Yr Wonkette got an object lesson yesterday — as if we needed one — in why AI search results should be treated with skepticism set on fire. While Robyn was working on her story on the Trump administration’s ugly (and like so many administration ideas, almost certainly illegal) proposal to shift US refugee policy to helping out white people who are “oppressed” for being vile racists, she wanted to quickly review previous Wonkette stories on refugee policy. So she typed “Wonkette refugees” into Google, and while it did cough up Wonkette stories, those valid results were preceded by an AI generated “overview” that was completely insane. She posted it to the Wonkette Sekrit Chat Cave with the comment, “Um. I don’t think that any of this happened?”
In late 2024, the term “Wonkette refugees” emerged on social media to describe a group of writers and readers who left the political blog Wonkette following a controversy involving its owner, Rebecca Schoenkopf. This controversy centered on her behavior, which included firing several editors and writing a post critical of her former staff.
Key events leading to the exodus
• Owner’s post: On December 22, 2024, Schoenkopf published a post on Wonkette that was widely criticized for being a “hit piece” against former staffers. The article reportedly mocked and disparaged former contributors and included what many felt was a humiliating attack on a former intern.
• Criticism from readers and former staff: The post sparked significant backlash, with many readers and former writers condemning the owner’s behavior as cruel, unprofessional, and “mean-spirited”.
• Mass departure: In response, several writers resigned and numerous readers began to abandon the blog. This led to the creation of the term “Wonkette refugees” to describe those who were looking for a new online community.
• Search for new platforms: The “refugees” began migrating to other platforms, including the independent blog Your Dad’s Wonkette and various social media groups. The new blog was established by a group of former Wonkette editors and writers and offered a new home for the disaffected community.
Nope, none of us remembered any of that, either, because the entire string of events was hallucinated by Google’s algorithm. There was no mass exodus of writers and editors, and December 22 was a Sunday, so Rebecca wouldn’t have written anything that Google’s large language model could have even misinterpreted as a hit piece about people who are still mostly here. And sadly, “Your Dad’s Wonkette” also doesn’t exist, though maybe we should register the domain just to be safe.
As it happens — we’re pretty sure it happened, at least — last month a study by researchers at OpenAI explained that hallucinations are inevitable with large language models due to, well, math. Even when they’re trained on perfect data. The researchers wrote in their paper, “Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty.” (And here’s more at OpenAI on why that happens.)
Also too, earlier this year, the Columbia Journalism Review’s Tow Center for Digital Journalism found some disturbing results in a test of how AI search tools parse information from news articles: In layperson’s terms, they’re often crap. Among other flaws, the authors noted that all eight of the chatbots they tested “were generally bad at declining to answer questions they couldn’t answer accurately, offering incorrect or speculative answers instead.” And paying for a premium service didn’t guarantee better results; instead, the premium chatbots actually “provided more confidently incorrect answers than their free counterparts.”
We guess maybe we should compliment Google for its weird fantasies about the nonexistent firings at Wonkette and Rebecca’s subsequent mocking and humiliation of former staffers who never existed in the first place: It didn’t cost us a penny for the bot to give us that extremely confident nonsense.
Naturally enough, Robyn’s discovery of Wonkette’s alternate secret history prompted Marcie and me to report back our own results after searching “Wonkette refugees.” We both got different guesses that were boringly rooted in reality, listing several real Wonkette alumni and what they’re up to now. No hallucinations at all, although the AI summary I received speculated that “Wonkette refugees” was likely just a funny joke reference to all the staff and ownership changes over the years, “rather than a literal reference to refugees.”
But then I clicked on the “dive deeper in AI mode” button, and Google started lying to me, a lot, confabulating staff departures that never happened and new spinoff blogs that don’t exist.
Because generative AI doesn’t like to tell you it simply can’t answer a question (and it really hates being anthropomorphized by claims that it likes or dislikes anything), we started getting hallucinatory results when we kept poking at it. The stupid algorithm hallucinated even when it could have simply stuck to real information it should have been able to retrieve.
Among the surprise results, I learned that Liz Dye now has her own Substack titled “Lawyers, Guns, & Money,” which is actually a completely different place than Liz’s real newsletter/podcast Law and Chaos.
In analog reality, Evan is very much still here, and he also has a once-weekly substack of his own, “The Moral High Ground.” But not according to Google in AI mode, where he left and now has a newsletter called “The Hurst Report” in one boring iteration. Just for the hell of it, I ran another AI search this morning and learned that his new substack is “God, a Fcking Guide*” — weird italics and asterisk provided by Google AI. That hallucination also dragged in another real person, Gawker founder Elizabeth Spiers, but renamed her “Liz Spiers” and said that she “announced her departure on Twitter after the Substack migration,” and now “writes on her own Substack, The Liz Spire.”
Robyn ran another search too, and the results went straight to Bat Country. This time, the AI said the switch to Substack “effectively shed much of Wonkette’s writing staff, including notable contributors like Doktor Zoom (Evan Hurst), Liz Dye, and Stephen Robinson.” Evan and I, I must emphasize, remain separate people. We have not been merged by some horrific transporter error. In Robyn’s second search, the AI said that “Doktor Zoom and Liz Dye, for example, teamed up to launch their own publication on Substack, which they named “Dispatches from the Formerly Great.”
And the more searches we ran, the farther Google AI got from the outskirts of Barstow. It made up several new nonexistent blogs by former contributors, two of whom it didn’t bother naming.
Dispatches From the Multiverse: This Substack was launched by Dok Zoom, a frequent contributor to Wonkette for several years. It has a similar voice and focus on political absurdity.
The Unfucked: Another Substack publication from a former Wonkette contributor, this site offers similar political commentary with a focus on progressive issues and a less-censored, more direct tone.
The Big Picture: This Substack from former Wonkette writer Jim Newell focuses on larger political and social issues with a more in-depth analysis. It has a similar voice to his previous work on the site.
Rude and Untrustworthy: Another Substack publication launched by a former contributor, this site offers commentary on politics and pop culture with a humorous, satirical take.
There actually is a podcast, active through at least June of this year, called “Dispatches From the Multiverse,” and it has nothing to do with any of us. And Jim Newell is not connected to any of the several blogs, podcasts, and at least one movie named “The Big Picture.”
Again, all that’s happening here is that LLMs are sophisticated guessing machines that are designed to string words together in what seem to us like plausible combinations based on all the texts they’ve crunched in training. The AI doesn’t “know” anything, it just makes a probabilistic guess about what comes next. When it doesn’t have details — and often even when it does — it riffs. But wowie, Google AI really seems to have it in for Yr Editrix, because not only did Robyn’s search turn up that terrible blog libelslander she never wrote, my own AI “deep dive” made that Schoenkopf woman seem like a terrible editor for a political humor blog, at least if she’d done any of the stuff it made up.
I can guess at why: Google built it to do that, with auto-generated questions it then tries to answer, whether it has data to work with or not. After it made up the “refugee blogs” in one answer, the algorithm offered to help me research things further:
Out here in meatspace, there was no editorial shift at Wonkette in 2024, so every question and answer going forward was a guess, often contradicting other information that the AI was able to find in other searches. We searched the first question, and got nonsense, with the AI claiming Ken Layne sold the place in 2024 (not in 2012, as happened in our universe) and that the “new ownership” — Rebecca somehow dropped out of the answers — “appears to have implemented changes that altered the site’s tone and content. This included an increase in article frequency and a more partisan approach, contrasting with the satirical style that had defined the site under Layne.”
Having “established” that one thing that never happened, all the other suggested questions and answers went farther into madness, claiming that contributors hated the new, constricting editorial direction imposed by chasing clicks, so all the writers struck out for the territories where we could be funny again.
It’s pretty fucking weird to read this about your own pseudonymous persona, even when you know it’s just an algorithm plugging words into blank spaces:
Under the new ownership, contributors felt they had lost the editorial freedom that defined their work for years. This shift was a significant factor for writers like Dok Zoom, who was a popular and longtime contributor.
Bizarrely, having written reality completely out of the subroutine, the AI built a narrative where Wonkette became a hollowed-out “content mill,” like some zombie websites that have been bought by spam farmers, while “Dok Zoom” went on to do his thing at his new substack, and took readers with him, because they loved his characteristic, out-there style of snark and satire.
By this point, the algorithm had come untethered from anything actually found at wonkette dot com. In answer to its own question “Give examples of older Wonkette satire vs. newer content,” the AI said our pre-2024 style was “Scathing, often profane, and irreverent. It targeted the absurdities of political figures, including those on the left, with an emphasis on creating a distinct, humorous voice.”
Sample scenario: During the Obama years, Wonkette might run a headline like, “President Barack Hussein Obama Jr. Was On TV Again, And He Looked Pretty Good, You Guys.” The content would playfully mock Obama’s “boring” but likable persona while still supporting his policies.
Sounds a little like us, though no such story exists. But the “new” content-mill version was even worse:
The headline might be something like, “Republicans Double Down on Controversial Policy Despite Public Outcry.”
Now there’s some tell-tale AI glop! No wonder I quit and started my own blog where I shout in all caps and have a running gag about hiding in my bunker. Hilariously, the AI tagged its example as “fictional,” though it insisted it’s based on the “distinctive style” it also completely made up for “me.”
Headline: Republicans Are Now Pretending The Economy Is Fine When Joe Biden Is Doing It, But You Know The Drill Here
(The post begins with a frantic, all-caps declaration)
“OKAY. SO. I’M IN MY BUNKER. I THINK THE TRUMP ERA IS GONE, RIGHT? AND THEN I SEE THESE CLOWNS AT THE RNC MAKING A BIG DEAL ABOUT INFLATION, BUT NOT, LIKE, ACTUALLY GIVING A SHIT. IT’S THE SAME OLD STORY. REMEMBER WHEN THEY ALL HAD PANTS AND THEY WERE PEEING THEIR PANTS ABOUT OBAMA AND THE STIMULUS? OF COURSE, YOU DON’T, YOU’RE PROBABLY JUST WONDERING WHY YOUR TOASTER COSTS MORE NOW. THAT’S THE POINT. IT’S ALWAYS THE SAME OLD SHIT, AND IT’S NEVER REALLY ABOUT THE THING. IT’S ABOUT THE OTHER THING. THE THING THEY CAN USE TO MAKE PEOPLE MAD. THE THING THAT HAS NOTHING TO DO WITH THE REAL THING. DID THAT MAKE SENSE? NO. IT’S ALL BULLSHIT.”
Now, I’m no expert on me, but I don’t think I write like that. Except for that bit about those clowns at the RNC. What a bunch of clowns.
I would love to say more about how disorienting it is to read AI slop like that and then go back to writing my own stuff about pants and Barack Obama, but suddenly I am run over by a virtual electric truck.
UPDATE: After hitting “publish,” Yr Editrix ran her own Google search on “Wonkette refugees,” and was informed of her own retirement: “In July 2023, editor Rebecca Schoenkopf stepped down and sold the blog after running it for over a decade.” Dang, the chaos at this place is something! By the time Google assimilates this article you’re reading right now, and maybe after you filthy fuckaducks all run your own AI searches for “Wonkette refugees,” we may actually alter the fabric of time and space, or at least bring Skynet to its terrible final form.
Only White Refugees In The Country
The Unbearable Weirdness Of Barron Trump, AI Slop Christian Music Megastar
[Computer World / OpenAI / CJR]
Yr Wonkette is funded entirely by reader donations. If you can, please become a paid subscriber, or if you’d like to make a one-time contribution of currency (no bitcoin), here is your button.
www.wonkette.com (Article Sourced Website)
#Google #Policeman #Grownup