Key Points:
- The EU’s AI Act mandates AI companies to disclose training data. AI companies fear revealing trade secrets could lead to competitive disadvantages.
- The Act addresses the potential misuse of copyrighted content for AI training.
- Both small startups and major companies like Google and Meta will be affected.
- French officials advocate for innovation-first approaches to avoid stifling the industry.
A new set of regulations in the European Union will mandate greater transparency from AI companies regarding the data used to train their systems, unveiling one of the industry’s most closely guarded secrets. The EU’s recently passed AI Act addresses growing concerns over how AI companies acquire training data, including the potential misuse of copyrighted content like bestselling books and Hollywood movies without creators’ consent.
Since the public launch of Microsoft-backed OpenAI’s ChatGPT 18 months ago, there has been a surge in engagement and investment in generative AI, which can rapidly produce text, images, and audio content. However, the industry’s rapid growth has raised questions about data sourcing and copyright infringement.
The AI Act, which will be rolled out over the next two years, requires organizations deploying general-purpose AI models, such as ChatGPT, to provide detailed summaries of the content used for training. Following stakeholder consultations, the newly established AI Office plans to release a template for these summaries by early 2025. Despite the push for transparency, AI companies resist revealing their datasets, citing trade secrets and competitive disadvantages.
“It’s like cooking,” said Matthieu Riouf, CEO of AI-powered image-editing firm Photoroom. “There’s a secret part of the recipe that the best chefs wouldn’t share, the ‘je ne sais quoi’ that makes it different.”
The extent of transparency required will significantly impact small AI startups and major tech companies like Google and Meta, which are heavily invested in AI technology. Over the past year, several tech companies, including Google, OpenAI, and Stability AI, have faced lawsuits from creators alleging improper use of their content for AI training.
In the U.S., President Joe Biden has issued executive orders addressing AI security risks, but copyright issues remain unresolved. Calls for tech companies to compensate rights holders for data have received bipartisan support in Congress. In response to increasing scrutiny, tech companies have signed numerous content-licensing deals with media outlets and websites, such as OpenAI’s agreements with the Financial Times and The Atlantic and Google’s deal with NewsCorp and Reddit.
Despite these efforts, OpenAI faced criticism in March when its CTO, Mira Murati, declined to confirm whether YouTube videos were used to train its video-generating tool, Sora, citing terms and conditions violations. Further backlash ensued when an AI-generated voice resembling actress Scarlett Johansson’s was featured in a ChatGPT demonstration.
Thomas Wolf, co-founder of AI startup Hugging Face, supports greater transparency but acknowledges industry-wide uncertainty. Dragos Tudorache, one of the lawmakers behind the AI Act, insists AI companies should publicly detail their datasets to allow creators to verify if their work was used.
The AI Act aims to balance trade secret protection with the rights of copyright holders. However, French officials, including President Emmanuel Macron, have expressed concerns that stringent regulations could hinder the competitiveness of European AI startups. French finance minister Bruno Le Maire emphasized the need for innovation before regulation, warning that poorly informed regulations could stifle technological advancement.
