GPT-2: Too Dangerous To Release (2019)

In 2019, OpenAI initially withheld the full GPT-2 model (1.5 billion parameters) due to concerns about its potential for misuse, releasing only a smaller version as an experiment in responsible disclosure. After nine months of monitoring and learning, OpenAI released the full GPT-2 model, finding that while the outputs were convincing and could be fine-tuned for harmful purposes, there was no strong evidence of actual misuse. The experience informed OpenAI's approach to later models like ChatGPT, which incorporated safeguards against specific misuses such as impersonation.