Your AI Can't See Gorillas
owl writes:
https://chiraaggohel.com/posts/llms-eda/
One paper that caught my attention a bit ago was Selective attention in hypothesis-driven data analysis by Itai Yanai and Martin Lercher. In their study, students who were given specific hypotheses to test were much less likely to notice an obvious "gorilla in the data" compared to students who explored the data freely.
[...] I use large language models relatively often to assist with smaller portions of my daily bioinformatics work and have been interested in studying their ability to perform complete bioinformatics analyses. A key part of any analysis is exploratory data analysis (EDA), and I wondered how well large language models would perform at this task. This naturally begs the question: are large language models able to notice the "gorilla in the data" given the same prompts given to human students?
[...] Furthermore, their data analysis capabilities seem to focus much more on quantitative metrics and summary statistics, and less on the visual structure of the data. In some ways, this could be seen as a feature rather than a bug. While humans are hard-wired to see faces in clouds and trends in random noise, these models appear to err in the opposite direction.
I have a few thoughts on potential implications:
First, it suggests that current LLMs might be particularly valuable in domains where avoiding confirmation bias is critical. They could serve as a useful check against our tendency to over-interpret data, especially in fields like genomics or drug discovery where false positives are costly. (But also it's not like LLMs are immune to their own form of confirmation bias)
However, this same trait makes them potentially problematic for exploratory data analysis. The core value of EDA lies in its ability to generate novel hypotheses through pattern recognition. The fact that both Sonnet and 4o required explicit prompting to notice even dramatic visual patterns suggests they may miss crucial insights during open-ended exploration.
Read more of this story at SoylentNews.