Hi all,
I’m reaching out to ask for help clarifying if the project I am working on will fall within OpenAI’s usage guidelines before I invest more energy into it. I’ve read through the usage guidelines but still have some questions.
Here’s a brief description of what I’m working on: an application where users input the audio of a podcast that they have created and receive back a summary of the discussion, styled like an article. The tool would help podcasters produce written content from audio they’ve already created that they can easily share with audiences that like to read instead of listen.
I’ve made a fine-tuned model for doing this. Here’s a quick overview of the implementation: use another API to transcribe the audio. Batch it into chunks of ~1500 tokens. Use a fine-tuned Davinci model to translate each chunk of the discussion into text that is styled like an article and return up to 500 tokens. Then combine these to form the article summary.
There are some sections of the usage guidelines that I have questions about:
In “Article writing / editing”:
- "Please limit max user input tokens to 50. Including more tokens in the prompt is okay (for instance, examples), but the tokens user input should only be 50 (~200 characters), or slightly more at maximum.”
- “We generally do not permit tools that generate a paragraph or more of text, unless the output is of a very specific structure that couldn’t be re-purposed for general blog or article generation (e.g., a cover letter, a recipe, song lyrics).” in “Article writing / editing”
In “Blog intro paragraphs”
- “Make sure that your tool does not generalize to writing a paragraph on any topic - only an intro-like paragraph - as we do not permit open-ended generations of this length.” in “Blog intro paragraphs”
These sections sound like I may not be allowed to use GPT3 to do the type of podcast-to-article writing that I want. However, my understanding of the spirit of those rules is to prevent replication of Playground and “the misuse of tools to generate mis- and disinformation at scale”. User text input would be small (i.e. the url of where their audio is hosted), but they would have control over what is said in the podcast they upload. In that sense, they can’t directly replicate Playground but their words will influence model output. Additionally, it would be hard for users to abuse this tool to create mis/disinfo at scale because they would need to spend a lot of time creating a brand new podcast for each new article they wanted to create. It wouldn’t scale well if they are trying to cheaply spin up a bunch of content fast.
Given that, would this type of project be acceptable if I brought it to pre-launch review? I would be building in the recommended safety filters like rate limits, filters for unsafe content, and a human-in-the-loop to review output. If not, are there any modifications I could make to keep the general idea of the project while making it compliant? Thank you in advance for any help!