A recent study from a team of AI researchers has cast doubt on the transparency of so-called open large language models (LLMs). The researchers, hailing from prestigious institutions including Cornell University, the Signal Foundation, and the Now Institute, published their findings in the journal Nature. Their central argument is that the claims of openness made by creators of popular LLMs, such as ChatGPT, may not fully align with reality.
According to the researchers—David Widder, Meredith Whittaker, and Sarah West—merely making the source code of an LLM available to the public does not equate to true openness. A significant issue arises from the lack of access to the underlying training data that is essential for understanding how these models function. Furthermore, the researchers point out that most developers lack the necessary resources to independently train their own LLMs, which limits the ability to fully engage with the technology.
As LLMs have gained traction over recent years, their increasing popularity has raised concerns among both the general public and professionals in various fields. Questions surrounding the implications of AI research have emerged, including fears of privacy erosion, job displacement, and the challenge of distinguishing between authentic and AI-generated content. While these concerns remain largely unanswered, the creators of LLMs have made efforts to appear more transparent by offering their models for public access.
Users can visit the websites of LLM creators to view or download the source code, with the possibility of modifying it for their own uses. However, the authors of the study argue that this does not constitute genuine openness. They emphasize that the source code for an LLM differs fundamentally from that of traditional software applications, such as word processors. In the case of a word processor, downloading the code provides everything necessary to use and modify the software effectively. In contrast, downloading an LLM only grants access to the code, without the accompanying knowledge that results from the extensive training conducted by its creators.
The researchers identify three primary factors that influence the perceived openness of current LLMs:
- Transparency: The degree to which the LLM’s creators disclose information about their systems varies significantly. For instance, the developers of Llama 3 restrict access to their model through Application Programming Interfaces (APIs), which the authors label as a form of “openwashing.” This term refers to the practice of presenting a product as open and accessible while implementing restrictions that inhibit true engagement.
- Reusability: The usability of open-source code is contingent upon its design and implementation. If the code is not well-structured or documented, it may pose challenges for developers attempting to utilize it effectively.
- Extensibility: This factor addresses how easily the code can be adapted or expanded upon. The ability to build upon existing models is crucial for fostering innovation and collaboration within the AI community.
The researchers’ findings highlight the complexities surrounding the concept of open-source LLMs. As these models become increasingly integrated into various sectors, understanding the true nature of their openness is essential for developers and users alike. The implications of their findings could influence future discussions on AI transparency and accessibility, prompting a reevaluation of what it means for a technology to be ‘open.’
This ongoing dialogue about the openness of LLMs is vital as society navigates the rapidly evolving landscape of artificial intelligence. With the potential for AI technologies to significantly impact numerous aspects of daily life, the need for clarity and transparency in their development and deployment has never been more critical.
