An illustration from the Cornell University paper shows how more copying is proven by comparing AI output against larger amounts of original content.Since earlier this year, the McKinley Park News has been raising a klaxon alarm about how the theft of our Service by Large Language Models (LLMs) and the "artificial intelligence" (AI) services powered by them undermine our local news enterprise and the practice of journalism.
Just last week, AI security researchers revealed they've proven what many have suspected: The LLMs "memorize" — or copy verbatim — the content they've stolen to use as training data. The security researchers were able to get the AI to spit out big chunks of text written exactly as presented in source training material by submitting specific, and somewhat nonsensical, queries.
Cornell University Paper
The paper "Scalable Extraction of Training Data from (Production) Language Models" from a research team at Cornell University provides technical specifics of the methods they've used to demonstrate this verbatim copying and republication. A supporting abstract from the research team and comments and notes from third parties (such as the video embedded below) provide more information.
Giant Plagarism Engines
This research demonstrates massive implications for working with AI, including security, confidentiality and, especially for our purposes, copyright and theft of service. The Cornell University team has illustrated how these LLMs have stolen and stored our and others' work product outright, which is then leveraged into for-profit services that compete directly aganist us while providing no remuneration.
There's no question now that AI services built on larcenous LLMs are nothing more than giant plagarism engines based on content stolen from publications like the McKinley Park News. But there are still few in our industry responsibly opposing this destructive theft.
We Demand a Moritorium
The McKinley Park News calls upon institutions of journalism currently engaged in promoting, training or teaching AI to put a moratorium on AI use unless the source data used to train the LLMs can be qualified.
Journalism institutions that don't refrain from trafficking in stolen goods are not only training others to steal from us, they're qualifying and participating in toxic activity that fundamentally undermines the industry and practice of journalism. By choosing big tech over journalism, they show which side they're on.
I'd like to personally thank the Cornell Research team for their important research that I think is going to have a significant and positive impact in the service of journalism. We appreciate your work!
— Justin Kerr
Publisher, McKinley Park News







Log In to comment on this item.
Subscribe
Report
My comments