Well, if everyone decided to use a successful AI poisoning mechanism ... one would assume that would kill off AI.
In theory someone who wanted to use AI as a tool in their operations - say a manufacturing company using performance data from "things" they owned and operated - could enable and support their own internal AI activities. A wider input from all machines of the same type operated by many clients might be better analysis in some cases but perhaps not critically so.
I'm not sure how a photographer could successfully achieve a similar result.
It might work for some but not necessarily all. Product or portrait specialists for example.
But I'm not sure it would be great for all.
That said, if it did not exist at all it would not much matter I suppose.
Assume he is a wedding photographer and he runs his own website (so no copyright terms and conditions from Flickr and alike to adhere to) and is showcasing images from a wedding of John and Jane, who granted him the right to do so, they might possibly not want to serve as an input for generative AI tools learning algorithms.
John Friend has a point here, arms race. The poison tool is in research state currently, if I understood right. Maybe, if a tool to poison image files exist in the future, it could be invoked by the "Open with application" function of export recipies. ImageMagick, I am looking at you.
I don't think a raw converter and library is necessarily the answer to all image related questions. Another example is sky replacement (stay polite if somebody wants to beat me for this :-))
I hope you realize that data poisoning in images like this (if it's real) that is undetectable by humans is just an arms race. If this is used widely, then the AI models will just be modified to detect its presence and see the image like humans see it (ignoring the poisoning). Maybe it's has short term value (if you're trying to ruin AI training), but probably not long term value.
Unfortunately abiding by the robots.txt file is completely honor-based. It’s known they were ignored by these LLM models before. They claim to honor them now but I personally put no faith in that.
Well, if everyone decided to use a successful AI poisoning mechanism ... one would assume that would kill off AI.
In theory someone who wanted to use AI as a tool in their operations - say a manufacturing company using performance data from "things" they owned and operated - could enable and support their own internal AI activities. A wider input from all machines of the same type operated by many clients might be better analysis in some cases but perhaps not critically so.
I'm not sure how a photographer could successfully achieve a similar result.
It might work for some but not necessarily all. Product or portrait specialists for example.
But I'm not sure it would be great for all.
That said, if it did not exist at all it would not much matter I suppose.
Assume he is a wedding photographer and he runs his own website (so no copyright terms and conditions from Flickr and alike to adhere to) and is showcasing images from a wedding of John and Jane, who granted him the right to do so, they might possibly not want to serve as an input for generative AI tools learning algorithms.
John Friend has a point here, arms race. The poison tool is in research state currently, if I understood right. Maybe, if a tool to poison image files exist in the future, it could be invoked by the "Open with application" function of export recipies. ImageMagick, I am looking at you.
I don't think a raw converter and library is necessarily the answer to all image related questions. Another example is sky replacement (stay polite if somebody wants to beat me for this :-))
Jon,
Does that mean you will never use any tool that has any association with AI?
I hope you realize that data poisoning in images like this (if it's real) that is undetectable by humans is just an arms race. If this is used widely, then the AI models will just be modified to detect its presence and see the image like humans see it (ignoring the poisoning). Maybe it's has short term value (if you're trying to ruin AI training), but probably not long term value.
Unfortunately abiding by the robots.txt file is completely honor-based. It’s known they were ignored by these LLM models before. They claim to honor them now but I personally put no faith in that.