Generative Artificial Intelligence (GAI) has in recent months exited the status of expensive luxury toy to enter the scene as one of the most powerful next steps in the Deep Learning Revolution. Whereas AI’s impact in the last decade should not be understated, it was limited to the analytical, with recommenders and classifiers powering popular platforms such as Netflix and TikTok. The Generative era promises to be even more powerful, no longer limited by labor-intensive creatives that bottlenecked speed and scale. When content in a variety of fields from text to image to video to audio to even code can be produced by anyone easily, endless possibilities are unlocked. New product experimentation can be faster than ever, and digital experiences can be more tailored to individual users than ever before.
The outputs of a GAI, or AI Generated Content (AGC), have applications that can affect every worthwhile human endeavor, but thorny questions abound regarding the intellectual property both powering and produced by these models. Unlike the platforms with powerful analytical engines, generative models have typically been trained by scouring the vast reaches of the internet. Whereas terms and conditions have been well defined about how a social media platform or a streaming site can use the data from users to train their algorithms, it is far less clear whether anyone has granted the right to use their data to power these more far-reaching algorithms. On the flip side, while AGC is produced by an AI, the human user leveraging the output desires some intellectual property (IP) protection over what’s created.
This piece will aim to first present an overview of these issues and conclude with a vision for how the AGC can not only power new opportunities, but with the right approach create more value for traditional creators.
Model Training
GAIs are trained on vast amounts of data for a given task and a significant amount of the heavy lifting of developing state of the art models is to amass the largest data set possible. A combination of scraping publicly available datasets and licensing proprietary data is commonly used, with more licensed data available as a company’s budget increases.
For natural language models such as GPT-3 and Co:here, a critical dataset is CommonCrawl, which archives much of the internet to be available to the public. Image data used for models like Stable Diffusion leverages datasets like LAION, which is built on top of CommonCrawl’s work, generating samples from 5 billion image-text pairs from CommonCrawl’s HTML image tags. These resources are built off scraping much of the web, among which many copyrighted and trademarked works are present. They use a fair use defense for the web scraping and dataset creation but stipulate that leveraging their datasets must respect existing IP rules.
This doctrine of fair use has been critical for many breakthroughs in media, think of how much of YouTube is powered by the remixing and critiquing of copyrighted content, and seems to have good standing with respect to the curation and maintenance of large datasets. The doctrine of fair use states that the force of copyright law is limited when the use of the work is transformative and does not detract from the value of the original work. Common fair use defenses include parody, criticism, research, and education. The research work that goes into creating and maintaining large datasets that include within them copyrighted works is evidently transformative, for being used to train an AI was not the original intent of the work itself and does not detract from the value of the original work by being used in building the dataset itself.
While simpler for datasets, the models built on top of them raise more complex issues. For example, a generative art product that makes images resembling those of a well-known artist does have a similar intent and might in fact detract from the value of the artist’s work. If it can be established that there is a direct relationship between the AGC and the copyright protected work, infringement may in fact be occurring. Some of this can be established observationally, as with the ghostly imprints of Alamy or Getty that appear in some prominent outputs. Even more broadly, trade dress protections, a subset of trademark laws that deal with “look and feel”, have been expanding in the US, making certain more qualitative judgements of style potentially infringing.
Striking hard on infringement in these early days of a new and fast-growing area of research would be a major blow to the very innovation that IP laws are meant to encourage. If training data and model development was restricted to compliant work, then only the largest firms would be able to innovate. AIs are data hungry and limits to what can be used based on risks that currently represent minimal commercial harm to creators is undesirable.
These exceptions, however, can’t last forever. The awareness that outputs can infringe on copyrights and trade dress is something that engineers must work to address as the state of GAI improves. Some techniques like adding hidden prompts can help minimize the presence of copyrighted works. In some cases, the final implementations of models may focus only on monetizing infringement without a genuinely transformative experience. For example, if I leverage a text to image model and customize it to make everything look like a Banksy piece. The look and feel of a Banksy work is protected, and by specifying Banksy it is drawing on a pool of images that are necessarily copyright protected. This application would not afford fair use protection if commercially used.
Model Outputs
On the flip side a question is raised about the protections afforded to the AGC itself. If the AI wrote a blog or created an image or developed a game, would any protections be afforded to it? Much buzz was created by a ruling earlier this year that an AI cannot copyright the works it creates. This ruling, however, was widely misinterpreted to mean that AI-generated works lack copyright protections. In fact, the case simply reasserted a legal norm that only humans can possess IP. That a GAI itself cannot own the work is in line with saying that an AI-powered self-driving vehicle is not responsible for the accident. The human in the loop, and there is always a need for some human in the loop, is responsible.
AGC involves humans at many stages, from producing the training data, to developing and maintaining the models, to the acts of prompt engineering and synthetic curation, new skills in the world of GAI-enabled creation. Acknowledging this clears up that there are many potential vectors on which humans are involved and to whom IP protections can be afforded.
A little bit needs to be stated on prompt engineering and synthetic curation. Prompt engineering has proven to be an important driver of the quality of AGC. It is the art of communicating in a way that can ensure that the GAI produces what you desire. Many of the products built on top of large language models and image generators rely on expertly crafted prompts to ensure that high quality outputs in a relevant domain are achieved. The best generative artists work tirelessly to craft ideal prompts to regularly generate beautiful artworks. The prompts themselves have a creative component to them, and their uniqueness can contribute to their eligibility for IP protection. Work is already be done to establish ownership of prompts and improve the search space for the best prompts that were used. Courts have established creative input tests to establish the eligibility for protection of short stories or taglines, and a similar test can assist in protecting the unique prompts that drive high quality output.
Synthetic curation is another important skill in the world of AGC, though as an art form it is in earlier stages than prompt engineering. AGC can be multimodal, taking inputs that are texts or images or audio or video and weighting them in some combination to create works that are any combination of these formats. The AGC itself can then have a human stitch them together or arrange them in a unique way or even feed them back into a GAI to achieve their final vision. The more synthetic curation that is done, the more it can be well established that IP protections should be afforded.
It is important to want to afford ownership for AGC and establish fair tests for determining eligibility as this would enable further growth in the field and expand economic opportunity. We can imagine GAIs becoming quite critical in scientific fields from new materials research to drug design as they become more crafted to understanding physical and biological properties. If the generations were not ownable and the patent protections surrounding the outputs not possible, this would discourage their use in these fields limiting their overall benefit.
Across the board, from marketing creatives to game design to merchandise production, the ability to establish rights over the works used is critical to the business operations that AGC may power. To accelerate the development of sound business models leveraging these novel tools, which would in turn benefit the potential copyright holders of the underlying data they were trained on, maximizing the speed and scale of implementation should be a key direction of IP policy concerning these technologies.
A Vision for the Era of AGC
Understanding the nature of the IP issues surrounding generative models allows us to chart a path to making them maximally
beneficial for everyone involved. An approach to be drawn on here is Data as Labor, pioneered by economist Glen Weyl and technologist Jaron Lanier. Data as labor identifies the production of the underlying data in large datasets as an act of labor, which grants creators certain rights, both economic and social, over its use. An analogy for data is music mechanicals, a component of the overall musical work which are often stitched together from the labor of various contributors. With this framework data can be represented the way that music is, with legal developments such as the Music Modernization Act helping to ease the ability for small creators to be compensated for their work.
There are several considerations that must be addressed to ensure that AGCs both produce value for generative creators and for the creators as data laborers.
First, it is evident through the vast amount of AGC already produced that not all generations have economic impact. Setting rules about what happens with created works that are not used, and even establishing a difference between works that are used privately and commercially sets the stage for where compensation occurs.
Second, not all works are exploiting of another’s labor. Watching an artist produce a work and taking inspiration from them does not in fact violate their rights but reproducing their key elements and motifs does. GAIs themselves can play a role in establishing the degree to which creation is indebted to a data laborer. In the image domain CLIP interrogators can be used to establish the degree of relationship between a piece of AGC and the corpuses of various artists, highlighting whether any singular work or artist or group of artists were the main drivers of the final output. In the field of text and video, similar models can be easily identified, with the broad range of AGC feasibly being brought within this testable domain. In addition, the prompts themselves can be queried to see if they make direct reference to protected works or their creators to establish their rights.
Next, compensation must be dependent on the commercial viability of the output. Protecting the IP of AGC is prior to protecting the rights of the data laborers who enabled them. If no profit can be made, then nothing can be compensated back. Establishing an effective rights management solution to this is one of the most promising applications of Non-Fungible Tokens (NFT) today. While NFTs are not a replacement for traditional legal rights, they can make enforcement of these rights easier. By establishing a digital ownership chain that can split up among multiple rights holders, NFTs can make managing rights for large volumes of content with a myriad of contributors easier. This would help creators no longer feel exploited by the derivative works GAIs enable, but rather feel a part of that relationship. Like how sampling has created expanded economic opportunities for musicians whose work constantly breathes new life, so too can AGC develop entirely new revenue streams for traditional creators.
Behind this, a rights management layer must exist to assist creators in being able to be recognized, managing the NFTs and payment schemas, standardizing the evaluation criteria for derivative works, and bargaining with platforms to ensure their policies respect the wishes of their creators, and that derivative works are not used in a way that goes against their wishes. The concurrent technological work done with CLIP similarity and NFTs makes this management ecosystem more feasible than ever before, and we can envision it maturing as quickly as music rights management has in the age of streaming.
A final important note on this vision is that it references only creators both traditional and generative. The middlemen of data collectors, model infrastructure providers, and productizers have not received the same treatment. In the case of data collection, it is evident that their data collection falls under fair use specifically because they themselves take no stake in the creative output and genuinely use the protected material in transformative ways. If they were to have rights over derivative works, this status would be called into question, changing incentives in a way that may not benefit the research community.
Similarly, those that produce and maintain generative models allow a diverse amount of work to be done, much of which leverages completely public domain work or that is used for non-commercial applications. It makes for more sense for them to be neutral to the downstream applications of their tools and charge on a usage basis. Entering into ownership agreements muddies whether usage-based pricing is eligible for being claimed by data laborers.
Finally, the end products that make GAIs accessible cannot uniformly have ownership stakes. The diversity of products, ranging from general purpose use like those made by WOMBO to marketing specific like those of Jasper, each have different relationships to commercial use. Some of these may output works that are generic enough that no IP protections are even possible, while others may tailor their outputs to a very specific catalogue and must pay traditional royalties for access. The challenge of developing standards is to make them clear enough that innovation is possible without making them so rigid that it’s only allowed in a limited way.
Nonetheless the vision established above outlines how AGC can create new economic opportunities for creators, expand creation, and drive product innovations while respected the rights of those work GAIs have learned from. Sketching out visions early is an important exercise in guiding both technical and legal development towards positive sum outcomes.
Conclusion
The attention being drawn to generative artificial intelligence from technologists, investors, and the public is justified by the vast array of new tools and applications that this technology can support. With GAIs becoming better, cheaper, and more accessible it is important to grapple with the intellectual property issues raised by artificial intelligence generated content.
In the early days of any new space weaker rules should dominate that allow people to move fast and break things, but concurrently it is important to theorize about where things are heading and what should things look like when they mature. As AGC adoption increases the importance for end users to understand their IP protections and obligations becomes vital to scale.
By looking at the models themselves as mechanisms for improving enforcement legal experts and policy makers can craft rules that are more unorthodox and ambitious in protecting rights. The vision laid out here argues that rather than in conflict the rights of traditional and generative creators can only be respected in tandem.