Already a time capsule, but back in January a bunch of us working at the intersection of the humanities and AI/ML came together to sketch out eight provocations from the humanities for genAI research. Here's a 🧵 1/ arxiv.org/abs/2502.19190 - ThreadSky

laurenfklein.bsky.social • 6 hours ago

Already a time capsule, but back in January a bunch of us working at the intersection of the humanities and AI/ML came together to sketch out eight provocations from the humanities for genAI research. Here's a 🧵 1/

https://arxiv.org/abs/2502.19190

Comments

triproftri.bsky.social•5 hours ago

Sharing all of this with the Responsible Computing Club @ SJSU (working through the Mozilla Foundation)!!

lsangha.bsky.social•5 hours ago

📌

smernit.bsky.social•6 hours ago

Excited to read this.

laurenfklein.bsky.social•6 hours ago

1. Models make words, but people make meaning. This one draws from C20 literary theory to explain how meaning is always context-dependent, and how the output of genAI models requires people to determine its significance 2/

laurenfklein.bsky.social•6 hours ago

2. AI requires an understanding of culture. This one unpacks two related yet distinct definitions of culture: the first is people as cultures, which is widely operational in genAI research; and the second which is objects or expressions of culture that emerge from a particular group, which is not 3/

laurenfklein.bsky.social•6 hours ago

3. AI can never be “representative.” This one draws from Black feminist and archival theory to reframe bias as an effect of unequal structural power. We show how focusing on the social, historical, and political conditions that produced the bias can be more illuminating than trying to fix it 4/

laurenfklein.bsky.social•6 hours ago

4. Bigger models are not always beer models. This one comes right from @zephoria.bsky.social and @katecrawford.bsky.social's original "Critical Questions" paper, just reframed for the landscape of genAI and supplemented with lots of good digital humanities examples 5/

laurenfklein.bsky.social•6 hours ago

5. Not all training data is equivalent. This one is related to #4, and points to practices from the humanities for detailing, contextualizing, and otherwise documenting the data creation process 6/

laurenfklein.bsky.social•6 hours ago

6. Openness is not an easy fix. Here's another one that should be familiar to folks #onhere. We expand the convo to consider the long history of data extraction and why we might want to be careful with open models, as well as current arguments in favor of open models for research use 7/

flootzavut.bsky.social•2 hours ago

📌

Comments

Posting Rules

Reply