🚨 New paper, at CSCW'24!



We examine how existing teams discuss ethics, and how their power dynamics limit what ethical critique can be raised.

And we ask: can speculative futures games help expand these limits? (Hint: 🚫No, not directly)

🧵... arxiv.org/pdf/2403.19049…
🃏 We observe 3 corporate teams and 1 activist group play a game designed to help ask: what could go wrong with AI?

🗣 After, we interview each person individually:  what did they feel able to say? how did this differ vs their teams' usual ethics discussions?

🚨New paper! 🚨

On *incentives* in AI/ML communities, why AI is … the way that it is. :/

This work is an "origin story" of LLMs as told by researchers around since long before the public eye and ✨hype✨

And many are conflicted… 🧵

📄pdf: arxiv.org/abs/2310.07715

https://twitter.com/_sireesh/status/1712467156120772745

For example, interview participants spoke of how they're incentivized to produce marginal improvements to get papers accepted, like a 2% model accuracy increase.

Numbers go brrrrr…

tip to new PhD students from a 5th year: *show your advisor tables*

it really doesn't matter what you put in the tables.

PhD advisors LOVE tables. you can put data in tables, and they love that, but don't stop there.

you can put action items, paper sections, RQs, related work, whatever. put it in a table. 100% success rate.

⛓What do modular software supply chains mean for "Ethical AI"?

Developers release discrete modules, or compose existing modules into finished AI systems.

@dawnnafus and I show that this makes it hard to imagine or take responsibility for AI harm.

📄: arxiv.org/abs/2209.09780 From interviews with developers building "fundamental" libraries, end-user products, & things in between, we show:

High in the "supply chain", harm feels remote and unimaginable.

Lower in the "supply chain", devs feel unable or unwilling to pay ethical debt accrued upstream.

.@clegoues and I spent the last 45 mins in lively conversation, primarily arguing what a software "bug" is.

Q: #SoftwareEngineering practitioners and researchers:
How would you define "bug"? @clegoues Please RT! and hottakes/informal defs are very welcome!

there are research sensors in our new CMU offices:

microphones measuring sound (not voice), 8x8 low res infrared camera, accelerometer detecting doors closing and other vibrations.

this is opt-OUT not opt-IN: data collected & is used for research, except mic, which is opt in. one use case the faculty member in charge of the research suggested was tracking if our offices are vacuumed by cleaning staff.

this seems like algorithmic management of lowly paid employees to me.