The advent of AI has brought a new “interpretation” of copyright law from big tech companies. Suddenly, it seemed to be acceptable to pirate millions of books, movies, and Open Source software programs to train your AI. The resulting AI would then create content that looked eerily similar to what humans had previously created. Especially in the visual world, the results were striking. From slightly modified Mikey Mouse characters to a flood of Studio Ghibli anime-lookalike profile pictures on social media.
In the name of progress, we artists, developers, and content creators were supposed to accept the erosion of our rights. After all, artificial intelligence promises a better future for everyone. Yet it left a bitter aftertaste: Billion-dollar Big Tech companies are stifling the little guy out of their income.
At the same time, there is a push to make AI content copyrightable, with little or no human creativity involved. After all, if you want to replace human creators and engineers with AI, you need to ensure that you can still extract the value from your customers.
However, two barely functional mockup clones of proprietary software have exposed the problem behind this thinking. More importantly, they have shown that Big Tech boards and management have forgotten how their companies are actually making money and why it might be a bad idea to slaughter the golden goose in the name of AI.
The Need For Content In AI Training
AI models are complex. Depending on the area of expertise, you analyze billions of data points to build each model. For ease of illustration, let us assume that “language” in a language model refers simply to human languages. Each decision point in a language model then describes which word is likely to follow the previous one, taking into account the knowledge within the model and the prior decision flow. So when you prompt a chatbot, you get a slightly dated-sounding but mostly correct answer. Of course, in areas where specialized knowledge is required, but some word structures are all too common, such as case names in law, language models can quickly invent knowledge just because a string like “Tree Trunk vs. United States” seems reasonable to the word flow.
Yet, most crucially, the data needed to train these models is mostly hidden away in books, source code repositories, and research papers. Most of it is protected by existing copyright laws. Yet AI developers have tried to push aside copyright claims by comparing AI training to learning or by ignoring complaints about copyright violations.
Given the amount of data needed for training, the financial motivations are relatively clear. These companies could not survive if they had to pay for every piece of data they ingest. Yet weakening copyright and trademark protections for training data applies equally to everyone.
The Responsibility For AI Outputs
The willing stretching of copyright exceptions is just the start of the pattern of trying to rewrite the legal frameworks that governed companies in society for the past century. Responsibility for AI outputs is a similar issue. Whether driven by user prompts or misbehavior, AI can create questionable and illegal outputs. Yet, AI companies are pushing for a situation where no one is responsible for what the AI creates. After all, if any error is just a Deus ex machina, and not a deliberate user action or developer oversight, we can all keep using AI without regard for errors or social costs.
This extends to how AI companies view copyright output. Unless the copyright holder is Disney, AI companies do nothing to ensure AI output complies with the copyright laws of the input. Even if the output is a blatant copy of some Open Source code, the view of Big Tech is that AI changes the training data far enough that any resemblance is purely coincidental.
Thus, existing copyright protections and copyleft provisions should not apply to the generated output. Instead, the new work is a creative process of the machine that the user should be free to use, copyright, and, ideally, feed back into the next iteration of training data.
The Risks For The Cash Cow
Yet, this thinking has one significant problem. Most AI companies are not cash flow positive. Thus, venture capital and Big Tech revenues are feeding the need for cash. Much of this is financed through the proprietary ecosystems Big Tech has created over decades. Nothing symbolizes this more than the subscription licenses for Microsoft Office and the Adobe Creative Cloud. Both charge heavily for the “industry standard” software, which has changed little over time and whose interface remains familiar to billions of office workers worldwide. Despite better alternatives, customers still pay high prices for the interface’s familiarity.
Yet a single vibecoding developer has exposed the problem in Microsoft and Adobe’s AI-first thinking. Two barely functional clones of the flagship products now exist on GitHub. While the software’s functionality barely exists, the interface is an almost perfect copy of the existing Cash Cows. Yet if the excuse for buying Big Tech proprietary software is familiarity with the interface, a similar interface might be enough to push enterprise buyers to use a free alternative.
The Catch-22 In AI Copyright
This brings us to the Catch-22 of AI copyright. To fund further AI development, Big Tech needs revenue from locking users into its existing ecosystem. Yet, to develop and market its AI, they need to weaken copyright protections on the input and ensure favorable copyright protections for the AI-generated output. In short, the same protections they need to keep their income-producing business afloat are the protections they need to fight to keep the AI race going.
Right now, no one can predict with certainty whether or not companies will be able to resolve this issue. Yet it clearly shows that the push for faster AI development has left governance and risk management behind. It may be time to think about the consequences of our actions instead of constantly hunting for ever bigger and more expensive models.
Leave a Reply