or will be inbuilt into many photo editors and the degree to which one uses AI will be a spectrum.
I just think its important these two issues are not conflated.
I still want to map out this issue across apps, but here is the UI for Krea for example :
Agreed. With most current generative AI image tools Iāve tested at least.
I imagine this not to be the case in the near future though. I am yet to come across a good AI drawing tool. AI image generators arenāt really comparable to a person sketching. I imagine something more comparable will come on the market soon if not out there already which could be more additive than subtractive, as you mention @eyekosaeder.
I linked to the chat directly so you can see.
Here it is again.
No, I premised the link in a new paragraph stating āSee the discussion in this threadā. I made no claims about the purpose of the thread other than the implication that I was linking to it because the discussion included content relevant to the issue of generative AI inventing details in images.
I apologize if I did not explicitly indicate āthere is some additional relevant discussion on this thread on a somewhat different topicā but this did not seem to me to be necessary. The point of providing a link (rather than quoting) is to allow people to read the linked material themselves in its entire context and decide whether they feel it offers any insights. Would you have felt less misrepresented if I had not linked to the thread as a whole but to some of the individual posts that discussed AI-generated details?
It seems to me that your objection is perhaps not so much because I misrepresented the discussion in the linked thread, but because you are unhappy about the tenor of that discussion, which was fairly critical of your example of the use of generative AI for photo āupscalingā, and because linking it in the context of this discussion emphasizes some of the positions taken by other participants in the thread which you disagree with.
You also seem to feel that your use of generative AI for photo editing is fundamentally different from the use of generative AI for image creation. I disagree. I think they are analogous, and it is not misrepresenting the thread to point out its relevance here. Multiple users in your thread considered that the use of generative AI in the example you provided resulted in a invention of details and changes to the image in ways that are problematic for iNat; these are the exact same issues that I would expect to see in an image that has been completely generated by AI rather than just one that has been edited using generative AI.
Just to clarify, you mean something like downvoting āEvidence of Organismā in the DQA, right? That would be acceptable, although itās a bit of a weird use case. A lot of people mistakenly think that checking āCaptive/Cultivatedā is a good way to make any problematic observation Casual, but it should not be used that way.
I think Iāll simply donāt upload them in iNaturalist, in the end. Iāll keep them in notes and just have the place holder without evidence to remind me what I saw for future reference.
And the glowworm larva shows it clearly:
Is this a glowworm larva or a myriapod? I count seven pairs of legs, which immediately rules out any kind of insect, and no clear distinction between thoracic and abdominal segments. Whereas the drawing accurately distinguishes between thoracic and abdominal segments, and makes no claims about legs. The drawing also more accurately reflects the near-absence of antennae. In short, the drawing is clearly much better in terms of identifying the specimen.
No.
Itās simply because I felt you mis-framed the thread and in turn conflated things which are to my mind distinct issues - use of AI upscaling tools and use of generative AI image-making tools.
As stated. It wasnāt an issue of you misrepresenting me personally.
However, now suggesting I am unhappy about that discussion and feel threatened at being linked to it wholly misrepresents me. I see no serious disagreement on my part in that thread or indication of me taking umbrage at anyoneās ātenorā.
Indeed I posted that image up because I myself was critical of the potential of the tools.
Please reread my opening words of the thread, as you seem to have mistaken my intention :
I shall leave the discussion hereā¦as I am repeating myself, this is wildly tangential to the OP and this feels like itās stepping into ad hominem arguments which I have no wish to partake in.
But one final time fwiwā¦
One may add or subtract key detail on your existing observation.
The other forms a totally new observation from scratch.
One will already be widespread in usage to different degrees and is deeply problematic
The other is likely very rare in upload and requires deliberate malice on the users part.
Yes, there is overlap in that both may generate incorrect aspects to observations.
But they are ultimately quite different issues and conversations to have.
That was all I was trying to say.
Please also read the opening to the other thread before you further build on this conflation of topics and misframing of my intent, @jasonhernandez74.
Multiple users?
I was the user who was raising this issue in the first placeā¦ 
Anyway yes, signing off from this thread to focus on enjoying the real world.
Have a great weekend iNatting yāall!
Whatever may have been your original intent, in this thread, you claimed that your drawing was āworse, arguably,ā than the AI image. I explained why it wasnāt. Your drawing did not hallucinate details like extra pairs of legs. Also, as I read through your discussion with ChatGPT, you tried to correct it regarding the three larger segments, and it did not make that correction; whereas your drawing already had that part correct.
The point you were trying to make in this thread, if I understood you correctly, was that drawings shouldnāt count because they are no better than images generated by AI:
But your example actually illustrated the contrary of what you were trying to say. Even after three rounds of refinement, the AI image was still further from the truth than the drawing that you think is so bad.
To be honest, after reading carefully the whole thread, my conclusion is that nor drawings nor AI illustrations should be a valid evidence at all, although drawings are the less worse option. The rest is nitpicking.
Nowā¦ thereās a reasonable consensus that drawings are not harming the validity of the iNat database due to its small size and sensible use, so let them be.