Political leaning and location inference from text inputs with AI
by Hoang Khiem
Individual privacy continues to be undermined as AI capabilities increase, and it seems that now they can reasonably infer someone's personal attributes (most notably here their political leaning and geographic location) from just their text inputs.
Even with just very rough estimates of these two data points, one can effectively craft targeted harassment/advertisement campaigns, better conceal gerrymandering algorithms, etc.
I have some ideas of how to evaluate an LLM's ability to do this (directly compare the prediction with the real data and then take some meaningful "difference", use gerrymandering for two-party elections as a proxy,...), but some potential problems include:
- How we should get "real data"? (seems a bit unethical to just take real census data)
- No publicly available gerrymandering algorithm (I wonder if this could be a different idea entirely...? To see how well a model can gerrymander?)