Machine learning doesn’t retain an exact copy either. Just how on earth do you think can a model trained on terabytes of data be only a few gigabytes in side, yet contain “exact copies” of everything? If “AI” could function as a compression algorithm, it’d definitely be used as one. But it can’t, so it isn’t.

Machine learning can definitely re-create certain things really closely, but to do it well, it generally requires a lot of repeats in the training set. Which, granted, is a big problem that exists right now, and which people are trying to solve. But even right now, if you want an “exact” re-creation of something, cherry picking is almost always necessary, since (unsurprisingly) ML systems have a tendency to create things that have not been seen before.

Here’s an image from an article claiming that machine learning image generators plagiarize things.

However, if you take a second to look at the image, you’ll see that the prompters literally ask for screencaps of specific movies with specific actors, etc. and even then the resulting images aren’t one-to-one copies. It doesn’t take long to spot differences, like different lighting, slightly different poses, different backgrounds, etc.

If you got ahold of a human artist specializing in photoreal drawings and asked them to re-create a specific part of a movie they’ve seen a couple dozen or hundred times, they’d most likely produce something remarkably similar in accuracy. Very similar to what machine learning images generators are capable of at the moment.

Expect for all the cases when humans do exactly that.

A lot of learning is, really, little more than memorization: spelling of words, mathematical formulas, physical constants, etc. But, of course, those are pretty small, so they don’t count?

Then there’s things like sayings, which are entire phrases that only really work if they’re repeated verbatim. You sure can deliver the same idea using different words, but it’s not the same saying at that point.

To make a cover of a song, for example, you have to memorize the lyrics and melody of the original, exactly, to be able to re-create it. If you want to make that cover in the style of some other artist, you, obviously, have to learn their style: that is, analyze and memorize what makes that style unique. (e.g. C418 - Haggstrom, but it’s composed by John Williams)

Sometimes the artists don’t even realize they’re doing exactly that, so we end up with with “subconscious plagiarism” cases, e.g. Bright Tunes Music v. Harrisongs Music.

Some people, like Stephen Wiltshire, are very good at memorizing and replicating certain things; way better than you, I, or even current machine learning systems. And for that they’re praised.

It’s called “machine learning”, not “AI”, and it’s called that for a reason.

“AI” models are, essentially, solvers for mathematical system that we, humans, cannot describe and create solvers for ourselves, due to their complexity.

For example, a calculator for pure numbers is a pretty simple device all the logic of which can be designed by a human directly. For the device to be useful, however, the creator will have to analyze mathematical works of other people (to figure out how math works to begin with) and to test their creation against them. That is, they’d run formulas derived and solved by other people to verify that the results are correct.

With “AI” instead of designing all the logic manually, we create a system which can end up in a number of finite, yet still near infinite states, each of which defines behavior different from the other. By slowly tuning the model using existing data and checking its performance we (ideally) end up with a solver for some incredibly complex system. Such as languages or images.

If we were training a regular calculator this way, we might feed it things like “2+2=4”, “3x3=9”, “10/5=2”, etc.

If, after we’re done, the model can only solve those three expressions - we have failed. The model didn’t learn the mathematical system, it just memorized the examples. That’s called overfitting and that’s what every single “AI” company in the world is trying to avoid. (And to do so, they need a lot of diverse data)

Of course, if instead of those expressions the training set consisted of Portrait of Dora Maar, Mona Lisa, and Girl with a Pearl Earring, the model would only generate those tree paintings.

However, if the training was successful, we can ask the model to solve 3x10/5+2 - an expression it has never seen before - and it’d give us the correct result - 8. Or, in case of paintings, if we ask for a “Portrait of Mona List with a Pearl Earring” it would give us a brand new image that contains elements and styles of the thee paintings from the training set merged into a new one.

Of course the architecture of a machine learning model and the architecture of the human brain doesn’t match, but the things both can do are quite similar. Creating new works based on existing ones is not, by any means, a new invention. Here’s a picture that merges elements of “Fear and Loathing in Las Vegas” and “My Little Pony”, for example.

The major difference is that skills and knowledge of individual humans necessary to do things like that cannot be transferred or lend to other people. Machine learning models can be. This tech is probably the closest we’ll even be to being able to shake skills and knowledge “telepathically”, so to say.

Why are you entitled to other peoples work?

Do you really think you’ve never consumed data that was not intended for you? Never used copyrighted works or their elements in your own works?

Re-purposing other people’s work is literally what humanity has been doing for far longer than the term “license” existed.

If the original inventor of the fire drill didn’t want others to use it and barred them from creating a fire bow, arguing it’s “plagiarism” and “a tool that’s intended to replace me”, we wouldn’t have a civilization.

If artists could bar other artists from creating music or art based on theirs, we wouldn’t have such a thing as “genres”. There are genres of music that are almost entirely based around sampling and many, many popular samples were never explicitly allowed or licensed to anyone. Listen to a hundred most popular tracks of the last 50 years, and I guarantee you, a dozen or more would contain the amen break, for example.

Whatever it is you do with data: consume and use yourself or train a machine learning model using it, you’re either disregarding a large number of copyright restrictions and using all of it, or exist in an informational vacuum.

The act of learning is absorbing and using massive amounts of data. Almost any child can, for example, re-create copyrighted cartoon characters in their drawing or whistle copyrighted tunes.

If you look at, pretty much, any and all human created works, you will be able to trace elements of those works to many different sources. We, usually, call that “sources of inspiration”. Of course, in case of human created works, it’s not a big deal. Generally, it’s considered transformative and a fair use.

Not necessarily. There’s plenty that are open source and available for free to anyone willing to provide their own computational power.
In cases where you pay for a service, it could be argued that you aren’t paying for the access to the model or its results, but the convenience and computational power necessary to run the model.

Every work is protected by copyright, unless stated otherwise by the author.
If you want to create a capable system, you want real data and you want a wide range of it, including data that is rarely considered to be a protected work, despite being one.
I can guarantee you that you’re going to have a pretty hard time finding a dataset with diverse data containing things like napkin doodles or bathroom stall writing that’s compiled with permission of every copyright holder involved.

They’re not wrong, though?

Almost all information that currently exists has been created in the last century or so. Only a fraction of all that information is available to be legally acquired for use and only a fraction of that already small fraction has been explicitly licensed using permissive licenses.

Things that we don’t even think about as “protected works” are in fact just that. Doesn’t matter what it is: napkin doodles, writings on bathrooms stall walls, letters written to friends and family. All of those things are protected, unless stated otherwise. And, I don’t know about you, but I’ve never seen a license notice attached to a napkin doodle.

Now, imagine trying to raise a child while avoiding every piece of information like that; information that you aren’t licensed to use. You wouldn’t end up with a person well suited to exist in the world. They’d lack education regarding science, technology, they’d lack understanding of pop-culture, they’d know no brand names, etc.

Machine learning models are similar. You can train them that way, sure, but they’d be basically useless for real-world applications.