Last year's survey helped us improve everything from the dates of the event to how the sandwiches on the saturday were provisioned, and I'm sure we'll hear about something we didn't notice this year!

And there were some “awards” of various dubiousness. #nbpy #excavacon

They also shared their slides! speakerdeck.com/annthurium/n... are @annthurium.bsky.social 's #nbpy slides. Thanks, Tilde!

Other types of bias seen here include hemisphere bias (like disproportionately snowy images for "January") , locality bias ("Asian" disproportionately gives East Asian), and gender bias derived from languages' default noun gendering. - @annthurium.bsky.social at #NBPy

Both Stable Diffusion and Dall-E refused to generate "sensitive" images like "a gay couple" 🙃 Firefly did a better job, though! - @annthurium.bsky.social at #nbpy

When you specify that discrimination is illegal and undesirable, multiple research teams found that several different models return less discriminatory results. - @annthurium.bsky.social at #nbpy

The Claude team ran an experiment like this, but didn't incorporate aspects of real-world discrimination like intersectionality. - @annthurium.bsky.social at #nbpy

There's not any one commonly-used means of measuring bias in AI model output, but one common strategy is to give the same prompt with different demographic implications (for example, "John Smith" vs "Jane Smith" both as the actor in an otherwise-identical sentence). @annthurium.bsky.social at #nbpy

We're starting with examples of prompts! Prompts are what you give the model to request it complete a task. You can include examples with strategies like "one-shot prompting" (one example), or use chain of thought prompting to ask for step-by-step reasoning. - @annthurium.bsky.social at #nbpy