There is no tool being handed period, these rulings already exist as case precedent. People are always trying to make a buck and it's easier to do that if the can shortcut the process and use other peoples works. The commercialization aspect is what crush's any attempt at justifying using copyrighted material without a license.
I even linked the legal framework of Fair Use doctrine, it's very easy to understand the four point test the courts have created. Of all the aspects, purpose and economic harm are weighed the most. Was the unlicensed use for commercial or educational use? Did the unlicensed use have a negative economic impact to the copyright holder? If the answer is "commercial" and "yes" then it's a slam dunk case.
'Fair Use' (and copyright in general) deals with
distribution of copies. Training a NN is not distribution of copies. If the
training dataset is distributed and contained copyrighted works then there's an issue, but trying to use copyright to fight distribution of not-even-close-to-copies is where you end up far away from all existing case law.
You may argue that the initial gathering of publicly available images and creating of even an internal and non-distributed training dataset may be argued as not covered under fair use, but there is actually case law in opposition to this, e.g.
Kelly v. Arriba Soft Corp., where it was upheld that a search engine can download entire full-resolution images and index them (and even distribute thumbnails of them) without infringing copyright.
::EDIT:: Oh, and "commercial" has
absolutely noting whatsoever to do with fair use. You can be non-commercially infringing (e.g. if you record a movie onto VHS and then hand a copy to a friend, that's infringement) or commercially non-infringing (e.g. the thumbnail example above).
The entire "but humans can blah" is also a dead argument as there exists precedent on transformative and derivative works. If a human reads a book and memorized every sentence, then later writes those sentences and recreates the work to sell, the human is still guilty of copyright violation. Who and what is doing the copying and distribution is irrelevant, only that copywrited material was copied and distributed without a valid license.
Despite being a straw man, this is also wrong: if a human reads a copyrighted work and creates a different work, there is no copyright violation. JK Rowling cannot sue you for writing any story about wizards even if you read Harry Potter beforehand. Likewise, Disney cannot sue you for copyright infringement for drawing a mouse in any art style, or for drawing your own character in a 'Disney style', but solely for drawing specifically a Disney character. NNs are not photobashing or cut-and-pasting sentences and paragraphs from some archive of source images or texts. That is fundamentally not how they function: you can download an entire trained NN and comb through them would and not find a single image or line of text.
I might be whispering in the wind here, but the principle of Fair Use doesn’t actually exist outside the US legal system.
'Fair dealing' is what most other countries refer to that concept as under Copyright legislation. The US is the odd one out.
::EDIT:: Any before someone trots out the lazy "AI BRO" argument in lieu of thinking: my opinion is that the current wave of MLNNs has very limited application, and the 'AI boom' will end once everyone figures out nobody actually
wants these LLMs, and moreso that nobody wants to pay for using them. Large Language Models have the least applicability, at best producing the crappest tier of marketing copy and only with extensive copyediting, and a very limited usage in turning correctly formatted and structured pseudocode into actual code (and turning poor pseudocode into crap code). Image processing models have broad applicability, but did so before the recent 'AI boom' and have been inactive use for many years anyway (e.g. multi-exposure and multi-lens image synthesis in smartphones). Image generation has some utility for artists (as Photoshop did, and more generally as photography did), but the idea that somehow it will 'end art' is about as valid as for every single other time it has been claimed by a new analog or digital tool.
I am far more concerned about the impact of the corporate manufactured panic (and subsequent legislative capture) over MLNN tools than I am over the MLNN tools themselves.
Facebook
Twitter
Instagram