Researchers are creating incredible new advances in "artificial intelligence" almost every week. While computers do not exhibit generalized intelligence, they can use advanced statistics to look at relationships in extremely large bodies of work to generate text and images that, in some specific cases, can be very useful. (In other cases, they can be quite disappointing). We first used AI three weeks ago in a workshop, and got initial feedback from members. A week ago, we introduced an experimental feature called Neopoet AI. The feedback has been generally positive, but not universally so.
It's exciting to be the first poetry community to experiment with AI.
This is, of course, an experiment. I wanted to see if AI could help poets improve their work. Some experiments fail and some succeed. I expect we will continue to tune our approach based on feedback from members.
I wanted to mention a few things we've done so far to address concerns:
- The Neopoet AI feature currently uses an OpenAI model through an API. We've confirmed that, contractually, they cannot use submissions to train or improve their AI. We've been testing a number of different prompts and other variables, so response quality has varied.
- We've created an opt-out feature.
- We've refined the way we work with the model. On Neopoet, you've always been able to select prompts like "is my internal logic consistent", and we are now including these questions as questions for the model. This has created some unexpected issues, however. When we first introduced this feature, it caused the AI model to misunderstand our ask - unfortunately this led to a few responses that were nonsensical (we quicky fixed this issue). In the interest of transparency and to foster a full discussion, we haven't taken down the poor response examples.
- Our stance is that, ethically, it must be clear to all users that computer-created content is computer-created. We now have disclosures both before and after the computer-generated text, and a link to submit feedback. We've refined these based on community input.
We welcome your thoughts on Neopoet AI as this experiment progresses.