Largely because they feared that large language models would be useful for large scale disinformation, election interference, impersonation, automated phishing, etc.

https://openai.com/index/better-language-models/ See "Policy Implications" and down.