Future of Web-Scale Training Sets: Unpacking Data Poisoning Concerns

[Frabjous]

I read an article in the April 6 edition of The Economist (regretfully behind a paywall) about poisoned datasets. Here’s an arxiv article it referenced.
https://arxiv.org/abs/2302.10149

What is the future of web-scale training sets? Is data poisoning a start-up pang, a long-term issue or an overreaction.

 

[jack action]

Data poisoning started way back when with keyword stuffing to trick search engines. Nothing new here. If the public web is the source, AI companies will have to program their AI to avoid certain patterns like search engines already do today.

So I guess my opinion is a mix of long-term issue and overreaction.