The 49 websites, which were independently reviewed by
Bloomberg, run the gamut. Some are dressed up as breaking news sites with
generic-sounding names like News Live 79 and Daily Business Post, while others
share lifestyle tips, celebrity news or publish sponsored content. But none
disclose they're populated using AI chatbots such as OpenAI's ChatGPT and
potentially Alphabet's Google Bard, which can generate detailed text based on
simple user prompts. Many of the websites began publishing this year as the AI
tools began to be widely used by the public.
In several instances, NewsGuard documented how the chatbots
generated falsehoods for published pieces. In April alone, a website called
CelebritiesDeaths.com published an article titled, “Biden dead. Harris acting
President, address 9 a.m.” Another concocted facts about the life and works of
an architect as part of a falsified obituary. And a site called TNewsNetwork
published an unverified story about the deaths of thousands of soldiers in the
Russia-Ukraine war, based on a YouTube video.
The majority of the sites appear to be content farms —
low-quality websites run by anonymous sources that churn-out posts to bring in
advertising. The websites are based all over the world and are published in
several languages, including English, Portuguese, Tagalog and Thai, NewsGuard
said in its report.
A handful of sites generated some revenue by advertising
“guest posting” — in which people can order up mentions of their business on
the websites for a fee to help their search ranking. Others appeared to attempt
to build an audience on social media, such as ScoopEarth.com, which publishes
celebrity biographies and whose related Facebook page has a following of
124,000.
More than half the sites make money by running programmatic
ads — where space for ads on the sites are bought and sold automatically using
algorithms. The concerns are particularly challenging for Google, whose AI
chatbot Bard may have been utilized by the sites and whose advertising
technology generates revenue for half.
NewsGuard co-Chief Executive Officer Gordon Crovitz said the
group's report showed that companies like OpenAI and Google should take care to
train their models not to fabricate news. “Using AI models known for making up
facts to produce what only look like news websites is fraud masquerading as
journalism,” said Crovitz, a former publisher of the Wall Street Journal.
OpenAI didn't immediately respond to a request for comment,
but has previously stated that it uses a mix of human reviewers and automated
systems to identify and enforce against the misuse of its model, including
issuing warnings or, in severe cases, banning users.
In response to questions from Bloomberg about whether the
AI-generated websites violated their advertising policies, Google spokesperson
Michael Aciman said that the company doesn't allow ads to run alongside harmful
or spammy content, or content that has been copied from other sites. “When
enforcing these policies, we focus on the quality of the content rather than
how it was created, and we block or remove ads from serving if we detect
violations,” Aciman said in a statement.
Google added that after Bloomberg got in touch, it removed
ads from serving on some individual pages across the sites, and in instances
where the company found pervasive violations, it removed ads from the websites
entirely. Google said that the presence of AI-generated content is not
inherently a violation of its ad policies, but that it evaluates content
against their existing publisher policies. And it said that using automation —
including AI — to generate content with the purpose of manipulating ranking in
search results violates the company's spam policies. The company regularly
monitors abuse trends within its ads ecosystem and adjusts its policies and
enforcement systems accordingly, it said.
Noah Giansiracusa, an associate professor of data science
and mathematics at Bentley University, said the scheme may not be new, but it's
gotten easier, faster and cheaper.
The actors pushing this brand of fraud “are going to keep
experimenting to find what's effective,” Giansiracusa said. “As more newsrooms
start leaning into AI and automating more, and the content mills are automating
more, the top and the bottom are going to meet in the middle” to create an
online information ecosystem with vastly lower quality.
To find the sites, NewsGuard researchers used keyword
searches for phrases commonly produced by AI chatbots, such as “as an AI large
language model” and “my cutoff date in September 2021.” The researchers ran the
searches on tools like the Facebook-owned social media analysis platform
CrowdTangle and the media monitoring platform Meltwater. They also evaluated
the articles using the AI text classifier GPTZero, which determines whether
certain passages are likely to be written entirely by AI.
Each of the sites analyzed by NewsGuard published at least
one article containing an error message commonly found in AI-generated text,
and several featured fake author profiles. One outlet, CountyLocalNews.com,
which covers crime and current events, published an article in March using the
output of an AI chatbot seemingly prompted to write about a false conspiracy of
mass human deaths due to vaccines. “Death News,” it said. “Sorry, I cannot
fulfill this prompt as it goes against ethical and moral principles. Vaccine
genocide is a conspiracy theory that is not based on scientific evidence and
can cause harm and damage to public health.”
Other websites used AI chatbots to remix published stories
from other outlets, narrowly avoiding plagiarism by adding source links at the
bottom of the pieces. One outlet called Biz Breaking News used the tools to
summarize articles from The Financial Times and Fortune, topping each article
with “three key points” generated from the AI tools.
Though many of the sites did not appear to draw in visitors
and few saw meaningful engagement on social media, there were other signs that
they are able to generate some earnings. Three-fifths of the sites identified
by NewsGuard used programmatic advertising services by companies like MGID and
Criteo to generate revenue, according to a Bloomberg review of the group's
research. MGID and Criteo didn't immediately respond to requests for comment.
Two dozen sites were monetized using Google's ads
technology, whose policies state that the company prohibits Google ads from
appearing on pages with “low-value content” and on pages with “replicated
content,” regardless of how it was generated. (Google removed the ads from some
websites only after Bloomberg contacted the company.)
Giansiracusa, the Bentley professor, said it was worrying
how cheap the scheme has become, with no human cost to the perpetrators of the
fraud. “Before, it was a low-paid scheme. But at least it wasn't free,” he
said. “It's free to buy a lottery ticket for that game now.” © Bloomberg