From Tilly Norwood to Sora 2, AI’s copyright reckoning is overdue
We don’t need to ask permission to use publicly available content for training our models — that seems to be the AI industry’s approach to training data.
“Tilly Norwood is not an actor.” This desperation laden blunt statement from the Screen Actors Guild–American Federation of Television and Radio Artists, certainly reopened a debate, after the AI actress’ creator Eline Van der Velden suggested multiple talent firms were interested in signing the new artificial intelligence creation. If this didn’t do enough to rekindle the conversation about artificial intelligence (AI) companies and a seemingly interminable interest in using copyright data to train their models, OpenAI’s blunt message that the onus is on rights holders to explicitly ask that their copyright material is not featured in videos their new Sora 2 tool generates.
This isn’t a unidimensional development, even as social media feeds get filled with AI debris, such as fake videos of Sam Altman getting caught stealing GPUs from a shop. There are arguments for and against the fair use of just about anyone’s data to make a case for AI progress, how many lawsuits are playing out in courtrooms, and the India context, wherein AI companies operating in India aren’t storing or training with data here and therefore the Indian copyright laws don’t apply to them.
Onus on artists and creators
We don’t need to ask permission to use publicly available content for training our models — that seems to be the AI industry’s approach to training data, an ideology that technological progress justifies scraping all publicly accessible data, for the potential good of humanity. Perhaps when there is enough compute and GPUs available. It would be fallacious to say this philosophy has already reached its apotheosis, because new milestones are created regularly.
“Several high-profile lawsuits have been filed against companies like OpenAI, Meta, and Stability AI, alleging unauthorised use of copyrighted content to train their generative AI models,” notes Negar Bondari, who is on the executive board of the University of Southern California’s Gould School of Law.
Suno, facing lawsuits from major record labels, openly admitted last year that it trained the AI music generator on copyrighted songs, and constituted fair use under U.S. copyright law. Nevertheless, the AI music company has gone ahead and launched Suno Studio this week, which it calls the world’s first generative digital audio workstation.
OpenAI is facing multiple cases for using copyrighted material without permission to train its GPT models. Google too finds itself in several legal battles, for training its models on work without the creator or copyright holder’s permission. For instance, artists allege Google’s Imagen model was trained on their copyrighted images without permission, thereby creating infringing “derivative works”.
Amazon-backed Anthropic’s $1.5 billion author settlement in September, to avoid a trial scheduled for December. becomes the largest publicly reported copyright recovery in the US, approximately working out to around $3000 per eligible book, and mandates Anthropic to destroy the datasets containing what were essentially pirated versions of books. Writers have also sued Microsoft for the use of nearly 200,000 pirated books to create an AI model, Megatron. Whether the Anthropic settlement has any bearing on this case, remains to be seen.
Is the law supporting an AI age?
Disney, Universal and Warner Bros. Discovery sued Chinese AI firm MiniMax, whose service called Hailuo AI generated characters such as Darth Vader and Wonder Woman, without permission from the studios. They seek $150,000 per infringed work. “MiniMax’s copyright infringement is wilful and brazen,” note the companies, in their complaint.
This week, Disney sent a cease and desist letter to Character.AI claiming that it has chatbots based on its franchises, including Pixar films, Star Wars and the Marvel Cinematic Universe. The AI company has since removed Disney characters, including Princess Elsa, Moana and Spider-Man, from its chatbot platform. “In sum, Character.AI chose to systematically reproduce, monetiSe, and exploit Disney’s characters, that are protected by copyrights and trademarks, without any authorisation, in a way that is anathema to the very essence of the Disney brand and legacy,” the letter states.
What are the AI companies saying? They are of course lobbying hard for training on copyright data to be considered as ‘fair use’.
For instance, OpenAI CEO Sam Altman believes we’re in the “Intelligence Age”, and the company in a proposal submitted for the US AI Action Plan wrote, “We propose a copyright strategy that would extend the system’s role into the Intelligence Age by protecting the rights and interests of content creators while also protecting America’s AI leadership and national security. The federal government can both secure Americans’ freedom to learn from AI, and avoid forfeiting our AI lead to the PRC by preserving American AI models’ ability to learn from copyrighted material”.
Case closed? Not quite.
The judge who approved Anthropic’s settlement tried to attribute human characteristics to AI growing up, noting “like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different.”
Earlier, writers including Sarah Silverman and Ta-Nehisi Coates, lost the case against Mark Zuckerberg’s Meta, with the judge noting that the authors had not presented enough evidence that Meta’s AI would cause “market dilution” by creating works similar to their originals.
A unique India context and way forward
Virag Gupta, AI and Cyber Law Expert, explains to HT that AI companies such as OpenAI telling Indian courts that since they don’t do AI training in India, there is no copyright infringement, is akin to “somebody saying on affidavit that yes, I am a criminal. This will be a challenge for India’s regulators and legal framework to resolve, before AI companies can be potentially pulled up for copyright infringements. Gupta is pointing to OpenAI’s submission in the courts stating they are not responsible for operation of services provided by ChatGPT, nor is it the owner or lessee of servers on which the training data is stored, and that none of the training of its AI model takes place in India.
Gupta also questions a broader official push for OpenAI’s ChatGPT when data storage as well as localisation remains unclear as per the Public Records Act, and hopes the jurisdiction will be enforced. “Exclusive AI platforms and many verticals of tech companies including Google, are not compliant with the Indian law. They don’t deserve conditional protection under Section 79 that is provided for safe harbour because even as AI platforms, they are IT intermediaries,” he says.
AI companies frequently assert that their models don’t replicate copyrighted works but instead “learn” from them by extracting patterns, linguistic structures, and contextual insights. However, the fundamental purpose of generative models is to statistically mimic the system that produced the training data. This allows language models to be prompted to write in the style of a specific author or mimic an artist’s work, as they have encoded patterns from the original material. Case in point, ‘personalisation’ that AI companies often pitch.
As USC’s Bondari notes, legal frameworks will need to keep pace. “For lawyers, artists, tech companies, and policymakers, the intersection of AI and copyright law represents one of the most dynamic and consequential legal debates of our time,” she says. They’ll need to be quick. Emily Blunt, in a podcast recently, said this when shown a picture of Tilly Norwood — “Does it disappoint me? I don’t know how to quite answer it, other than to say how terrifying this is.”
Blunt sums up what most creators worldwide are thinking.
