Amid growing controversy over generative AI copyright infringement, the Vermillio platform claims it can track the percentage of AI-generated images derived from pre-existing materials. We explore how this technology reveals reliance by tools like Sora and Gemini on famous artworks such as James Bond and Doctor Who.
- The problem of originality in generative AI
- The copyright battle
- Vermillio platform: shedding light on closed models
- Neural fingerprint: how does the technology work?
- Case studies: Doctor Who and James Bond
- Training data: the source of controversy
- Industry response and “fair use” defense
- Call to action and protecting creators
- FAQs
The problem of originality in generative AI
If you ask Google’s AI video tool to create a movie of a doctor traveling through time and flying in a blue British police box, the result, unsurprisingly, resembles “Doctor Who.” And if you ask OpenAI’s technology to do the same thing, something similar happens. You might ask: what’s wrong with that?
The answer may be one of the biggest problems facing AI leaders. Generative artificial intelligence is supposed to be just that—generative, meaning it develops new answers. But how much of that output is truly original?
The copyright battle
The core issue lies in determining the extent to which tools like ChatGPT, OpenAI’s Sora 2 video generator, Google’s Gemini, and Google’s Veo3 video tool rely on someone else’s art to arrive at their own inventions, and whether the use of source material constitutes copyright infringement.
Professionals and creative industries, including authors, filmmakers, artists, musicians, and newspaper publishers, are demanding compensation for the use of their works to build these models, and that the practice stop until they grant permission. They also argue that their works are being used without compensation to build AI tools that create works in direct competition with their own.
Vermillio platform: shedding light on closed models
One of the main points of contention is the heavily guarded models of AI giants, which make it difficult to know how much their technology relies on the works of other creators. However, one company claims it can shed light on this issue.
American technology platform Vermillio tracks clients’ intellectual property usage online and claims it is possible to track the percentage to which an AI-generated image relied on pre-existing copyrighted materials.
Neural fingerprint: how does the technology work?
In research conducted for the Guardian, Vermillio created a “neural fingerprint” for various copyrighted works before asking AI to generate similar-looking images. This fingerprint acts as a unique identifier for the original content, allowing Vermillio to compare it with AI-generated content and determine the degree of similarity.
For “Doctor Who,” Vermillio entered a prompt into Google’s Veo3 tool. The resulting “Doctor Who” video matched 80% with Vermillio’s “Doctor Who” fingerprint, meaning Google’s model relied heavily on the copyrighted work to produce its output.
OpenAI’s video, taken from YouTube and watermarked with OpenAI’s Sora tool, had an 87% match, according to Vermillio.
Other examples use a James Bond neural fingerprint. The Veo3 James Bond video had a 16% match when prompted for a “famous James Bond movie scene.” Meanwhile, the Sora video had a 62% match. Images generated by Vermillio using ChatGPT and Gemini had 28% and 86% matches, respectively. The examples also showed strong matches with Jurassic Park and Frozen.
Training data: the source of controversy
Generative AI models must be trained on a massive amount of data. The primary source for this information is the open web, which contains a wide range of data from Wikipedia contents to YouTube, press articles, and online book archives.
Kathleen Grace, chief strategy officer at Vermillio, whose client roster includes Sony Music, says: “We can all win if we pause for a moment and figure out a way to share and track content. This would incentivize copyright owners to release more data to AI companies and give AI companies access to more interesting datasets.”
Industry response and “fair use” defense
A Google spokesperson said their policies prohibit the infringement of intellectual property rights. However, Google-owned YouTube says its terms and conditions allow Google to use creators’ works to build AI models.
OpenAI said its models are trained on publicly available data, a process it claims complies with the U.S. legal principle of “fair use,” which permits the use of copyrighted work without the owner’s permission under certain circumstances.
Call to action and protecting creators
The Motion Picture Association trade group urged OpenAI to take “immediate action” to address copyright issues surrounding the latest version of Sora, citing videos showing copyrighted characters.
Baroness Beeban Kidron, a prominent figure fighting UK government proposals to amend copyright law in favor of AI companies, says that “it is time to stop pretending that theft is not happening.” She asks: “If Doctor Who and 007 cannot be protected, what hope is there for an artist working alone, without the resources or expertise to pursue global corporations that take their work, without permission and without payment?”
FAQs
What is the main problem between generative AI and copyright?
The problem is that AI models are trained on massive amounts of copyrighted content without permission or compensation, and then use this knowledge to generate new content that may closely resemble original works.
What is the Vermillio platform?
It is a technology platform that claims it can determine the extent to which an AI-generated image or video relies on pre-existing copyrighted material, using a technology called a “neural fingerprint.”
What does a high match percentage mean in Vermillio’s analysis?
A high match percentage means the AI-generated output closely resembles the original copyrighted work, indicating that the model relied heavily on that work to create the result.
How do AI companies defend themselves?
Companies like OpenAI argue that using publicly available data to train their models falls under the U.S. legal doctrine of “fair use,” which allows for limited use of copyrighted material for transformative purposes.
