Is AI Stealing My Art?

[parkdalegardener]

The days of scraping of the net for training data like was done for the LAION or CLIP datasets is over. Privacy concerns and data acquisition laws world wide are taking care of that. NSFW imaging raises ethical and legal concerns. IP is being heavily defended. Those scrapes were finished a couple of years ago. They include no recent data. Time was needed to use the accumulated date to train the weights we are using today.

Instead you are training the dataset directly. Willingly. For free with no ask of recompense. "Big Data" wants you. Every day hundreds of thousands of people are using online chatbots and AI art generators. These sites are free in some cases and cheap otherwise. These sites use your input to train their models and you pay them to do it. Can you say Midjourney V4?  What about all those sites that suddenly offered to protect your art for a small fee; after deciding that everything on the site was up for grabs to train their pay text to speech generator. Hiya Deviant. Sites that encouraged you to display your AI art along with the prompts to show off and share with each other; and then used every one of those images to train their new for pay model and then removed everything except output from their proprietary model. Right Lexica?

The more folk that check out ChatGP, the more it learns. For Mickeysoft; who just dumped a boat load more cash into OpenAi last week. OpenAI. The company that makes ChatGP and Dall-E 2. You can expect to see MSOffice incorporating a chatbot similar to the annoying paperclip of years gone by to assist with MSWord docs and a Dalle type integration with MSPaint applications. Adobe has incorporated a for pay stable diffusion plugin and is now a part of Photoshop and is coming to After Effect and Premiere.

Everyone is getting on the bandwagon. Take free open source code and put it behind a pay wall. It has happened for years and will happen in perpetuity. Once this is gobbled by "Big Data" it is gone from the public view and you have no idea how it is being used, trained, deployed, or what have you.

This stuff is here. It isn't going away. The 2.x models may be less capable of "A naked portrait of (insert celebrity) in the style of (insert artist) and (insert animation or other companies IP) standing in a temple." than the 1.x models but they are still out there and people are still using them. The results don't have that overpolished smooth and unreal look that the newer models have.

[parkdalegardener]

Every time a person uses this stuff there is a risk. Are you uploading pictures of yourself or a loved one to a third party site to make you a cool AI Avatar? You just gave someone your likeness and a name. Maybe not a real name but does it matter? Your likeness in now in a database that is out of your control. Someone other than you now has your likeness they can manipulate as they wish and in a way no Photoshopped images ever could. Perfect? No. Convincing? Probably.

Most models that you find online are in a .ckpt (checkpoint) file format. Remember checkpoints? The save game of diffusion models. Well these files can contain malicious code in with the actual data. Most virus scanners have no idea what these types of files are. Python files like these are "pickled" and some sites have a pickle scanner check them. Some do not. You do not want any .ckpt files. Period. None.

Instead you want "safetensor" files. Safetensors are data, just the data, and nothing else. So help me dog. No executable code and only the weight data.

"And the beat goes on" to quote Sonny and Cher.

[parkdalegardener]

So; is AI stealing my art? Yes; every time you use an online generation service. Free or otherwise. No; if you think your art is part of some collage made by the AI or such thing and your art is somehow hiding there ready to come out at the drop of a prompt. Maybe; if you want to fit data yourself. It will never be as well integrated as something that was trained from the beginning with your data as part of the set.
 
Is AI going to steal my art? No; if it isn't organized in a way the training weights can understand such as image/text description pairs when crawling the net. Maybe/Probably, if you are using an online generation service. You are helping train their proprietary models.

Hiding one's head in the sand won't make it go away. Hopefully a simple(?) explanation can help make sense of the coming of AI "art" scene.

[parkdalegardener]

Some other things to note. Every AI generated image contains a watermark that other AI's see. This is to keep that feedback loop in the training I mentioned earlier from happening. It cannot be removed after generation. AI should not train AI on AI. Introduces bias and error to the weights. That's why you are training ChatGP and it isn't writing code (shudder) to train itself.

There is also the NSFW filter. This, like the AI gen watermark; is also enabled by default. Online generation services love this one. Pay extra for nekked anime gurrls. Only Fans for fake boobies and the like.

Both restrictions can be removed locally and I'll leave it at that.

[thelufias]

I'm a PSP (Paint Shop Pro) guy....Have been when JASC put it out to market and "J's" sister sent me her version 2 since she didn't use it anymore...  32 Discs to load...but it was worth it...and I've been updating PSP every year....

Some folks have just got to realize AI is here to stay no matter the argument.... This is what my 2022 version has:

 Paint Shop Pro 22

[parkdalegardener]

Exactly.

[thelufias]

Speaking of which....Just got done visiting one of my favorite places ~ SpaceWeather.Com....

They've put this up on their main page:

This is an AI Free Zone! Text created by ChatGPT and other Large Language Models is spreading rapidly across the Internet. It's well-written, artificial, frequently inaccurate. If you find a mistake on Spaceweather.com, rest assured it was made by a real human being.

[parkdalegardener]

Hehehe. That's great. No doubt in response to the Sports Illustrated foolishness last week where they were found to be publishing such articles.

I wrote the original thread so long ago I should update some things. Unlike a lot of large language models that were frozen once training started. Means that a lot of information may be out of date leading to errors as well.

I don't do X (Twitter) but Elon Mush; I mean Musk, has rolled out his AI hate bot. I mean Chat Bot. His claim is that with his new AI, GROK; he can keep LLMs (large language models) current. He trained it on 2 months of Twitter posts. He feels that letting it out in the world will provide edgy commentary and humor on X. Really? Bet a whole heck of a lot of his own drivel was used in the training set. After all; the two months of training data are on Twitter posts only made after he purchased the platform.

[thelufias]

Like you I do not associate with X in any way whatsoever.  Didn't like Twitter to begin with.  Started posting some of our winners up there but wasn't worth it.  To much other useless crap on it. 

[AmirA]

AI....Is it good or bad?
Well I am a fine artist that paints on canvas. We share our work everywhere to show, etc. As so many steals our work, I do add watermarks to my work or downscale the pixels for zooming in to make the image not usable or low reso.
Yes its here to stay and no we cannot get past it, but you still see it everywhere. Especially on Facebook with all the AI images going around. Looking close to the details of hands and eyes is freaky.

But as an artist you can use AI as a tool. But then some misuse it for their art, etc. Not sure if AI is used for renders here. For me I wont use AI for paintings or renders. I try to stay unique in my art styles.