Inevitably, all technology revolutions face their day in court. And now, generative AI is facing its day in court. The technology that composes poetry, paints portraits and writes code was developed based on an enormous reservoir of human creativity, and those who created it were seldom consulted. Now the issue facing the courts and developers is whether this learning from human creations is proper learning or massive theft.
The Original Sin of the Dataset
They are not drawing inspiration from a void. Instead, generative models are trained on billions of pictures, articles, books, songs, and even lines of codes, mostly pulled directly from the internet. The fundamental problem is that of consent, because the creators of such art were usually not informed about this practice, nor did they receive any form of compensation and were given an opportunity to opt out.
The developers argue that training is a process similar to human learning. An artist learns by observing paintings of other artists, a novelist – by reading many books and is not charged with a license fee for that. The comparison sounds convincing but it is quite flawed. A human gets acquainted with several thousand works during his/her whole life and the style acquired through this process is influenced by experience. A model learns millions of works in just a few weeks and is capable of reproducing the artistic style of a living artist instantly and in large quantities – something impossible for a person to do.
A Legal Landscape Still Taking Shape
The copyright framework is built upon a small number of foundational principles: copyright does not apply to ideas but expression, gives creators an exclusive right in reproducing their work, and allows for exceptions, like the fair-use doctrine in the United States and text-and-data-mining exemptions in some regions of Europe and Asia. The generative artificial intelligence tests all three of them.
Litigation filed by authors, visual artists, news outlets, and music publishers claims that the act of reproducing works to train a model constitutes infringement per se. Meanwhile, the defendants claim that it is transformative since they are only using the works as a means of extracting the statistics, but not redistributing and storing them. Courts have been inconsistent in their rulings, and there is still no established international rule. Some jurisdictions choose to allow more exceptions to foster innovation, while others move toward imposing transparency requirements and opt-out policies.
The Output Problem
The issue of training is just one part of the controversy. The other relates to what these models create. Now and then, the model simply produces plagiarized chunks of text or even images with watermarks, and it becomes difficult to argue that this is anything but plagiarism. Usually, however, what is produced is something new – but this something new directly competes with the creators of the work in the same market.
Recommended for You:
- Prompt Engineering & Optimization: The Strategic Discipline Behind High-Performance AI
- Best Way to Cope with AI Amid the Rapid Shift in Technology
- The Machine That Sees Tomorrow’s Weather: How AI Is Rewriting Climate Science
And this is where another question arises: who actually owns what the machine has created? In most cases, the law states that copyright presupposes human authorship. So, there is no clear position regarding pure machine generated works. A person who creates iterations of prompts, edits and curates is likely to own more rights than a mere prompter.
Beyond Law: The Ethical Dimension
Legal is not the same as legitimate. Just because it is legal does not mean that it is not damaging. When generative tools reduce the value of illustration, translating, voiceover and writing by using the market, the people who are going to be negatively impacted are the people who have worked hard to teach them. Ethics would take into account whether something is legal, but also cost and benefit.
There are many other effects to consider. Models can mimic someone’s voice and likeness without their permission which can lead to identity theft and other damage to their reputation. Models can also take bias from their training data and make it sound like they know what they are talking about. They also cause confusion between real work and fake work which undermines the trust that journalism, scholarship and the arts need. Attribution is an important part of creative work but cannot be used on systems that do not know where their information came from.
Paths Toward a Fairer Settlement
However, the discussion need not be at a standstill. A number of effective approaches are coming into being:
– Rights licensing markets, where the rights owners get remuneration for the use of their catalogues in training, as has been implemented in the music and stock photo businesses.
– Effective opt-out/opt-in approaches, making consent a process rather than an afterthought.
– Transparency of datasets, where the developers reveal the materials used in training their models.
– Standards of provenance and labeling, which will make it possible for consumers to differentiate between human and machine work.
– Compensation systems, inspired by the model through which broadcasting royalties support creators.
None of these is without flaws, but all recognize creators as stakeholders, not as raw materials.
The Verdict Is Still Out
The use of generative AI technology is not the harbinger of doom or salvation for humanity. On the one hand, this technology allows making creativity more accessible and increasing people’s aspirations. At the same time, however, technology based on the cumulative human culture has to give something back to this culture. The question of how to maintain the incentive to create and the right to innovate at the same time is one of those questions that need to be answered by those societies that see them as mutually exclusive.
