OpenAI, the Microsoft-supported research lab, has acknowledged that it is “impossible” to develop high-quality neural networks without utilizing copyrighted material owned by individuals. Despite claiming to lawfully acquire such content for training its models, OpenAI asserts that using public domain material would result in inferior artificial intelligence (AI) software.
The issue of copyright infringement within the machine learning field has recently come to the forefront. In a report by the Institute of Electrical and Electronics Engineers (IEEE), it was revealed that Midjourney and OpenAI’s DALL-E 3, two prominent AI services, were able to recreate copyrighted scenes from movies and video games based on their training data. This raises legal questions surrounding the liability of AI vendors and their customers.
Gary Marcus, an AI expert, and Reid Southen, a digital illustrator, co-authored the report, which documented numerous cases of “plagiaristic outputs” from OpenAI and DALL-E 3. They strongly suggest that both companies trained their AI image-generation models on copyrighted material without the knowledge or consent of the copyright holders.
Furthermore, the report highlights a lack of transparency from OpenAI and Midjourney regarding the training data used for their AI models. Users are unaware if they are infringing on copyright or trademarks when producing images using these systems.
The issue of copyright infringement extends beyond digital artists. The New York Times has filed a lawsuit against OpenAI due to its ChatGPT text model reproducing verbatim copies of the newspaper’s paywalled articles. Similar claims have been made by book authors and software developers.
While OpenAI denies the validity of the New York Times lawsuit, it acknowledges that any generation of infringing content by its neural networks would be considered a “bug.” The company defended its practices, emphasizing collaboration with news organizations and the application of fair use defense under copyright law.
Legal experts argue that generative AI models should not be solely held responsible for plagiaristic outputs. They assert that the individuals prompting the AI model to reproduce copyrighted material also bear primary responsibility. Additionally, the prevalence of specific images in the training data set influences the likelihood of their reproduction by AI models.
Addressing concerns, experts suggest implementing measures to restrict the behavior of AI models. However, they caution against placing the burden solely on AI creators, as the distribution of images by copyright owners for publicity purposes indirectly contributes to the imitation of those images by AI models.
The ongoing debate surrounding copyright infringement and the responsibilities of AI creators and users poses complex legal dilemmas that will require careful resolution to strike a balance between innovation and intellectual property protection.
The source of the article is from the blog motopaddock.nl