In several recent matters, we have deployed AI to assist with analyzing important collections of content related to the claims at issue.
Open-vocabulary image object detection
In a consumer class action with product liability allegations, we supplemented keyword search analysis with an analysis of 100,000 images from social media and traditional media outlets. By using LLM analysis of image content at scale, we facilitated identification of different subjects relevant to the allegation. The purpose-built model created by the firm’s Data Science Center was smaller, faster, and simpler to produce than a larger, general-purpose LLM.
LLM-based text classification and topic modeling of product reviews relied on by expert
In support of an expert’s testimony, we employed an LLM-based topic modeling approach to product reviews. This approach featured academic support and—unlike traditional approaches—did not require specification of seed terms, pre-set topic numbers, or subjective human interpretation. This analysis demonstrated the heterogeneity of consumer preferences and lack of emphasis on the product feature at issue.
We applied that same LLM model to detect references to at-issue product traits and distinguish with simple references to the product name. To validate text classification, we repurposed rigorous inter-rater reliability standards found in the traditional, human-coded content analysis literature.
LLM sentiment analysis of social media posts
In a recent consumer fraud matter, we directed a general-purpose LLM to perform aspect-based sentiment analysis on hundreds of thousands of social media posts. When the opposing expert challenged our results using revised prompts, we demonstrated that our prompting strategy followed better practices, generated far fewer hallucinations, and—most importantly—yielded significantly more accurate output.