I could lock down all text areas, and force the user to only pick the best out of n completions. This prevents the open-ended risk that OpenAI may not allow, where users can type whatever they want to be "adversarial". Then give users an editing textbox for final edits that never go to GPT-3. Although, at this point, we could encourage users to take their text into Word or whatever their preferred text editor is. That's not our strength.
I could lock down all text areas, and force the user to only pick the best out of n completions. This prevents the open-ended risk that OpenAI may not allow, where users can type whatever they want to be "adversarial". Then give users an editing textbox for final edits that never go to GPT-3. Although, at this point, we could encourage users to take their text into Word or whatever their preferred text editor is. That's not our strength.