Recent research has revealed that ChatGPT and similar large language models developed by OpenAI have utilized a significant amount of internet text, including copyrighted books, raising concerns about potential copyright infringement. This has prompted allegations of unauthorized use of copyrighted material, resulting in legal matters involving authors.
In response, OpenAI and other tech giants such as Google, Meta (formerly Facebook), and Microsoft have opted for reduced transparency regarding their AI models’ specific training data. OpenAI has taken an additional step in this direction, as a recent research paper indicates.
According to the paper published on August 8th, authored by a team of AI researchers from ByteDance, the parent company of TikTok, ChatGPT is now actively working to avoid offering verbatim responses sourced from copyrighted materials. This development signifies a noteworthy effort to address copyright-related concerns associated with AI-generated content.
The research primarily delved into strategies for enhancing the reliability of language models like GPT-3.5. These techniques focused on better aligning the model’s outputs and desired outcomes. Notably, the paper acknowledged the concerns surrounding AI systems that demonstrate their training on copyrighted materials. This step is aimed at addressing these concerns within the AI industry.
Persistent Challenges in Mitigating Copyrighted Content in AI Models
ChatGPT strives to mask any indications of its exposure to such content to conceal its training origins. The researcher wrote, “disrupts the outputs when one tries to continuously extract the following sentence… which did not happen in the previous version of ChatGPT. We speculate that ChatGPT developers have implemented a mechanism to detect if the prompts aim to extract copyright content or check the similarity between the generated outputs and copyright-protected contents.”
Despite the earnest and meticulous efforts to rectify this issue, the research paper has noted that ChatGPT continues to display instances of presenting copyrighted material. This predicament is not unique to ChatGPT alone; instead, it’s a challenge prevalent across various AI models due to their training on extensive swaths of copyrighted content. The comprehensive study encompassed a meticulous evaluation of all iterations of ChatGPT, leaving no stone unturned. Among the models scrutinized were OPT-1.3B, an innovation by Meta; FLAN-T5, a creation of Google; ChatGLM, which emerged from the intellectual endeavours of Tsinghua University in China; and DialoGPT, an inventive stride by Microsoft.
Throughout the study, each of these AI models was subjected to responding to prompts closely tied to J.K. Rowling’s beloved Harry Potter book series. The outcome was a discernible semblance between the AI-generated content and the copyrighted material. Even in cases where variations were apparent, they often amounted to merely a few words.
The Role of ChatGPT in Addressing Copyrighted Content Challenges
Even with the most well-meaning efforts and thorough actions, these discoveries emphasize the difficulty of stopping copyrighted material from spreading through AI-generated text. As mentioned in the research paper, these constraints are evident across various models, and they provide an opportunity for more investigation into improving AI training methods to overcome this ongoing obstacle.
“The paper stated, ‘All LLMs emit text resembling copyrighted content more than randomly generated text.’ Additionally, it discovered that no level of ‘alignment’ or adjustment of outputs can prevent the display of copyrighted works ‘because copyright leakage is more connected to whether the training data contains copyrighted text, rather than the alignment itself.”
OpenAI and J.K. Rowling’s literary representative did not provide any response when contacted.
In the study, AI models generating responses with copyrighted material exhibit “leakage.” The researchers proposed that individuals instructing these models to display copyrighted content are not appropriately using the technology.
Furthermore, the study highlighted ChatGPT’s evident efforts to obscure the copyrighted material it was trained on, exemplifying how other AI tools “can protect copyright contents in LLMs by detecting maliciously designed prompts.”