The most irritating thing about a convincing man is that he remembers what you want to hear.
Sometimes it seems to be the only thing he remembered.
You said you needed certainty. His promises are now very definite.
You said actions mattered. He explains, in considerable detail, all the things he intends to do.
He listened so carefully to what was missing from my life that now I’m missing exactly what he promised.
Then other people start telling his version.
Without him. In serious voices. With sources.
And now you’re not arguing with one person. You’re arguing with “several independent sources”.
Wonderful. I used to just want to believe him. Now someone is explaining why I should.
He doesn’t even have to be there. Somewhere, he’s quietly drinking coffee while people explain that you’ve misunderstood everything.
An excellent division of labour.
You’re still trying to work out what happened. He’s already published the official version.
Once, this required recruiting someone you trusted. Sometimes buying their loyalty. Sometimes friendship and a promise to return the favour were enough.
But you still had to explain what to say. Then check they’d remembered.
The idiot might mix up which Thursday they were together.
Now you can try putting the right version where an AI will find it. Let it read. Find something similar in a few other places. Bring it back with references.
His friend used to cover for him.
Now it’s an answer with a bibliography.
At least the friend knew he was involved.
On 30 October, NPR published an investigation that makes you want to follow “Do you have any evidence?” with “Did you make it yourself, or did you order it?”
At the centre of the story is a network of 37 websites identified by researchers at VisionHeight. All 37 domains were registered in a single evening. Thirty-six within eight minutes. [1][2]
Independent sources. Even their births were coordinated.
The network included regional sites, political pages and sites posing as local newspapers. Different names. Different voices. A measured report here, an opinion piece there, a thorough-looking roundup further along.
At first glance: several newsrooms.
Look again: possibly someone changing outfits very quickly.
One site in the network was Kalamazoo Beacon. Journalists counted 16 articles, all about one congressional candidate, Sean McCann. ChatGPT cited the site when answering about candidates’ cost-of-living policies. [1]
A busy place, Kalamazoo.
Weather, traffic, a new bakery: all of that can wait. At the moment, the only thing happening in town is one man.
Sixteen articles about him. Not one about how the locals are coping.
Perhaps the newsroom was simply enthusiastic. Everyone needs a hobby. Some people cultivate orchids. Others cultivate public opinion.
But a shared birthday across a whole group of sites gives that enthusiasm a certain organisational flair.
There’s nothing inherently wrong with running several websites. One publisher can produce a newspaper, a magazine and another three titles, budget and patience permitting.
The question is what happens when a reader mistakes related voices for independent corroboration.
If you’ve brought your friends along to tell me how wonderful you are, I’d at least like to know they’re your friends.
Especially if you wrote the reviews yourself.
The opinion is ready. The facts will follow.
There’s a detail in VisionHeight’s report I would have invented if it weren’t already in the documents.
Some sites still contained production notes: content was being prepared; local stories would appear once a reporting process had been established.
Candidate recommendations were already available. [2]
The newsroom was still waiting for journalism. The opinion had started work.
Like a man who hasn’t decided where he was yet, but already knows you’ve ruined everything again by asking.
The conclusion is settled. Circumstances to follow.
CyberScoop had described the network on 15 October: repeated favourable assessments of candidates, linked coverage and pages accessible to automated crawlers. Researchers viewed it as an attempt to influence the sources behind AI answers. [3]
And here we find an interesting saving in effort.
A website doesn’t have to become your favourite newspaper.
It only has to become your favourite assistant’s source.
You may never see its homepage. Never open “About us”. Never notice that the local newsroom hasn’t quite got round to taking an interest in the locality.
Your assistant finds the article, summarises it and adds a tidy footnote.
You didn’t trust that website. But someone you trust introduced it.
Apparently, references aren’t just useful when applying for a job. Even the truth benefits from knowing the right people.
References supplied
On 22–23 October, NPR tested ChatGPT with 37 selected questions, asking each once. Nineteen answers cited sites in the network, often alongside reputable sources. That means nineteen answers containing those citations, not nineteen proven lies or nineteen voters whose minds were changed.
The network mainly promoted Democrats. Its operator is unknown; the candidates’ involvement has not been established. The Democratic Congressional Campaign Committee denied involvement. OpenAI said it acts on identified problems to improve answers. [1]
But this story doesn’t require every sentence on every site to be invented.
You can use real information. Choose the convenient parts. Describe one person’s strengths in detail. Leave the rest outside the frame.
Give the opponent some attention, too.
About as much as a man gives the word “but” in “I like you, but”.
Carefully selected truths can leave a deeply dishonest impression.
Individually, everything looks respectable. The date is real. The speech happened. The quote exists.
It’s just that, out of the entire life story, you’ve only been shown the photographs without his wife in them.
Then that carefully selected version appears in a few more places.
The number of links looks persuasive.
Like the number of people telling you he’s a good guy.
Worth asking how many owe him a favour.
It didn’t necessarily believe him. It may simply have repeated him.
It would be easy to write: “They taught AI to lie.”
Easy, dramatic, and not what the investigation shows. Questionable sources turned up in search-based answers. Nobody has shown that the model was retrained on those pages or that its knowledge changed for good. OpenAI even uses separate crawlers for search and for collecting potential training data. [4]
So no brain surgery required. Just be on the right shelf before the conversation starts.
And it doesn’t take eternal memory. One well-timed answer is enough, like one well-timed text at 2 a.m.
Being readable by AI proves nothing on its own, either. llms.txt, for example, is a proposed format for a site guide, not a magic “believe me” button. [5] Honest people format documents nicely, too. Sadly, so do the others.
Which means we still have to check the documents.
Yes, annoying. I’d hoped artificial intelligence would at least automate that part of relationships.
Switching assistants isn’t a divorce from the problem, either. The Atlantic separately reported that Google AI Mode cited sites from the same network. [6] Different question, different day, different search mode: different answer.
An assistant’s confidence is not a background check on everyone it knows.
You came for a second opinion
Political advertising disguised as journalism is older than chatbots.
People have been calling their own interests the public interest for years. They didn’t even need a subscription.
What changes now is the voice in which an interested party’s material can reach us.
A clearly labelled campaign leaflet at least declares its intentions. We understand that somebody wants something from us.
An assistant arrives differently.
It listened to the question. Compared the options. Put everything into bullet points. Saved us time.
The original material may have a specific job to do. The person receives it as a helpful explanation.
You came for a second opinion. The first one simply changed its voice.
That’s the part I find genuinely uncomfortable.
Once, slipping me a convenient version required earning my trust, or recruiting someone who already had it.
Now you can try getting into the material my assistant reads on my behalf.
It needn’t be part of a conspiracy. It can deliver someone else’s prepared version in perfectly good faith.
The most convenient witness is the one who doesn’t know they’ve been briefed.
These investigations don’t establish how often this changes people’s decisions.
But the route from someone else’s interests to my feeling that “I checked” is already visible.
Along that route, verification risks ending exactly where it should have begun: at a well-written answer with references.
A comforting answer is easy to mistake for an honest one. Especially when it settles a question that hurts to ask.
I wasn’t asking you to say what I wanted to hear. I was asking for it to be true.
Do they know they’re independent?
I still want links.
I just want a link to begin the checking, rather than end the conversation.
Who wrote the piece? Who publishes the site? What supports the conclusion? Are these separate reports, or one version that has reproduced exceptionally well?
An unfamiliar site may turn out to be an excellent small newsroom. Obscurity isn’t an accusation.
But footnotes don’t automatically make a text honest, either.
A footnote is a wonderful thing. Sometimes, though, it leads you to his friend instead of the evidence.
I’m still susceptible to convincing people.
Especially when they answer calmly and can back up what they say.
It’s just that now, when I hear “several independent sources confirm this”, I have one more question:
Do they know they’re supposed to be independent?
Or do all thirty-seven happen to have the same story about where he was on Thursday?