LLM privacy policies are difficult to read, study finds

Per Tech Radar, a recent study has highlighted significant challenges in understanding the privacy policies of large language models (LLMs), with many being excessively long and difficult to comprehend.

A cybersecurity firm, Bridewell, assessed the privacy policies of 20 popular LLMs and found the average reading time to be 20 minutes, with some policies exceeding 14,000 words. The average Flesch Reading Ease Score for these policies was 40.2, indicating a high level of jargon and complexity that makes them largely impenetrable for the average user. The study also revealed that 13 out of the 20 LLMs use user inputs and outputs for model training, and opting out of this process is not always clear or possible. This raises concerns for both individuals and businesses, as employees might inadvertently share sensitive or confidential information that could be used to train AI models.

Experts advise users to fully understand data processing by LLMs and recommend businesses implement clear internal guidelines and enterprise accounts with appropriate protections to mitigate risks.

Source: Tech Radar