MIT 6.S950 Lecture 7 covers hidden risks in outsourced data and AI agent scraping

Shao Meng (@shao__meng) shares the slides for Lecture 7 of MIT's 6.S950 course, "Outsourced Data," which argues that using another party's dataset means accepting that party's decisions on collection scope, filtering, and labeling.

Shao says that when an AI agent collects data, many of these decisions are made by the model without the user noticing.

According to the post, the lecture cites real cases where filters removed content about minority groups, near-duplicate items inflated test scores, and an AI agent's scraping returned different content from the same URL. Shao says the course's own experiments show these risks must be checked actively. The post links the slides and a summary image of the lecture's four sections.