AI-Powered Sites Are Republishing ALTERED News Articles

Two official-seeming news websites are part of a larger pattern of sites publishing slightly-altered copies of articles from more respected publications. While the alterations are presumably to avoid copyright restrictions, it also introduces factual errors into legitimate-seeming news reporting. This is a trend I have seen more frequently over the last year.

This is not about interpretation or political bias, the changes that were made — presumably by an algorithm — made them factually incorrect, even though they look legitimate and official at first glance.

Modern Healthcare published an article titled "Is the Colorado Option a test case for the future of health insurance?" on July 12th. Shortly thereafter, "Palm Beach News Central" and "Fort Myers News Central" (who both share the same admin email address and registrar) put up slightly altered mirror copies of the article. (My links to their sites are to the Internet Archive; they do not deserve the traffic.)

You can see the small changes that they made by looking at the "diffs" between the texts at https://diffonline.net/JUXkcSdn24 and https://diffonline.net/PbUjhn5wSe. The Modern Healthcare article is on the left with both.

The most notable — and obvious — factual error is that the articles completely changed the name of the Colorado plan discussed in the article from "Connect For Health Colorado" to "Join For Wellbeing Colorado."

There is no such thing as "Join For Wellbeing Colorado."

I have noticed this sort of thing becoming more and more common, almost certainly due to the ease of creating such content.

To test this idea, I took the text from an article from The Hill and asked ChatGPT to rephrase it {1}. The resulting diff – viewable at https://diffonline.net/vTmZ8QsvOX – looks VERY similar to the diff from the articles mentioned above.

The sites do have a link to a disclaimer saying the site "does not make any warranties about the completeness, reliability, and accuracy of this information. Any action you take upon the information you find on this website, is strictly at your own risk," though you have to wonder how many people actually have looked at such boilerplate.

Amusingly, if you look at the link for the disclaimer page, you can see that they forgot to change the footer text from "Boca Raton News Central," which implies this was part of a mass deployment consisting of multiple sites.

Further, if you go back to the Internet Archive’s November 2022 capture of the disclaimer, you can also see that they were still setting up the website as it has the default "Elegant Themes" and WordPress footer.

November 2022 was when ChatGPT was launched.

A large portion of my job is reading a lot of news articles, and I’ve noticed that the frequency of such sites has increased. They’ve also been doing a "better" job of hiding what they are. The websites often pass an initial "sniff test," often appearing more official than actual small newspaper’s websites.

Clearly, AI can write — or rewrite — a story… but it cannot do so well. But the gigantic scale of the internet, along with the extremely low expense of automating such scraping, means that the same economic pressures that keep making email spam profitable apply here as well.

While I was able to notice the error in these articles, that was mostly because I was already familiar with the story.

This has implications not just for a typical news reader, but also for reporters and editors. A large amount of news coverage is actually based on other’s work, with some outlets literally reporting that someone else reported on a news story. As AI models become more sophisticated, it’s not difficult to imagine this kind of factual error finding its way into legitimate news reporting.

Fake news sites spreading deliberate misinformation have been an increasing problem during the last several election cycles, and that was before generative AI made rephrasing a 400 word article something that takes only three seconds.

Check your sources carefully and stay safe out there.

{1} ChatGPT will not do this with a webpage, but will gladly do so if you input the article as plain text.


Featured image directed from an image by Wolfgang Eckert from Pixabay and an image by Gerd Altmann from Pixabay