Nick Buttrick
banner
nickbuttrick.bsky.social
Nick Buttrick
@nickbuttrick.bsky.social
Socioecological and cultural psychology. Inequality, guns, & bootstraps aka why America is strange. Asst prof at Wisconsin-Madison. Posts are not my own.
Understanding how given groups are represented in the broader culture is still interesting of course, but it is not necessarily the same as understanding ground-truth differences in how groups understand the world themselves. 9/10
January 23, 2024 at 2:24 PM
4) How do we know? LLMs are most powerful in studying hard-to-survey cultural groups or those with little prior data, but this is also where it can be hardest to know if the LLM is faithfully representing the beliefs of a set of people versus stereotypes on the internet. 8/n
January 23, 2024 at 2:23 PM
3) LLMs “know” cultural psychology. They may be learning from the cultural psychology literature itself and then repeating back to us things that psychologists have already discovered (in other words, the extant scientific literature could be influencing model output) 7/n
January 23, 2024 at 2:18 PM
2) Whose writing??? Being written ABOUT is not the same as doing the writing. LLMs may overrepresent stereotypes, especially where outgroup writings are more voluminous than writings from inside the group. 6/n
January 23, 2024 at 2:15 PM
1) The internet is not the world. What info makes it onto the internet is shaped by many forces, such as literacy, power, and the rewards that come with posting. Bias in, bias out. 5/n
January 23, 2024 at 2:10 PM
HOWEVER. LLMs are biased. 4/n
January 23, 2024 at 2:07 PM
LLMs can reproduce cultural differences: although WEIRD by default, they can be asked to respond ‘as if’ they were a person with certain properties (age, politics, etc). And their answers are remarkably accurate. 3/n
January 23, 2024 at 2:05 PM
Cultural psychologists have used compressed cultural artifacts (like Google Book’s nGram) to track cultural change over time. For instance, US culture shifted in viewing happiness as “luck” (in the 1800s) to more modern views. LLMs could let us do the same on a massive scale. 2/n
January 23, 2024 at 2:03 PM
Reposted by Nick Buttrick
Bob’s 1986 AER, from his dissertation, compared estimates from an experimental and an econometric evaluation of the same job training program.

It showed limits of econometric approaches and is one of the most influential validation exercises in labor.
www.jstor.org/stable/1806062
October 25, 2023 at 8:40 PM