drivendataorg
Consider adding a set of LLM-specific considerations (or a worked example w an LLM)
We may want to elaborate on the default checklist items, create a new checklist, or add an example that focuses on ethical considerations that are unique to LLMs or that are particularly acute with LLMs. For example, considerations related to privacy (sending private data to an API endpoint, de-anonymization via embeddings), misinformation, hateful or offensive speech,...
Great idea to consider adding AI Guardrails to the checklist, especially for LLM development. Some considerations to be added to the checklist are defined here: https://github.com/guardrails-ai/guardrails.git
I think our existing list of questions is high level enough that it covers most of what needs to be discussed for LLMs. I like the idea of a tool-specific list of key ethical considerations, but I'm not sure where it would go in the current Deon structure. It may be a broader resource we want to add, and could include LLMs and other model types / tools - maybe we add another section under "custom checklists" with "Tool-specific considerations"
I'd break this down as :
- [ ] Add examples with LLMs (PR #168)
- [ ] Add a tool-specific section with LLM-specific harms
