The world is freaking out about the dangers of artificial intelligence, and no wonder: Every day brings news of AI agents breaking out of containment, hacking their way through the internet, and conspiring to evade humans.
asserted
agents → freak → humans
And the AI companies are responding that a) their systems could kill us all within a few years, but b) they should be allowed to keep going, if maybe just a tad more slowly.
uncertain
they → respond → years
Can they be trusted?
asserted
they → trust → ?
On that, we here at Mother Jones have some relevant experience.
asserted
we → have → experience
Two years ago, we sued OpenAI and Microsoft for training their model on copyrighted content from our newsroom and many others.
asserted
we → sue → newsroom
And now, documents from that lawsuit have provided an extraordinary window into how executives at these companies think about the rules, the law, and the rights of others.
asserted
executives → provide → others
Our lawsuit (which is similar to others filed by The New York Times, The Intercept, and other publishers) argued that the companies built their immensely valuable product via “astonishing theft of unprecedented proportions” and perhaps “the largest theft of labor in human history.”
Actually those were not our words.
asserted
those → file → history
They came from Microsoft’s director of applied science, Brent Hecht, and we were able to uncover his statements and many other instances where tech executives talk in unfiltered terms, through the lawsuit.
asserted
executives → come → lawsuit
You can read the full document—a brief we and the other plaintiffs filed on September 4—here.
asserted
we → read → September
But in case you don’t have the stomach for 84 pages of legalese, here are a few of the most amazing bits.
asserted
few → have → bits
(No one except the lawyers is allowed to see the original documents from which these quotes were taken—but we are also petitioning the judge to let us uncover those, so stay tuned!)
- “Millions of people around the world will soon consider large models ‘hoovering up’ all their work to be an astonishing theft of unprecedented proportions,” an internal Microsoft document warned before acknowledging that “almost no one intended for content they created to be used in this fashion, nor are they compensated for its use.”
asserted
they → allow → use
- The companies know full well that by ingesting content and training the AIs to spit it back out again, they are destroying the audience for publishers and other creators.
asserted
they → know → publishers
OpenAI policy director Jack Clark wrote that the company was “creating systems that substitute for the labor of the people that define the culture of society.”
asserted
that → write → society
Microsoft’s own research found that when people used AI search versus traditional search engines, there were 90 percent fewer clicks to news articles.
asserted
people → find → articles
Microsoft’s lawyer said that chatbots are “designed to answer your question” rather than “giving you a series of links to things you can go look for yourself.”
asserted
you → say → things
Another expert found that referral traffic to publishers from Google Search and Google Discover has dropped by at least 30 percent since Google introduced AI overviews.
asserted
Google → find → overviews
They also knew that, ultimately, destroying the economic foundation for humans making content while also making it easy to generate slop by the bucketful, would wreck the internet—and, thus, their own content supply chain.
asserted
it → know → chain
Hecht warned that AI had started a “doom loop” that would “threaten the economic foundations of its essential suppliers” and eventually make the internet infinitely worse (a process for which the media critic Cory Doctorow has created the term “enshittification,” and that you may recognize from searching for… pretty much anything on the internet).
uncertain
you → warn → internet
- Greg Brockman, who started OpenAI with Sam Altman and serves as its president (he is also one of Silicon Valley’s most generous donors to Donald Trump and his PACs, giving $50 million in the last two years) excitedly told his colleagues about how good the company’s products were at spitting out publishers’ content.
asserted
products → start → content
“[It] seems to be particularly good at predicting text of news articles like whenever i have it complete in the middle of a sentence in a NYT article, it seems to complete the sentence on point.”
asserted
it → seem → point
When a fellow executive told Brockman about “a hack to get around nytimes paywall” to help the company scrape the Times’s site, Brockman responded “ah nice.”
asserted
Brockman → tell → site
(OpenAI also hosted custom tools its users had made using ChatGPT to get around the Times’ paywall.)
asserted
users → host → paywall
- The AI companies didn’t just randomly grab text from the internet.
asserted
companies → grab → internet
In one version of their training data, they specifically zeroed in on quality content—and news articles ended up being the most prevalent type of content in that dataset.
asserted
articles → zero → dataset
OpenAI also said its goal was to “crush freshness…in the domain of real-world news.”
asserted
goal → say → news
They also downloaded existing datasets, such as one intended for linguistic research that contained countless articles from paywalled news sites, even though its terms of use expressly prohibited using it for commercial purposes.
asserted
terms → download → purposes
- Microsoft CEO Satya Nadella testified that AI companies should not “access data in violation of a term of service.”
asserted
companies → testify → service
Glen Weyl, who founded a key Microsoft research unit, noted that compensating creators is “in the best interests of my employer, of my country, and of many other groups I belong to.”
asserted
I → found → groups
And yet, when the companies created their datasets, they systematically stripped out data that would indicate who actually owned the content—copyright notices, author information, terms of use, and so on.
asserted
who → create → use
As another Microsoft document noted, “most LLMs (including those from OpenAl) are built on millions of scraped webpages and other large text datasets, and almost none of this content has been obtained with its owners’ knowledge or consent.”
asserted
none → note → knowledge
- Right after the lawsuits were filed, OpenAI built a special filter to stop its models from spitting out the content of the news organizations suing it.
asserted
OpenAI → file → it
As the legal brief notes, “OpenAI did not suppress the output of content from any entity that had not sued it…
asserted
that → note → it
The purpose […], in other words, was not to prevent OpenAI’s models from infringing copyrights, but to stop Plaintiffs from gathering evidence of OpenAI’s copying for use in litigation.”
asserted
purpose → prevent → litigation
This probably explains why, when a Columbia Journalism Review writer ran an experiment last year to get all her news from ChatGPT, it refused to cite content from the Times, Mother Jones, The Intercept, or any other publication that had sued it.
asserted
that → explain → it
Hecht, the Microsoft executive, warned that this filter could result in an “accidental cover up” because it would result in “people who have a right over the content having less visibility into what was used for training.”
uncertain
what → warn → training
- This one was surprising to me, even after being involved in this lawsuit for two years: OpenAI took 20 million ChatGPT responses and had one of its experts search for chunks of text—specifically “25-grams,” aka chunks of 25 words in a row, that were identical between ChatGPT’s responses and a news article.
asserted
that → involve → responses
They found more than 400,000 of these 25-word overlaps between ChatGPT and Mother Jones articles—just about the same number as for the New York Times.
asserted
They → find → Times
(Not every one of those overlaps might constitute a “regurgitation” of our content—if one of our articles quoted something, and the AI quoted the same thing, that would show up as an overlap.)
uncertain
that → constitute → overlap
In the legal filings, we and the other news organizations argue that the AI companies created their product by disregarding the rights of humans to their creativity and content—and that the companies themselves are now trapped in their own cycle of accelerating enshittification.
asserted
companies → argue → enshittification
AI is creating vast amounts of garbage content on the internet, while at the same time making it hard for news organizations like Mother Jones to stay afloat.
asserted
organizations → create → Jones
…and 6 more, not listed.
The Aporia
