Creators will need to opt out of having their copyrighted character or other work recreated by OpenAI’s Sora 2, which launched today. If they don’t, the new text-to-video generator may create clips featuring that IP.
OpenAI is notifying talent agencies and studios about the updated Sora and the opt-out process, sources told The Wall Street Journal. It will not accept sweeping, blanket opt-out requests covering all of an artist’s or studio’s works, and rights holders will need to report specific violations they discover.
Creative industries have largely opposed opt-out systems for AI training data, arguing they put the onus on creators to exclude their content rather than requiring AI developers to seek consent. It could also erode their ability to control and profit from their creations, as companies would not need to negotiate with them for initial access.
Guardrails on styles and public figures
Interestingly, OpenAI has also been vocal in its disapproval of opt-out proposals, arguing that they present “significant implementation challenges” as they complicate the identification of legally usable content for commercial AI training. The firm certainly does not want anything to slow down its growth, especially when facing stiff competition from the likes of Google, which recently connected its Veo 3 video generator to YouTube.
However, OpenAI may be making an exception to avoid any lengthy, expensive copyright claim. It may also be operationally easier to block the generation of specified copyrighted characters/styles in Sora than to retroactively exclude them from training data.
The company faced backlash in March when users of the image generator built into GPT-4o started creating Studio Ghibli cartoons. (The demonstration of Sora 2 includes the Studio Ghibli style.) Engineers had to implement guardrails that prevented it from generating images in the style of a living artist, which are now also being applied to the new Sora, according to the sources.
Sora will also never generate images of recognizable public figures without their permission.
The ongoing debate over consent and compensation for AI training
The question of how creative industries can continue to exist alongside AI, which requires vast amounts of human-created training data to be useful to society, is being hotly debated. Creators need to be compensated, but allowing them to withhold their IP from model training if they disagree with the terms could lead to biased models, policy experts warn.
Legal battles and copyright debates are unfolding around the world as lawmakers grapple with how to resolve the fundamental tension between innovation and creative rights. Anthropic, Meta, Perplexity, Stability AI, Midjourney, Microsoft, and OpenAI (many, many, many times) are among the AI developers that have faced legal action from artists, news outlets, and musicians for using their work without consent.
In recent months, judges have ruled in cases against Meta and Anthropic that using copyrighted material to train AI models qualifies as fair use, provided it cannot recreate the original. This is something that OpenAI has, controversially, lobbied the US government to be included in the law.
Sam Altman’s startup has also signed several licensing deals with publishers to avoid further legal trouble.
US President Donald Trump has called it “impractical” to seek permission from every artist for the scraping of their work, with UK politicians holding a similar opinion. Political views on this issue are far from neutral: Tech companies openly resist rules they say would hinder innovation, while politicians are reluctant to deter these lucrative firms from setting up shop in their countries.
Companies including Microsoft and Cloudflare are looking to put the power back in creators’ hands by offering solutions that allow them to sell their content on a per-use basis.