Oh right, here's an update, for those curious on the role of AI LLMs (they're fucking LLMs, "AI" is inaccurate and used for novelty gdi) in domestic violence social services:
No California agency in attendance is using LLMs for direct service* (yet), THANK GOD
More admin than I expected DO use LLMs for things like grant proposals (FUCK) and more admin-oriented things like strategic planning; however, everyone did admit that LLMs are limited in effectiveness for grant proposals and require subsequent editing (NO SHIT)
EVERYONE reported survivors using LLMs, partly for DV education (oh god) but mostly for legal purposes (oh god); all of us direct service folks reported seeing how LLMs were fucking people over in their self-representation in legal processes (I have so many thoughts on this as a legal services supervisor, but I'll tuck those aside for now)
Tbh I was the most vocally negative and offered the following points lmao: LLMs cannot make moral decisions and can never be held accountable; LLMs risk increasing social isolation where DV already thrives in isolation; publicly accessible LLMs are run by corporations whose goal is profit, not safety or human values, and so by necessity they will always be in conflict with the purpose of our services; LLMs fuck pro se litigants over; confidentiality is unsecured against regular internet leaks and court mandate; LLMs exaggerate preexisting biases in input and risks marginalizing survivors even further, especially survivors already living on the margins; we cannot have land acknowledgments in our email signatures while simultaneously driving the push for more data centers in a state that's already burning for lack of water, guys, c'mon, wtf.
An advocate from an org serving South Asian survivors from a huge variety of cultural and linguistic backgrounds pointed out the problem of training LLMs to be culturally responsive and to make translation choices based on intent and meaning, not literalness, which threatens a survivor's safety in accessing resources and navigating an English-speaking legal system.
At the end of the day, the majority of attendees weren't exactly excited to implement LLMs more broadly, but we all had to acknowledge that we need SOME KIND of policies in place since survivors and some employees are using LLMs regularly. I argued that IF the industry moves towards greater adoption of LLMs, then it should be developed by the CPEDV**, divested from any corporate interests whatsoever, and trained and exhaustively tested by experienced DV advocates with thorough policy guidelines as guardrails.
We'll see what happens. 🫠
* I know of two AI chatbots specifically for domestic violence, one of them including the National Domestic Violence Hotline in the US. I have not used them and cannot provide insight into how they're trained or what their confidentiality protocols are like.
** CPEDV = California Partnership to End Domestic Violence. They're a state-level nonprofit meant to help steer the county-level DV orgs which provide direct services, offering policy guidance and professional development for DV org staff across the state.
