The topic of copyright issues in relation to AI continues to draw attention and spur discussions among technology companies and copyright holders. OpenAI recently submitted a proposal to the British House of Lords, urging the updating of copyright laws to accommodate the use of copyrighted material in training AI models. According to OpenAI, without the ability to use copyrighted material, their AI model, ChatGPT, would not be able to exist. They argue that relying solely on text in the public domain would limit the effectiveness of AI models in meeting the needs of users.
Critics, however, challenge OpenAI’s standpoint, pointing out that licensing copyrighted material for training purposes is a viable alternative. They argue that OpenAI’s claim of it being impossible to create AI models without using copyrighted material is disingenuous. Other companies have successfully developed generative AI models without resorting to unauthorized use of copyrighted material.
In the midst of these discussions, OpenAI also faced a copyright infringement lawsuit from the New York Times. OpenAI defended itself in a blog post, claiming the lawsuit was without merit. They maintained that the New York Times’ content did not contribute significantly to the training of their existing models and would not be impactful for future training either.
One of the challenges in negotiating these copyright issues lies in different perspectives on the value of data. Tech companies like OpenAI argue that they should only pay for the marginal information value that data provides the AI models. However, copyright holders are concerned about potential revenue loss if AI models replace traditional sources of information, such as news websites. Bridging this gap in expectations will require adjustments on both sides, much like what occurred in the music industry with the rise of streaming services.
Additionally, the issue of regurgitation or plagiarism in generative AI models has arisen. OpenAI argues that instances of regurgitation, where models produce verbatim copies of copyrighted content, are relatively rare and may be influenced by the repeated ingestion of specific text during training. However, recent research has shown that regurgitation is not an uncommon occurrence in many AI models, including those developed by OpenAI. Researchers demonstrated how easy it is to prompt AI models to generate images and text that closely resemble copyrighted content without explicitly instructing them to do so.
As copyright issues surrounding AI continue to unfold, there is a need for careful consideration and negotiation to strike a balance between protecting intellectual property rights and promoting advancements in AI technology.
The source of the article is from the blog japan-pc.jp