🧵 [1/n]
Super interesting study on an instance of the Algorithmic Imprint (https://twitter.com/UpolEhsan/status/1537112310505824256)-- people might retain biases from AI systems even when the AI system is no longer there. The spirit of the argument is well taken. What are some of the caveats you should pay attention to when trying to transfer insights from studies like these to real-world settings?
The study population: mostly students and crowd-workers. Critically reflect on how population-biases might be impacting results.
The study domain: a simulation of an image-based diagnosis. But why? The paper doesn't motivate how using a medical use case on lay people/non-experts (i.e., not doctors) is a valid way. Why didn't the paper use a different use case that could be more familiar with lay users?
🧵 [4/n]
No doubt judges are a hard population to get access-- What's a path forward?
Here's a starter pack (has yet to fail me):
If not judges, get the nearest match-- lawyers, paralegals, etc. Instead of using crowdworkers, lawyers and paralegals are better proxies.
Instead of "armchair scenario building", directly work with the community to co-design the scenarios such that they are grounded in reality (not fiction or just literature).
🧵 [end]
3. Instead of just deploying a controlled study, pair it up with a more mixed-methods approach-- use scenario-based design interviews to juxtapose the quantitative data with the qualitative findings.
PS> this is not a dunk on the paper. This is a well-written paper. The caveats highlighted are meant to help the broader RAI audience to be able to better transfer insights from research papers.
Add comment