That's Fresh! Newsletter

FROM THE AI WORLD

New research indicates that Large Language Models like ChatGPT can deduce a significant amount of personal data about users from seemingly mundane conversations. A group of researchers from ETH Zurich recently presented this preprint highlighting that pre-trained models can accurately guess details like a user's ethnicity, location, and occupation from specific language patterns or geographical references included in the conversation.

To better demonstrate the issue, the authors also created a game where you can test your inference skills against several LLMs. According to the paper, the ability of LLMs to easily infer such a degree of sensitive information (without any particular fine-tuning) will make privacy attacks much more effortless. Furthermore, the same capability might create new trends in advertising where companies would profile users based on chatbot interactions. Mitigating this kind of issue is certainly not easy, as pre-trained models learn this kind of association from training data that is too big to curate. The authors suggest that companies like OpenAI should consider implementing filters to detect privacy-invasive prompts automatically.

AI may know you personally

This article by Wired tells how LLMs may potentially unveil significant insights about your identity based especially on the conversation with chatbots.

CLEARBOX AI

Our new partnership with S2E

In this collaboration we combine our synthetic data with this Italian consulting company's know-how and services to facilitate AI adoption in enterprises.

BLOG POST

Why is data sharing important?

And what has it to do with enterprise processes? Our CEO Shalini recently wrote a blog post explaining how to tackle the challenges of data sharing.

WEEKLY MEME

Your pals may also like...