The importance of high-quality data is paramount in the world of AI. As I have been told many many times on my research calls with the great and the good of AI, Garbage In, Garbage Out. Large Language Models (LLMs) require vast amounts of data to learn and generate human-like text. In this context, social media platforms like Reddit have emerged as valuable sources of training data, leading to a new era of data licensing agreements between AI companies and content providers.
Reddit, a popular website with over 82 million daily unique viewers and 16 billion posts and comments, has recently entered into licensing agreements with several AI companies, including Google and OpenAI. These deals, worth hundreds of millions of dollars, highlight the growing value of Reddit's user-generated content as a resource for training AI models.
Reddit's unique value as a source for AI training data lies in its dual role as a content repository and a guide to quality information across the internet. The platform's karma system, which allows users to upvote content, effectively curates a vast collection of potentially high-quality material.
The use of Reddit data in AI model training has significant implications across various fields. OpenAI recognised Reddit's potential as a content guide when developing GPT-2, as outlined in their 2018 paper. This approach has led to advancements in natural language processing and generation, enabling AI models to produce more coherent and contextually relevant text.
However, the use of Reddit data raises important questions about copyright law and data ownership. Unlike traditional media outlets where content is created by paid staff, Reddit's content is user-generated. This has created uncertainty regarding the legal rights of AI companies to use this data for training without explicit user consent.
Despite the potential benefits, there are challenges and limitations to using Reddit data for AI training. While Reddit's terms allow it to licence user-created content, they don't transfer copyright from users. In the UK, non-exclusive licensees face restrictions when suing for copyright infringement under specific legal conditions.
The lack of transparency surrounding data used to train recent AI models like GPT-4 and GPT-4o has raised concerns among content creators and legal experts. This opacity may be partly due to the attention and lawsuits alleging copyright infringement that followed the success of GPT-3.
The recent deals between Reddit and AI companies like Google and OpenAI signify a major shift in the relationship between content providers and AI developers. Licensing AI training data has rapidly become a significant business as the AI Arms race continues.