THE LEGAL ASPECT OF SYNTHETIC MEDIA GENERATED BY AI

Machine learning algorithms have revolutionized the way we handle, process, and understand information on a very profound level. For the first time in history the capacity of current GPUs, is sufficient to digest, order and reorder an enormous amount of information and data which by far exceeds the capabilities of the human brain.

While the ability to process huge amounts of data has not yet reached the capacity to replace the human brain or consciousness (if ever), it must be said that certain functions of the human brain can already be reassembled. Amongst those (with limitations of course) is the field of visual content perception and creation. Recent GANs, generative adversarial networks, and DDPMs , denoising diffusion probabilistic models, effectively generate visual information with increasingly stunning accuracy and stochastic creativity.

From a technological and creative point of view, these developments are very exciting, but it is also a new frontier for the legal assessment of how we handle visual content. These developments have the potential to redefine our understanding of copyright laws, tangent’s privacy rights, and what place “inspiration” has in the legal definition of visual content. Additionally it also may recategorize a neural network and bring it to the same level as a human brain from a legal point of view.

DISCLAIMER: I AM NOT A LAWYER AND THIS IS NOT LEGAL ADVICE. It is simply my personal interpretation of the current situation. Consult a lawyer in your relevant jurisdiction.

Let’s begin with a brief explanation of how synthetic media is created via machine learning algorithms.

1 The Process
I do not necessarily like the term “intelligence” when talking about many machine learning algorithms. While the process reassembles the neural structures of an organic neural network (aka CNS or Brain), the basic functionality can be, for the sake of clarity, explained as an A to B abstraction generator. That means, that if you feed an algorithm (model) with a certain amount of data, this algorithm then creates a trained model — a condensed neural abstraction / representation of this fed content. Then you can generate variable output based on this which reassembles the input data.

Here are several important steps in this process:
1. The source data — where does it come from and how is it prepared? (Both aspects have legal implications.)
2. Who wrote the used Model/algorithm/framework, code, and from where does it come?
3. How is the copyright assessed on the output data?
Let’s start with the first and probably most difficult step.

2 What is creativity?
Before we dive into the specifics, let`s analyze how art or content is generated by humans nowadays and how this process interacts with the law.
Any new creation is nothing more than a rearrangement of existing content. Everything a human can create is an abstraction or reassembling of some existing piece. We take a reference point A, then create variations of it, modify the context with element B, then jump between various interpretational levels and so on. “Inspiration” or “creativity” does not refer to any form of “original” ideas, but to the interpretation of an existing concept modified into a new context or form.

Just as in an ML Model, the human brain learns to digest sensory information through basic pathways that are pre-determined by our bodies. Eyes can see in this light spectrum; ears can hear that frequency spectrum, and other basic interpretations of this information on a CNS level determined by genes — fear of spiders, etc. Infants learn quite young how to recognize all incoming data — from colors, shapes, and sounds. As they grow they will learn to organize and label this data correctly adding context to the psychological construction of “I” and their own interaction with the world. Any feeling, idea or inspiration we have is usually born through a constellation of existing labeled elements or the discovery of a new “label” or entity through the previously unknown combination of two or more existing elements. (For instance, a small child recognizes stripes as a visual element and also the shape of a horse so when they see a zebra, they will combine these two data points into a new creature.) This process can stretch in a human brain through many different layers of meaning. You can symbolize an abstract idea of, for example, God, by very basic labeled elements such as “sunlight breaking through clouds”. Another example of the combination of two different elements of different importance layers is a famous painting from the black metal scene where a man shoots an angel from the sky with a gun. Both elements are not usually connected and identified as not compatible but create a new context reflecting a certain emotion.

Human being sare capable of creating stunningly beautiful artwork where the original source of inspiration is no longer evident, but in other cases, the final published artwork is extremely close to the original inspiration and the artist basically copied it. In many cases, the only reason why this has no legal consequence is because the original artist or copyright holder cannot prove that the new work is based upon his art. A human brain is like a closed black box — it is difficult if not impossible to prove that the creator of the new work simply did not have the idea on his own.
However, in machine learning the situation is a bit different because datasets on which the model are based are fixed entities of information. They exist and therefore the original copyright holder can access, at least in theory, the necessary information to prove that the new work has been copied. In this case we would talk about a “derivative”. If you use this data in your dataset and face a legal problem, the only thing you could do is “lose” the original dataset or block access to it in any other legal or physical way, but both will not shed a good light on you and ruin the positive outcome of any due diligence.

3 The source Data. The fuel for any AI

No AI/ML is functional without being fed with a massive amount of source data. Just like a child’s brain, the model must learn with the fed datasets but unlike a human brain, which can switch between different hierarchies and layers of logic and selective perception, an ML model is usually tied to the level of information it is fed. Consequently, itis less efficient at learning as it requires a great deal more data to understand the context of a certain type of information than the human brain. But AI/ML has the capacity to digest much more data than the human brain and store it in a reliable way.

In this article we are focusing on synthetic media, so let’slook at human face recognition generated by AI.

When AI is tasked with learning to generate human faces it needs hundreds of thousands or even millions of images as examples. This is the A input in our A to B content generator. Everything relies on this.

In a practical way, this data must be sourced in the first place from somewhere and here is where the potential for problems can arise. We can scrape images from Flickr or Facebook Profile pages — but what implications does this have?

a) Are we legally allowed to use a third-party service to download and source data? Does the TOS of Flickr or another platform allow data mining from their site?
This aspect is often overlooked perhaps because ML is still relatively new, but within a few years most internet platforms will, most likely, adapt their TOS to either permit or prohibit data mining from their services. Most of the larger stock photo agencies already block data mining of their images. From a legal point of view it is clear since the wording of most TOS is very specific — especially if the subject is directly targeted.

b) Copyright
No matter which site the data is sourced from, someone usually owns the copyright to every photo found online. Before we discuss these copyright issues, it is worth mentioning that there is one exception. If an author publishes his work under the CC0 creative commons license (or similar) he agrees to abandon all rights to his creation. Creative commons licenses are also irrevocable — the creator cannot undo this step for licenses already issued and the content becomes public domain. Additionally, the CC0 license is not interpreted the same way in all countries, so it may be a valid or invalid legal concept that somebody can choose to “abandon” or “transfer” the copyright of his work.

The CC0 creative commons license can be found here:
https://creativecommons.org/publicdomain/zero/1.0/legalcode

Due to this factor, for all subsequent commercial usage of copyrighted visual content we are obliged to observe the valid legislation in place how to respect the original author’s copyright.

Derivative work, transformative work, and fair use.
One of the most important questions concerns the ownership of creative work when using an original, copyrighted piece of art. Unfortunately, there is rarely a clear answer to this question. At times these disputes result in court hearings which spark outrage either from copyright holders or from entities using an existing work commercially and facing a massive fine.

Derivative work
The United States Copyright Act of 1976, 17 U.S.C. Section 101 states:

“A “derivative work” is a work based upon one or more preexisting works, such as a translation, musical arrangement, dramatization, fictionalization, motion picture version, sound recording, art reproduction, abridgment, condensation, or any other form in which a work may be recast, transformed, or adapted. A work consisting of editorial revisions, annotations, elaborations, or other modifications which, as a whole, represent an original work of authorship, is a “derivative work.”

One of the most critical legal aspects in machine learning and dataset sourcing is that any ML algorithm depends 100% on the source dataset. There is no genuine or original creative aspect to it. A GAN like Stylegan or BigGan cannot add to this source data by itself. All outcome has its root in the original data plus the interpretations of the AI and the “noise” that has been added.
Because ML is relatively new, creative content provenance is an unexplored frontier legally. It will depend on each jurisdiction and on the individual court within that jurisdiction and how they will interpret each case.

If the jurisdictions determine or rule that all AI output is a derivative of its underlying datasets any entity creating synthetic content must be the copyright holder of the original datasets or use CC0/public domain content exclusively.

One of the first court rulings in this area was the case of the Authors Guild which sued Google for using original copyrighted books to train their book search engine algorithm. The case went through three different examinations and the court finally interpreted Google’s use of copyrighted works to create a commercial algorithm as “fair use”.

Here the original text concerning that case:

“Google’s unauthorized digitizing of copyright-protected works, creation of a search functionality, and display of snippets from those works are non-infringing fair uses. The purpose of the copying is highly transformative, the public display of text is limited, and the revelations do not provide a significant market substitute for the protected aspects of the originals. Google’s commercial nature and profit motivation do not justify denial of fair use. Google’s provision of digitized copies to the libraries that supplied the books, on the understanding that the libraries will use the copies in a manner consistent with the copyright law, also does not constitute infringement.”

And here is a corresponding article:
https://towardsdatascience.com/the-most-important-supreme-court-decision-for-data-science-and-machine-learning-44cfc1c1bcaf

However, this was one interpretation of one specific case and more importantly — the output type of the ML algorithm was of different nature than the input type. Google used input material in the form of text to create an abstraction of this text, but did NOT generate text with its new model, but rather used it to fuel its search engine function.

In the case of synthetic media, it is quite a different scenario because visual input data is used to generate visual input data, and this is a whole different story.

Another example is when you scrape images of faces off the internet to train the model to identify faces. If you teach the algorithm with images to generate images, it is much more likely to be considered a derivative work.

Fair use
The concept of fair use has quite a few different legal definitions around the globe and is handled differently in different countries. At this time there is no clear international mandate of what, or what is not, considered to be fair use and what can be copyrighted by the creator of the non-derivative work.
In 2013 for example, the second circuit ruled in the case of Cariou vs Prince (https://en.wikipedia.org/wiki/Cariou_v._Prince) that the following work displayed below is not a derivative work, but an artwork on its own due to the transformative nature and the newly created original expression of art.

But the question is, can this also be applied to ML algorithms?

Maybe — or maybe not.

As discussed earlier, a straightforward GAN which generates synthetic images relies 100% on its original dataset and simply creates variations of the input data. There is no genuine or original idea which may render the argument of transformative work useless. On the other hand given a sufficiently large dataset, the AI output data may look nothing like any particular image in the original artwork.

Due to the lack of legal clarity in this situation and the varying interpretations of the various courts and jurisdictions, it’s best not to count on any specific interpretations of Fair Use in the context of machine learning until there are more reliable court rulings or laws in place.

Due to the unclear situation of the copyright status and the technical advantages a wholly owned dataset offers, at vaisual.com we create our own datasets exclusively and own the full copyright to any creation.

c) Privacy laws
It took a while for the jurisdictions to wake up and address these issues but it has happened. The European Union implemented its famous GDPR law to protect the personal data of its citizens, and the USA has various data protection laws in place (TFC act, Privacy act, COPPA, HIPAA, FCRA, GLBAand more). In Europe, there was an extensive discussion especially about photography. Wedding photographers in Germany now are required to ask permission from all attending guests to be able to photograph them!
In a machine learning algorithm if you were to feed Stylegan exclusively with, about 1000 images of the same person, what would the generated synthetic image look like? You probably can guess the answer — just like the original person with slight variations and the typical technical flaws and issues when a model is trained on too little data.

We are now facing two legal questions when we use other people’s faces to train our model: their biometric data and their identity.

1. Do we need permission from the individual person?

Under the current interpretation of privacy laws each individual is the sole owner of their own biometric data. This is implemented (or not) in various ways but ultimately it boils down to this — we can expect legislation to follow more and more this paradigm in the decades to come. My personal interpretation is this… we do need permission from the person to be able to use their biometric data in commercial ML projects — a model release. In Visual Ltd. for example, you must obtain a signed contract with two witnesses, before using any image or personal data of anyone in a machine learning project.

2. Can the original photograph be traced back from the output data and does the output closely resemble that person?

It is possible, in most cases, to recognize an AI generated image nowadays and depending on the algorithm and the size of the dataset the AI output may look a little or a lot like the original input. It is just a matter of how many images you generate, and certain aspects of the original face could be reassembled, although it is never a perfect one to one copy. It remains to be seen how the judicative powers around the globe will interpret this part. In Visual.com, our model releases would give us permission to publish the original face without any modification.

d) The Software
Another issue of importance is the software used in the process. Today, code for Machine Learning consists of many layers of code created primarily in an open-source environment often by hundreds of different people, companies, and entities. In general, the software itself has no copyright claims on the content created with it. For example, Microsoft is the owner and copyright holder of Microsoft Word, but that has no legal implications on the text written with it. If an author writes a book in Microsoft Word, the author is ultimately the copyright holder. However, any software developer may create its own license for its own usage. Student licenses with reduced prices often exclude commercial usage of their software in their TOS. Nvidia also restricts the usage of their open-source project Stylegan 2 to non-commercial projects.

e) Who owns the copyright of AI generated synthetic content?
Who created the content and how? It may be that the AI output of synthetic media is considered a derivative work of the underlying dataset and in that case the copyright is carried forward from the dataset to the AI output. This is our working assumption at Visual right now. Alternatively, the data scientist who adjusts the hyperparameters in the model training or during the training process via preparation of content is considered an artist by applying their skills in a relevant manner to influence the output. And indeed, these adjustments make a significant difference. Other parties could also argue that AI output simply is not subject to copyright — the key argument with which these claims can be reasonably defended.

What’s on the horizon?
Now after looking at the individual elements, zooming out to see the bigger picture, another important question remains: Does the law ultimately recognize the creations of a neural network to be equal to those of a living person? Just as with a human brain, a neural network learns by absorbing information from the outside world and then rearranges and interprets it — the process is virtually the same. This is how technology is born-with the one key difference being that a human may claim that he had the original idea independently of other artists while in a neural network there is evidence regarding that in the form of the original datasets. There is always room for interpretation and if future courts rule that the output of AI is to be treated the same way as the output of a human artist then we will face a whole new era not only in legal ramifications, but also in how we define creativity.

With such a ruling, it would be the time when AI will be deemed equal to a human being.