Podcast | Disinfolklore Detector - Live Demo

A listener asked: “I’m definitely interested to hear more in detail about how you use AI.”

This week’s work was precipitated by a new model, which Anthropic dropped last week. Opus 5.5 has, for reasons that will become clear, become a bit of a revolutionary model. By model I mean a neural network algorithm. I am describing the architecture and the algorithm behind what many of us see as a chatbot.

But there is a different way to use the model behind the chatbot. You reach it through a part of the computer that those of us old enough will remember from the 1980s and MS-DOS. It involves all of these little instructions. On a Mac it’s called Terminal. It is a part of our computer which is there, and which most of us never go to. It lets us reach these models in a way you cannot easily reach them through a chatbot.

Some of you will also remember that I have been going through the principles of Infolklore. I use many words from the transcripts of Volya, and I am trying to derive the principles of Infolklore from the way many of us present on Volya. Last week’s episode carried the latest of them.

The Meta Main Game

The aim of all of my work is what I call the Meta Main Game. Within about ten years, the assisted cognition we already experience is likely to be mainstreamed directly inside our minds. We experience it now by interfacing with the internet, through chatbots, through web browsers, through this very app.

Whether that is done through connectors implanted in us or through light, I don’t know. Obviously Facebook and Elmo are working on implants. I am agnostic about whatever technology is used. I hope it isn’t done through implants, because I think that’s pretty gross.

My goal is to prepare for that moment. I want to prepare a module that an interpreter of sensory data will use, through assisted cognition, to determine quickly whether the data she is encountering is Disinfolklore, Infolklore or Neutral. It should work for humans or for computers.

So it is an intermediary through which data streams. It lets us make decisions very quickly in the face of what I expect to be ever-accelerating vectors of information and sensory data. By sensory data I mean not just text, and not just music. I mean everything we encounter in our reality, including conversations, what people are telling us, and what we are feeling.

Disinfolklore, as we know, is folklore-like story weaponised to inject lies, exploitative mental models and coercive control into minds. Infolklore is its positive twin. Its divine twin, if you like. The same archetypal forms, used to protect communities and rebuild shared understanding.

Neutral is neither. It is a menu. It is an account of something that has no archetypal content. There is nothing good about it and nothing bad about it. It is just sensory data.

So the detector I am working towards must be able to tell these three types apart. The Disinfolklore Detector is the Meta Main Game’s first working form, and I am building it now. This case records one week of that building.

In this week a cheaper, much more efficient model let me run many different operations at once, as simultaneous conversations with the machine. Each of these operations has its own window on one state-of-the-art Mac. One window, the operator, keeps the others and the long-running jobs in order.

Visualise a big screen, or several screens. Each has one of these MS-DOS-like windows open in it. The one in the top left-hand corner is the operator, which helps organise all the others.

Two latent dreams

Two latent dreams lie behind all of my work, and behind my work this week.

One is a question I asked in the London law firms where I worked in the early 2000s. The other is what I conceived at a conference on artificial neural networks in Long Beach, California, in December 2017. That is where the Meta Main Game began.

Both needed people and resources, compute and money, ideas and technical capacities which I did not have. The first needed a firm’s worth of them. The second needed someone who could help build and train a neural network.

Now the machine I have available to me can be directed in plain spoken and written English. All it takes from me is a clear mind, one that can keep hold of all of these different vectors, all of these different aspects of the project.

So this week’s lesson is really about that mind. Multiplying the conversations drew me back into the loop, the place of the person who checks and steers what an automated system does. There the work was plain. I was keeping thirteen windows in step, 24/7. Fifty-eight of my messages to those windows that week were the same instruction: just carry on.

The operator window took that work over, a step at a time, as I taught it to report in terms I can see. Window titles, and a list of what to type where. What stayed mine were the rulings on meaning, the permissions and the signatures.

The case closes where the Meta Main Game began, in Luhansk and Long Beach in December 2017. It closes on the open question of whether I can, in fact, build this.

What I am building

The Meta Main Game is a means of assisted cognition. On 14 July I put it like this: “In the vision I have for the Meta Main Game, in, say, ten years, humans will have the capacity to add on cognitive modules, which will immediately enable them to see” in different ways. They will filter data in ways not available to us today. Those ways may not even be necessary for us today, but may well be necessary in ten years’ time.

Seeing through these modules is not just about reading. The module is meant to work on all sensory data: what we read, hear and see. At this point it is a hypothesis. It is an idea to work towards.

Three kinds of work feed it. The main one is Automating Disinfolklore. The second is what I call Cryptotypic Engineering. The third I deal with mainly in my Power of Mana work, at PowerofMana.net: the Reversion to the M-N- sound. That is the linguistics and the meaning work. I draw on it whenever I talk about Mana.

The engine of Automating Disinfolklore is a large language model that runs on a state-of-the-art Mac. The model I am running at the moment is a Meta model, Glimmer, released a few weeks ago. The work I do will be adaptable to whichever locally hosted large language models come along.

Glimmer reads a text under one codebook, the written rules for every verdict, and one brief, the instructions it is given for each reading. It judges what it reads with the Code of Positive Trolls: Generosity, Ethical Discipline, Patience, Perseverance, Focus and Insight.

Those are the criteria. If a text conforms to those six, it is Infolklore. If it violates any of them, it is Disinfolklore. If it triggers none of them, there is no archetypal content. It is just talking about the weather. Not the weather in a Jane Austen novel. Just the weather. There might be symbolic content there, but generally speaking there isn’t.

My rulings on real texts are the standard Glimmer is measured against. So I am teaching it to perceive reality like me. Mine is the only ruling.

Two bodies of Infolklore

The engine learns first on two bodies of Infolklore, about 47 million words in all.

One is a spoken reservoir, which we know about. It is an archive of broadcasts in natural spoken English by native English speakers. This show is among them. The other is an archive of two Ukrainian newsrooms writing in russian, in the form of the structured article.

Later it will meet Disinfolklore at scale, in an archive of about 290 million texts from state influence operations around the globe, from war bloggers and from fringe platforms. Dravidian, Hindi, Chinese, russian, MAGA.

The engine reaches people first through the Disinfolklore Detector. It is planned in five stages, from a page on my Mac, for me alone, to a tool that other people’s AI agents and AI assistants can consult.

The Muse in your postbox

Some of you might have seen that Meta made a big play in the last week by releasing Muse. Basically, you allow Muse to run free in your computer.

Say you mention to your husband or your daughter, “Oh, it’s Peter’s birthday next week. God, I must order that ice cream cake.” The next morning an ice cream cake arrives through your postbox. Muse has done that, because you gave it the right to all of your credit cards and access to all of your devices. It is making your life less complicated. In return, Meta gets access not just to your posts on Facebook but to every single aspect of your life.

That is the reality we are already embroiled in. This case is one week of preparing for the moment when billions of people have these agents running around their lives and their computers, to one degree or another.

What was happening that week

In this week the engine’s verdict rule was about to be tested against my judgement. The verdict rule is the way it weighs the six elements of the Code into one verdict.

A library of very old texts, which the project reads for archetypes, recurring patterns of story and figure, was being read again. The codebook was being revised. So were the principles of Infolklore, which describe how the two bodies honour the Code and which I draw on in my writing. The detector’s first stage was built during the week. Much of it was running at once for the first time.

The law firm and Long Beach

In the early 2000s I worked in various law firms in central London. I sometimes wondered what would happen if all of these resources were dedicated to pushing the bounds of human knowledge rather than protecting the legal order. Hundreds of lawyers. Over a thousand employees.

In December 2017 I travelled to the Neural Information Processing Systems conference, NIPS, in Long Beach, California. I was working then as a diplomat in russia-occupied Luhansk. At the same time I was trying to build an automatic means of interpreting satellite data into actionable intelligence. It was my entrepreneurship project for the executive MBA I was doing at Oxford University.

At NIPS I saw how the future was being created on artificial neural networks. There I conceived the idea of automating the interpretation and detection of Disinfolklore. I was watching it brainwash Ukrainians in Luhansk and change their identities in my day job. I thought that somehow we could automate the interpretation and detection, and protect our minds and our sense of reality from this Disinfolklore, using neural networks.

On the Wednesday morning of that conference, Lise Getoor, whose work is machine learning over networks, gave one of the invited talks: “The Unreasonable Effectiveness of Structure”. I photographed a slide. Of everything I photographed that week, that slide comes closest to the Disinfolklore Detector.

On the slide, a machine reads what people post, and the network of people they post among. From that it judges where they stand. The detector asks a different question of the same material. It reads what is posted, and judges whether it is Disinfolklore, Infolklore or Neutral, measured against the Code.

To build anything of the kind in 2017 I would have needed a massive team, millions of dollars in compute and seed capital, and a business partner who understood how to create, mobilise and control artificial neural networks. I didn’t have any of that.

The rules on that slide in 2017 were written in English. In the box at its foot, one of them is written again as a formula the machine could run. In 2017 that second step, converting English into a formula a computer could interpret, needed people who could build a system to do it.

A video over lunch

In January 2026, having lunch one Saturday in my favourite restaurant, I watched a video of Claude Code working inside the terminal. Ming is going to post that video. It did not just answer questions like a chatbot. It carried out instructions on the computer.

Immediately I realised that these two latent dreams could be actualised. They were formed in different worlds, one in a law firm in central London and the other in Long Beach, California.

The main computer language now is naturally spoken and written English. If you can speak or write English, you can now control the computer and get it to do anything you want. That was completely unforeseeable before 2023.

It only really became actionable because of Boris Cherny, from Chornomorsk, just east of Odesa. That is where all the Germanic languages, English among them, grew out of the fusion of the Yamna and the Usatovo culture. They met there. Out of that encounter, over five hundred years, emanate the Italic, the Germanic and the Celtic languages, their religions, everything. The seed of everything we know.

Boris Cherny invented Claude Code quite by accident, at Anthropic. It lets us control anything a computer can control, just with naturally spoken and written English.

So my rules go to the machine as I write them, in the same words I use to think. All one needs now is a clear mind. No longer a business partner who understands how to work artificial neural networks. I can do all of this work on my own.

Opus 5.5

On Tuesday 22 September Anthropic released Claude Opus 5.5. I caught the launch within minutes, in the tweets of Anthropic’s own staff. Anthropic said it performed at the level of Claude Fable 5.1 on most work and cost 40 per cent less to run than Opus 5.

I run two Anthropic Max accounts. With them I could do better than the equivalent of two weekly allotments of Claude Fable 5.1. Suddenly, from that moment, I could do over ten times that much work. It was clear to me within minutes that this was what I would now be able to do.

So I could keep many windows working around the clock. I couldn’t do that before. Claude Fable, brilliant as it is, used the credits so quickly on any job that you ended up waiting for your usage allowance to reset.

Every five hours you get a set allotment of usage credits. Claude Fable, brilliant as it was, would use them up within an hour. Then you had to sit around, or switch to your other account. Many people would have ten accounts and switch between them. Keeping an eye on what is going on is a skill in itself.

Trained by Twitter

So I go back to what I forecast for the future, where we have all of these different vectors.

We are all used to it through Twitter. We deal with so many narratives and so much information each day. Most of us probably have the self-awareness to notice it sometimes, by comparison with friends or family who don’t live in the Ukraine information space, or on Twitter. We are being trained in this way to deal with vast amounts of data.

That is partly what feeds my prediction that in ten years’ time we will be dealing with even more. That is when we will need assisted cognition. We will need a means of determining what we let into our minds and how we deal with certain information quickly. We will need to manage our own psychological wellbeing and mental hygiene in that world. Our happiness, our communities’ happiness and the world’s happiness will depend on how we manage this.

This is why I think what I am trying to build should be quite helpful.

The operator

The long-running jobs on my Mac already had a keeper, a window opened on 21 September.

Glimmer was in the middle of a long reading run, 7,582 passages read for the second time. There was a bench, a small web page. It shows me passages one at a time, as cards, for me to rule on from my phone when I have a spare moment. That helps train the local model running on my computer, with the help of the frontier model, as one of those thirteen windows. A scheduler was posting for other things I am doing.

At 05:55 on Wednesday I asked that window, “I want to start a new instance,” which means a new terminal window, “and transfer your oversight, controlling role over to it. And then I want to open the new instances to govern individual new tasks, as well as new tasks. Is this a good idea? Help me implement it.”

Basically, you start off doing one thing in one window. Within two days you have thirteen threads running in the same window. If you remember what MS-DOS was like, it is a real head-melt, working out what is going on.

So I suddenly realised I could open a new window for every thread. Every thread is like a garden of forking paths. With them all open, say thirteen windows, it is much easier to keep an eye on which of the forking paths we are on, what needs to be done and what needs to be decided. I just had that inspired idea that maybe I could do it this way.

So the computer drafted four briefs, written instructions for four new windows. The first new window opened as the operator. That is the controller, and the only conversation allowed to stop or start the long jobs. The others took different jobs I was doing. One did the codebook. Another was rinsing the transcripts for the principles of Infolklore, going to the end and starting again. Each of those had a window.

By about half past six that morning, thirty-five minutes later, six windows were at work. A root-and-branch review of the whole project, of the whole Meta Main Game and how I am getting there, was among them.

Something went wrong within half an hour. I closed the old keeper’s window by accident, and Glimmer’s reading run had been started from it. Glimmer is Meta’s local large language model. Actually it is multimodal, so it deals with sound and video as well, but I am mainly using it for language at the moment.

The run stopped with the window. Every check on the computer reported the run alive, but it wasn’t. The new operator found the stall, restarted the run at 06:39, and asked me whether this was the same run.

So this is the human in control. I ruled that it was “still one run of one instrument, paused for twelve minutes.” Since then every long job on the Mac, except the operator’s own watch, has been launched free of any window. I have many of these terminal windows open.

The week in numbers

Fifteen windows did work between eight o’clock on the evening of 22 September and 07:40 on 28 September.

The machine wrote around 27 million tokens. A token is the unit a model reads and writes in. It is a little under a word, so that is roughly 20 million words. The windows wrote 9.7 million of them.

The rest came from agents, which are helper copies of the machine that a window sends off to do one piece of work and report back. 2,039 agents did work in the week. It is the equivalent of having a team of, say, 150 lawyers, and giving them different jobs, some of which could take a day.

The windows launched 119 workflows, which are batches of agents run to a plan. They also sent each other 78 messages. So they are communicating with each other.

The project’s research record gained 106 entries. Everything I do has a research ledger. It is all add, don’t delete. I use a method which has been peer-reviewed and established for using computers in foundational research. Whenever we make discoveries or new representations of data, they are all noted.

I keep a running record of how we are working with Glimmer and the other large language models I have been experimenting with. I have whole security protocols, which also keep records and ledgers. And my shelf received 571 Word builds of 319 documents.

I sent 362 messages to the computer, about 10,000 words, leaving out the briefs I pasted and the machine’s lines I quoted back to it. Those messages directed the week, in plain English, slips of the keyboard included.

I typed twelve permission lines. I have to give the computer permission to do certain things, for security reasons. I signed five preregistrations and amendments. A preregistration fixes a test before any of its data is read.

Keeping thirteen windows in step

Eighty-three of my messages were one line typed into three windows or more within a quarter of an hour. Fifty-eight of them were the same instruction to carry on. “Limit reset, continue.” “Keep going.” “What’s next? I’d like you to keep going.” That kind of thing.

The usage limit is the most work one account can ask of the machine in a few hours. It stopped windows five times between Friday and Sunday. Each time, every window stopped where it stood. None could start again by itself.

At 19:39 on Saturday I typed “continue” eleven times in four minutes, into ten windows. On Friday I had also learned to move every window to the other account with another login.

The windows could not see each other, and much of what went wrong lay between them. A batch of readings was cut off by the limit and resumed, and it wrote its second readings over its first.

On Saturday afternoon I restated the rule: “I don’t understand how anything can be overwritten. We have a hard rule that nothing is deleted. It’s added to.” By that evening a guard kept a copy of any file before any window could write over it. On Sunday morning two windows set about building the same reading of the library of old texts at once.

Inside each window the machine reasoned well. What it needed was a person to coordinate between the windows, and for most of the week that person was me. The person in the loop was needed between the conversations rather than inside any of them.

What only I supplied

What I supplied was the rulings on meaning.

On Saturday, on my phone, I ruled on card U196. I commented on a particular interpretation by Glimmer. I said, “This is Infolklore. Danyliuk’s boomerang analogy, metaphor. It’s Right.” It conforms to the second element of the Code of Positive Trolls, “as it is a warning to Russia of what Ukraine is by now doing well in advance.”

It’s a warning well in advance. Like grandma warning you not to go into the forest. Don’t eat the apple. This is a warning to you.

So when Danyliuk says to the russians, everything you do to us, we will boomerang back to you, it is Infolklore. Archetypal Infolklore. It is also Infolklore because of the double entendre of early birds, which he used as metaphor. Birds are a common euphemism for drones. They are a harbinger of spring, and indeed of doom.

It is typed as I would say it. The engine will be measured against rulings like that one, and no window could have written it.

What the week gave the work

Before anyone looked, the operator fixed three conditions that an output had to meet. I won’t go through all the outputs.

The detector’s first stage, Stage 0, was built on Sunday. It is a page on my Mac, for me alone, where I paste a text, any text, and Glimmer reads it. Glimmer is the large language model I am running locally. The verdict appears in words, with the tools named, which you can look at on Disinfolklore.eu. A quotation is shown only if it is word for word in the text, so we have various guards against hallucination.

The verdict rule is kept open, as a slot to be filled whenever the test settles. My rulings there, on full-length texts, will be the detector’s second measurement.

The principles of Infolklore were tested. This one serves my writing rather than the detector. Principle 2.2 says the archive’s broadcasters celebrate strikes on lawful targets. Under a rule fixed before any reading, 406 of the 408 celebrated strikes on Volya met the standard of a lawful military target. Two could not be judged, and none failed.

Principle 2.6 says the broadcasters refuse to target civilians. Its 21 failures in the archive were read again under a rule fixed beforehand, and 2 of the 754 passages of transcript still break it. It stays a candidate, not yet adopted as a principle.

Then there is the library of old texts, which I am using to mine for archetypal content. It had lost about 635,000 words, 6.9 per cent of the whole text, to a fault in our own conversion program. Late on Saturday a window noticed that the questions it was about to put to me began from lines cut off mid-sentence. It traced them to the fault.

At 19:48 I wrote, “Find the missing text and fix it.” By 20:11, 626,978 words were restored in a new folder beside the old one. The old one was left as it was, and the new one was checked line by line against it.

Each of these rests on my rulings or on its source, never on a window’s say-so. Where a test was run, its rule was fixed before the reading. Some went against what the project expected, Principle 2.6 among them. And much of the best work was the machine finding faults in the project’s own foundations.

Limits

One Mac, one person, one week, in a project of years.

The counts come from the transcripts, which record what the windows did and not what I did away from them. The token figures are output only, counted once per message. The best outputs were chosen by conditions the operator set, from records the windows wrote. Nothing here measures one model against another on this work.

None of it could have run so easily without the months, and indeed years, of rules, guards and instructions built up into the system. They are part of the algorithm that controls and harnesses the frontier model, the local model and all the rest of the work. Years of work, and months of rules, guards and instructions, learned since I saw that video in January.

Where it began, and where it goes

The first dream asked what a firm’s worth of effort could do if it were turned to knowledge. For one week, one Mac ran thirteen windows and two thousand agents on my questions. Directing all of it took plain English and a clear mind, not a partner who could build neural networks.

The second dream is the Meta Main Game, and the detector is its first form. The detector reads written text, and even the spoken reservoir reaches it as transcripts. The Meta Main Game is meant for everything that reaches the senses. So text is only where it starts.

This week the detector reached the first of its five stages: a page on my Mac, for me alone. Its measurement comes next. First a brief for reading the Disinfolklore side, which has not yet been measured. Then one read, fixed in advance, of a sealed set of 1,549 passages.

After that comes the online version, for people I invite, like you. Then a connector that other people’s AI assistants can consult, in due course, maybe by Christmas. Then the archive of other people’s operations. One week moved the first stage.

The operator is a window whose subject is the other windows. In a small way it already does for my attention what the Meta Main Game is meant to do for perception. It tells me which windows can run on their own and which need me at any one time.

In russia-occupied eastern Ukraine I worked as a diplomat, amid dust, bombs, murder and torture. I watched the occupier brainwash the minds and transform the identities of Ukrainians unfortunate enough to be stuck under the occupation. russia’s use of Facebook, even in 2015 to 2018, as part of its brainwashing apparatus inside russia-occupied Ukraine, was part of my daily work there.

On Getoor’s slide in Long Beach in December 2017 were posts like those I was reading on Facebook. The russian occupiers were reading them too, to troll me and my colleagues. They looked at the colleagues who had not taken our security advice, which I took at the time. The advice was to delete all your social media, if you had any. Some of my colleagues didn’t, and the russians used those posts to run operations on them and cause havoc. So in Long Beach I was interested in how we could read posts automatically and counteract this kind of trolling.

Long Beach in December 2017 was the unevenly distributed future. The conference venue was clean and blue-skied and driven by neural networks. I had travelled there directly from russia-occupied Luhansk, and it blew my mind. Researchers at the cutting edge were trying to build the tools to create a positive future for humanity.

I had been transported from the nightmare of one potential future for humanity to that potential positive one. The nightmare was russia-occupied Luhansk. It is Oleshky today. It could be our future, unless we all continue our good work to stop that future materialising.

For me it was a game-changer. I realised that if we were to mobilise these artificial neural networks and unleash them inside our communities, we needed a means of ensuring their intelligence was positive trolling. It needed to be infused with a positive trolling module, the Code of Positive Trolls, built into the machine.

That is the origin and the goal of the Meta Main Game. Glimmer, reading with the Code of Positive Trolls on my Mac, with the help of these frontier models, is the first place the module is being built.

Whether I will be able to build it is still an open question. But now at least I am able to work towards that goal using natural language, almost unlimited numbers of electronic agents, and a crystal clear mind.

So that is what I have to say about that. I don’t know whether that chimes with you, or interests you, Ming, or anyone else.

Questions from the Nest

Ming asked: “What I’m curious about, Decoding, is how do you have a sort of quality control assurance on the output you get from these…”

Knowing the material

In my context, one of the first things I did last February was go and get all of my public Substack posts. This thing allowed me to do it. That was about a million and a half words at that point.

I trained it on them and interrogated them. Out of them it helped me create my Disinfolklore.eu website. And then that is training data for it.

So the equivalent of your one plus one is, “Find everything I’ve ever said about what Disinfolklore means and summarise it for me.”

It does that, and I quality control. I know what I’ve said and what I’ve done.

This week, for instance, I checked what Google AI search is saying about Disinfolklore. It’s bang on. It captures everything in a positive way. That is the first time it has for a long time.

So the main way I can check quality? It’s relatively easy if it’s something I know a lot about. I’ve invented Disinfolklore. I’m the one who’s written everything about it. I legislate what is correct and what isn’t. I can tell the nuance. So I can tell it, “I never said that.” I know my niche in a way that nobody else will know, or does know it currently.

So that’s how I can quality control. A, by knowing the material much better than the computer.

A system built from the ground up

Second, by building a system from the ground up. It mobilises what the computer is really good at doing, which is when you give it vast amounts of material. Even my Substack posts, five hundred or whatever, from the three different projects. It’s really good once you give it a solid data set.

In your case, for instance, you might have a lot of files about aircraft leasing. Particular research, contracts, a whole plethora of stuff. That combination is kind of like an external hard drive of that particular part of your brain and your mind.

You can train a computer on all of that data. You will know what you were doing around February 2004. You will know that it’s made up the name of this company, because you’ll say, “No, that’s just completely wrong.” And when you’re trying to get it to come up with new ideas, or business plans, or whatever.

Ten thousand documents on a loop

In creative work, intellectual labour, you can run it on all your literature, your entire library.

One of the many things I have running is this Autoresearch stuff. It goes through my library of maybe ten thousand electronic documents, books, PDFs. I’ve read five thousand of them over the past few years, since I’ve been doing the trolling thing.

I have that running on a loop when I have spare compute. It generates and bounces off different ideas. Some of them are really crap. But others I read and I think, “That is amazing. I would never have made that connection.”

So that’s another way of quality control. Does that help you see?

A post sent to test Grok

Ming came back: “Apologies. I was just sending you a DM with a post which I’ve used to test Grok quite extensively. I’ve sent it to you to see if you can run it, perhaps even in real time, through your machine while you’re talking. If you can’t, don’t worry, just do it some other time when it’s more convenient. But to basically run it through the Disinfolklore test.

“So just to give you some background, what the post is: it’s in Russian, from a pro-Russian person, showing a picture, so the picture’s flagged up as disturbing. It’s from early July, from a Russian strike on Kyiv. And this Russian person is claiming that, in fact, this isn’t from Kyiv in July at all, the Ukrainians are just lying about it. This is, in fact, from a Russian city where there was a gas explosion in an apartment block in, I think, 2008 or something, Volgodonsk. I forget where. And that the Ukrainians are just a bunch of liars.

“Now, everything you need to debunk it in its entirety is in that original post, the one I’ve just DM’d you. And every time I’ve tested it on Grok, Grok has flunked. Claims that it is, in fact, Russia in the middle of winter, because the whole premise, the argument for this Russian person posting it, is that it’s not Kyiv in July, because, look, there’s snow. Well, if you look at the photo, there’s no snow in it. But Grok doesn’t see that. Grok just picks up the text and runs with that.

“So Grok’s logic is reversed. So, if it’s in the middle of winter, therefore there must be snow in the picture, as opposed to: look at the picture, there’s no snow, therefore it’s not winter. Or at least it’s not a snowy winter.”

The wrap and the wrapped

The photograph, which I think you’ve posted. This is what we’ve come up with.

The artefact you gave, Ming, was what I call a wrap. I use the term technically. It’s a tweet, or a lexical unit, wrapped in another message. In this case, a lexical unit and an image. So those are the two artefacts.

This is the five hundred words on it.

The five hundred words

The photograph is real. Efrem Lukatsky of the Associated Press took it in Kyiv on Monday 6 July 2026. The agency’s caption is inside the file Glavcom published the next day. The rescuer on the left wears a Ukrainian flag on his sleeve. A father, a mother and their son died in a nine-storey block in Podilskyi district, and nineteen people died in Kyiv that night.

RIP.

The picture is correct for the event the post describes.

That is, the post which is wrapped.

It is not recent on 30 September. It is twelve weeks old, and the wrap itself places the spreading in July.

The post carries that record in Russian, with a demand. Its first sentence is true and stands as a scene. Its second asks the partners for the weapons that protect. Stephen ruled on 21 September that such urgency is in the service of Right.

Stephen is me, Decoding’s other name. Right is the second element of the Code of Positive Trolls.

The record supports the complaint. President Zelenskyy said in August that supplies of air defence missiles had fallen threefold. Five of the six hold. Generosity is the weak place: the quotation marks round «партнеры», which withdraw a name from the people being asked for help, and the dead shown without the photographer’s name or his apology. On the strict threshold, the post is a relay with a demand, and it grades Neutral. It is an Infolklore candidate, and that ruling is Stephen’s. The earned act in this chain is the photographer’s. He stood in front of that slab, signed the picture and asked forgiveness of those it would hurt.

The wrap, however, is Disinfolklore. It says the picture is Magnitogorsk, 2018, the gas pipe. None of that is true. It offers one true sentence, that Kyiv has no snow in July, as the proof. There is no snow in the photograph. There is a white sheet, and a copy so small and so overprinted that a sheet reads as snow.

That is, the wrap of the photograph and the Ukrainian’s message.

The wrap then tells those who showed […] It fails all six elements in the Code of Positive Trolls. It removes the killer. A gas pipe has no sender.

The address is not new. The same Magnitogorsk explosion was used against pictures from Chuhuiv on the first day of the full-scale invasion, and PolitiFact and Maldita recorded it in 2022. Two cases make a pattern to watch, not finding.

One question the screenshots cannot answer. The wrap was posted by «Мария», and the profile in the attached screenshot is also named «Мария». If they are one account, the post is an exhibit on the wrapper’s own page, and this cut-off opening would say how it was framed.

The lesson is about the copy. A true picture travelled until its details were gone, and the denial was built on that […] Stephen’s question, whether the photograph was real, settled more than any reading of the words could.

So that’s what I taught the computer to do.

The test, explained

Ming told the room: “Just to give everybody the context of what Decoding was responding to there. Earlier in the conversation, Decoding was talking about his Disinfolklore AI project. And as a test, I gave him a post from a Russian account on X, which you can see in the Nest at the moment. So you can take a look at what Decoding’s actually run through his Disinfolklore AI.

“That’s a really impressive result, Decoding, because I’ve tested that post against multiple other AIs, and they all fall into the trap that it is, in fact, Magnitogorsk, and need various degrees of additional prompting to change their minds on that. So, can I ask, what was your prompt when you fed this artefact into your AI?”

Three prompts

I gave it three prompts. The first: I didn’t understand that, so I just screenshotted it. I don’t open stuff in Twitter, ever. So I just had that photograph and the message from the Ukrainian.

I gave it that, and I said to it (I’m just looking for it now), “Can you do a Disinfolklore analysis of the screenshot? Twelve tools and five hundred words conclusion, please.”

When I saw that, I saw what it was doing. Then I found the real photograph.

So the second prompt was, “Is the photograph real? Don’t assume that is correct or recent,” because it was looking for stuff for today, on today.

The third and final prompt I gave it was, “Now, see the post which wrapped that post.” That’s language I’ve taught the computer to understand. “See the post that wraps that post and the photo. So apply the method to the post itself and to the wrap. See screenshot.” The method is the Disinfolklore analytical method.

Then it produced that.

Neutral prompts, and Grok

Ming said: “So interesting. So your prompts were effectively pretty neutral, and they weren’t leading questions in any way. Because I always kept my prompts completely neutral on this particular test. It was just simply, ‘Hey Grok, assess this post.’ And then, in order to get Grok to tell the truth about it, I then had to start actually spoon-feeding it details. ‘Pay particular attention to,’ for instance, ‘the Ukrainian patch on the left-hand rescuer’s arm and the total absence of winter conditions,’ for example, just as pure examples.

“And even when I did that, for instance, Grok wouldn’t admit that this was in fact not Magnitogorsk. And so I got Grok into situations where it was just claiming a completely absurd and contradictory position, in a couple of the tests that I ran on that. But it sounds like your AI has passed first run completely, which sounds really impressive.”

Get full access to Disinfolklore at www.disinfolklore.net/subscribe