GPT-2 6-month follow-up

OpenAI has released the 774 million parameter version of GPT-2 following a staged release process that began with the 124M model in February 2019 and the 355M model in May 2019. This release is part of a broader effort to balance the benefits of open research with the risks of misuse, such as the generation of convincing synthetic disinformation.

Key Lessons on Model Release and Misuse

OpenAI's experience with GPT-2 highlights three primary challenges in the deployment of large-scale generative models:

1. The Difficulty of Coordination

While multiple organizations have the capability to train models of 1.5 billion parameters or more, public release of such models has been limited. OpenAI notes that coordination between organizations regarding publication norms is difficult due to the proprietary nature of these systems and the lack of clear inter-organizational communication channels.

2. Human Susceptibility to Synthetic Text

Synthetic text is increasingly capable of deceiving humans. Research from Cornell University indicates that 72% of a test cohort judged GPT-2 synthetic text samples as credible, compared to 83% for real New York Times articles. Additionally, research from the Allen Institute for AI (AI2) and the University of Washington found that the "GROVER" system could produce news that is more plausible than human-written propaganda.

3. The Complexity of Text Detection

Detecting AI-generated text is a significant technical challenge. To be effective in real-world scenarios, detectors must achieve accuracy rates between 99.9% and 99.99% with very few false positives. Current ML-based methods typically only achieve accuracy in the low to mid-90s, and this accuracy further decreases when models are fine-tuned. OpenAI concludes that statistical detection must be supplemented by human judgment and metadata to effectively combat misuse.

Research Partnerships and Safety Analysis

To inform the potential release of the full-size 1558M parameter model, OpenAI partnered with four research organizations to analyze the 774M and 1558M models:

• Cornell University: Studying human susceptibility to digital disinformation.
• The Middlebury Institute of International Studies (CTEC): Exploring potential misuse by terrorists and extremists.
• The University of Oregon: Developing "bias probes" to analyze internal model bias.
• The University of Texas at Austin: Studying the statistical detectability of outputs after domain-specific fine-tuning and detection transfer across models.

Future Release Strategy

OpenAI's current plan is to release the 1558M parameter model in a few months, though this timeline remains contingent on the findings from research partners and the observed usage of the 774M model. OpenAI advocates for a combination of staged releases and partnership-based model sharing as a foundation for responsible AI publication, particularly as the challenges inherent to powerful generative models are expected to grow over time.