Many artists pointed out specific issues with the image: typical AI artefacts that come from image generation that a trained eye can easily see, but harder for a layperson to understand the significance.
Authors remained unconvinced: needed better proof.
Some artists tried to appeal to reason with AI detection software.
They were not listened to, but I understand: AI detectors are NOT reliable, and they have their own issues (often they train their OWN AI on the input images! ☹)
Mark Lawrence graciously asked the artist Shaun Mauss and the the author Mathew Prindle for some simple proof that it isn't AI.
They were given reasonable time to get home and provide it.
I dug into the psd file, and saw what I expected to see: the images had been copied and pasted from elsewhere, not drawn in the PSD (as the artist had claimed).
This discrepancy is, to me, the most repeatable "AI detection" methodology.
@Zn2plusC (who was cooking Lasagna at the time) discovered that some of the layers were photoshop "smart objects", and the filenames literally had prompts in them, which happens in MidJourney.
And finally, @TheBrianPenny put two and two together and realized "Hey, we can literally find the generation in the MidJourney Logs using the filenames".
So they did. Every detail piece in the image was AI [re]generated.
A minority of authors & artists feel there is some bad blood going around: feeling victimized, bad vibes, a witch hunt.
But a "witch hunt" implies innocence: this was an accurate accusation.
And I feel hopeful: it's clear to me now that evidence is VERY hard to fake. #TeamHuman
I no longer worry that artists have to "document every step" of their process.
It's good to do for social media and other reasons, but not necessary.
Because the proof is in the pudding: the .clip, .psd, .kra file, and a statement of their process.
If someone says they paint or draw, the file will always show evidence of their process.
If someone says they composite blender renders and do paint overs, there will always be evidence.
If someone says they collage and photobash, there's evidence.
If someone LIES, we CAN TELL
AI detecting software doesn't work.
I believe in community building and trust networks across domains. Getting second opinions. Because poetically, human connection is the best antidote to AI.
AI harms small creatives. Honest folks who just want to write, draw or make things.
So we should be kind to each other, approach each other with empathy, and listen.
The victims of AI are the Authors who are scammed, the Cover Artists who were cheated in a competition or lost out on a commission, and the readers who endure endless sludge
The house Judiciary committee had their first hearing of many on Copyright and AI. Photography and music were represented (3 people), as well as 2 lobbyists for AI.
AI companies are using every argument they can to dodge these 3 pillars of copyright. 2/5 (reupload w/ sub)
There are a lot of highlights, like allusions the the pittance that Spotify pays out to artists, and this little dig that Ms. Ross implies that Sam Altman has more than enough money to compensate artists with how much money their company is worth. 3/5
💸COMMERCIAL Machine Learning (AI) is rotten to the core in terms of respect for Copyright.
👁🤖I'm trying to track down the origins of datasets for things like motion tracking, object detection & general computer vision.
They almost ALWAYS lead to unlicensed Flickr images😠🧵
I can't find a Machine Learning product that properly licenses training data. If you know of any, send them to me! (This isn't about Image Generators, this is everything).
Every layer of dataset abstraction/annotation just further abstracts away the original rightsholders.
2/9🧵
I wouldn't mind if this stuff was done for Research to discover scope and requirements, then license data.
But people are developing huge products requiring mass compute to churn through human data we generated and selling it back to us. Consent should be a factor here.
3/9🧵