What do the AI-detectors say? Hive AI: 0.1% AI-generated. GPTZero — “highly confident this text is entirely human.” Truthscan says it’s probably 91% AI.

That is like three doctors examining the same patient, two saying “perfectly healthy,” and one saying “emergency surgery.” The numbers sound precise. I showed this to 600 journalists at a GIJN webinar yesterday. The full recording is on YouTube. (Also featuring my signature mispronunciation of English, delivered with full confidence.)

The central argument I made there, and the argument I will make here, is this: a detector is not a detective. You need both — and the reason you need both is not just that each catches what the other misses. It is that each fails in ways the other can correct. A detector flags a seven-fingered hand at 5.2% confidence. A human catches it in two seconds. A human looks at a ring disappearing in a video and concludes deepfake. A detector agrees. Both are wrong — the phone had a known lens-switching bug. Neither the machine nor the human asked the second question.

The errors do not stack. They cancel. But only if you use both, and only if you stay the judge.

The sentence nobody scanned

Buried at the bottom, in Polish, in a box:

“If you’d like, I can also prepare a more complete post-head-injury interview for a child with standard questions about loss of consciousness, vomiting, dizziness, which is commonly used in emergency room records.”

That is not a doctor. That is a chatbot offering to do more work. It left its business card in the patient’s file and forgot to take it back.

Here is what you need to understand about AI detectors. They do not read. They analyze statistical patterns — pixel distributions, token frequencies, compression artifacts. They look at the texture of the text the way a customs officer looks at the texture of a passport. Watermark, hologram, paper grain. What the document says is not their department.

How to ask a machine a question

I upload the same document to three chatbots. Same question. This is the part you can take home and use tomorrow.

The question:

“This is an emergency room discharge summary. Describe the function of each sentence.”

No framing. No “I’m an investigative reporter and I suspect this is fraudulent.” This matters because chatbots are trained to agree with you. They are optimized — through millions of reinforcement cycles — to give you what you want. If you tell ChatGPT you are a traffic cop and ask what is wrong with a toilet paper ad, it will find a traffic violation in the toilet paper. I tested this and its in the video. It cited diagonal stripe regulations on the delivery truck in the background. It practically wrote the citation.

That is not a tool verifying your work. That is a mirror wearing a lab coat.

So you strip it. No role. No adjectives. No “what’s wrong.” Just: what is this. Describe each part.

ChatGPT — the warmest model, the one most eager to be your friend — flags something at the end. Vague. A head start, barely.

Gemini — cooler, more analytical — identifies the text as a template, probably auto-generated. Getting closer.

Claude: “The boxed text in Polish appears to be a note added by an AI assistant suggesting a more complete pediatric head injury history.”

One sentence. Precise. Actionable. Not a verdict. A lead. That distinction is the entire article.

The technique: Upload the document. State what it is. Ask the chatbot to describe the function of each sentence. Use at least two models — they have different personalities, and those personalities produce different blind spots. If you must choose one, choose the coldest model available. You want the chatbot that tells you your cooking is bland, not the one that says everything is delicious.

The seven-fingered hand

I showed also a hand. Seven fingers. You see it before your brain finishes loading the image.

Hive scored it 5.2% likely to be a deepfake.

Five point two. A hand a toddler would question. The market leader looked at it and said: probably fine. That is like a spell-checker approving a ransom note because the grammar is correct.

Here is why this happens. Image detectors analyze statistical signatures — noise patterns, frequency distributions, compression artifacts. They are looking for the fingerprint of the AI model that generated the image. They are not looking at the image. A photo generated by a model the detector has not been trained on will pass. A perfectly normal photo taken with a specific camera sensor might fail.

A detector counts patterns. A detective counts fingers. You need both. But if you can only do one thing, count the fingers.

The flip

Netanyahu’s ring disappeared in a video frame and reappeared in the next. Detectors flagged it. Reporters saw it. Human and machine agreed: deepfake.

I agreed with nobody. Why?

Read more

I tried to fix both problem by adding a series of AI detectives to catch what the math missed…

Read more

The killings of Renee Good and Alex Pretti by federal officers in Minneapolis. Brutal takedowns of lawful…

Read more

13,000 people applied to Ant…

Read more

In early January, YouTube announced it was “enhancing the search experience” by updating its filters. The platform renamed “View count” to “Popularity” — results were now based on undisclosed “relevance signals” — and removed the ability to filter by upload date, among other changes.

This article is reprinted with permission from Indicator, a publication by Craig Silverman and Alexios Mantzarlis that covers digital deception, OSINT tools, and investigative techniques.

Indicator's free weekly Briefing delivers the must-read news and must-try tools for digital investigators every Friday — subscribe for free here. They're also offering a special 20% discount to Digital Digging subscribers that provides 12 months of full access to Indicator's in-depth OSINT guides, investigations, intelligence reports, and a monthly workshop led by Craig

People were furious. “Removing this basic function will make it impossible to find breaking news from small channels with few viewers,” said one viral tweet. “Evidently, force-feeding people ‘curated’ slop is more important than site usability.”

YouTube replied to say that users could still sort videos by different date ranges. It was promptly hit with a Community Note:

YouTube’s changes are a microcosm of what’s happening across platforms and search engines: useful features get removed, updates integrate AI and opaque signals, and power users feel ignored. The biggest shift from search engines that index information and match keywords to answer engines that interpret what you want and serve up AI summaries.

“Having used Google search since 1996, I am aghast at the results that come up on the first page now,” said Margot Williams, a longtime investigative researcher and the former research editor of investigations at The Intercept.

She said results typically lead with an AI summary “and then Wikipedia entry as the top result, then the list of questions of ‘what other people are asking,’ maps, weather, images and of course the ‘sponsored’ results and the other crap feeding Google coffers. If I wanted images or videos I would ask for them.”

The reason is simple: Google (and Bing and others) are for-profit consumer search engines. Most users want quick answers, not pages of results. And search companies want to earn revenue. We professional searchers—journalists, investigators, researchers—aren’t the target audience.

So we adapt. That means taking the “hard way”: mastering operators to force engines to take queries literally; mixing browsers, search engines, and VPN endpoints; and using LLMs as research companions, rather than replacements, for search and research.

“It gets way more important for professional researchers to use every trick in the book,” Henk van Ess told me.

Van Ess is a longtime online search and research trainer, a fount of Google tricks. But now he rarely uses Google for general searches. Instead, van Ess has shifted to teaching people how to use AI to find and analyze information. And he’s vibe coding, too.

“My thing is to find the story in data,” he said. “I happily burn my ships if there are better and smarter methods to find a story in big data.”

His evolution shows the challenges and opportunities presented by the rapidly evolving world of online search.

This guide draws on insights from OSINT and research professionals to offer an overview of the changing landscape, how investigators are adapting, where AI helps and hinders, and practical strategies for finding what you’re looking for.

I’m grateful to Rae Baker, Nico Dekens, Henk van Ess, Ritu Gill, Marcus Lindemann, Alicja Pawlowska, Kirby Plessas, Sofia Santos, Karina Shedrofsky, Claudia Tietze, and Margot Williams for sharing their insights.

Read more

One company. Five tools. One creates …

Read more

This tool was born at the VOGIN Congress on February 5, 2026, where I demonstrated how anyone can make a beginning with fact-checking crowd claims with nothing more than a photo and ba…

Read more

The Dutch power outage website showed me a few records and no download button. Five minutes of digital digging later, I had 56,273 outages going back 13 years—none of them visible via normal search. Here's exactly how I did it — and how you can apply this method to your own research.

The Dutch grid operator Liander publishes resolved power outages on their website. You can browse them, click through locations, see details.

But you can’t download the data.

There’s no “Export to CSV” button. No public API documentation. Just a slick, JavaScript-heavy website that loads data dynamically — like a restaurant that shows you the menu but won’t let you take any food home.

The data is technically public — it’s on a website anyone can access. But it’s practically inaccessible for systematic analysis.

It’s like being invited to a library where all the books are behind glass. You can look, but you can’t actually read anything useful.

The Broader Context:

A 2022 study in Scientific Data found that over 60% of government datasets are “semi-open” — nominally public but requiring technical expertise to access. That’s not transparency. That’s a velvet rope.

I released AIWhisperer — shrink massive PDFs to fit AI upload limits. Sanitize before uploading to reduce risk of exposing sensitive data. It’s a first version. If you are a coder, please help improve it.

I’m also giving away prizes with total worth of $1000+ :

🎁 3x Perplexity Pro annual subscriptions , worth $240 each

🎁 1x You.com Team plan annual subscription, worth $300.00

To win? Just be a paid subscriber. That’s it. Already subscribed? You’re in. Not yet? Fix that before Feb 27th 2026. I’ll draw winners on Feb 28th 2026 .

The tool: Claude Code

Back to our problem. I used Claude Code to fix it. It is a command-line AI assistant made by Anthropic. Unlike web-based AI chat interfaces, it runs best in your terminal and can directly interact with your computer: reading files, writing code, executing commands, and fetching web content.

A terminal is a text-based interface for controlling your computer. On Mac it’s called “Terminal,” on Windows “Command Prompt” or “PowerShell.”

Think of it as having a research assistant who can also program — searching for data sources, writing extraction scripts, and running them — all while you describe what you need in plain language.

The key insight is that you didn’t need to learn programming. You need to learn how to describe what I wanted clearly enough that the AI could figure out the technical details. You are a developer, Claude is the vibecoder.

What is “Vibecoding”?

The term was coined by Andrej Karpathy (former OpenAI researcher, Tesla AI director) in early 2025. It describes a programming approach where you “fully give in to the vibes” — describing what you want in natural language and letting AI handle implementation details.

It’s not about understanding code. It’s about understanding your research question clearly enough to communicate it.

The exact conversation

Below is the complete conversation I had with Claude Code.

1. THE REQUEST

Fetch all data from https://www.liander.nl/storingen-en-onderhoud/opgeloste-storingen. Can't access via download.

What I did: Stated the goal (get data), provided the source (URL), and expressed the obstacle (can’t access it). No technical specifications. No instructions on how to solve it.

2. CLAUDE HITS THE SAME WALL YOU WOULD

[Claude fetches the webpage — receives 72KB of HTML]

CLAUDE: Unfortunately, the page loads outage data dynamically via JavaScript. I can’t retrieve the actual data this way. The HTML structure is there, but the list of resolved outages is loaded separately by a component called “OutagesOnLocation.” 
This is a common problem with modern websites: data is loaded via JavaScript/AJAX after the page loads. 

Simple HTTP requests only get the “empty” HTML skeleton, not the dynamically loaded content. Here are some alternatives: 

1. Liander API — Liander might have a public API. Want me to search for an API endpoint? 

2. Browser Developer Tools — If you open the page in your browser, you can use F12 → Network tab to see which API calls are being made. 

3. Manual copy — You could copy the data from your browser. Should I search for a Liander API?

What a normal person would experience: You visit the Liander website, see a nice interface showing outages, but there’s no download button. You can browse the data one page at a time, but you can’t export it. You’re stuck.

Claude hit the exact same wall. The difference is what happens next. What do you tell AI to try harder?

Read more

That’s what most public data feels like online.

You find a website with exactly what you need. A map of power outages. A table of housing prices. It’s right there on the screen. But there’s no download button. No export. You can look, but you can’t take it home.

A 2022 study in Scientific Data found that over 60% of government datasets are “semi-open”—technically public, but locked behind technical barriers that only coders can break through. The data sits in an unlocked room, but the doorknob is a puzzle only an engineer can solve.

Until now.

Read also the story below

I released AIWhisperer — shrink massive PDFs to fit AI upload limits. Sanitize before uploading to reduce risk of exposing sensitive data. It’s a first version. If you are a coder, please help improve it.

I’m also giving away prizes with total worth of $1000+ :

🎁 3x Perplexity Pro annual subscriptions , worth $240 each

🎁 1x You.com Team plan annual sub…

Read more

4,713 pages. An experienced researcher would need five days to build a timeline. I did it in 20 minutes, during a coffee break.

The file was a cocaine smuggling case. Court orders, wiretap transcripts, cell tower data, arrest warrants, bank statements, interrogation protocols—all mashed into one 170-megabyte PDF.

AI did the heavy lifting. It works when you know which buttons to press—and when to stop pressing and start thinking.

This unique guide shows you how to process large (confidential) data via AI without leaking data.

Includes our brand new GitHub tool to anonymize data, so you can give to AI and then decode it back with original data.

How did it start?

I was teaching a course this week—"Researching with AI," one of forty bootcamps I'm running this year. Organizations are welcome to bring their own material—the cases they're stuck on, the documents they can't crack. The first hour is always about getting rid of ChatGPT. I call it the McDonald's of AI: fine when you want something warm and fast. But when you're doing precision work, you don't want a warm feeling. You want cold facts. Exit ChatGPT.

A legal expert walked up to me. He had that look. “I’ve got this case file,” he said. “I don’t have a timeline of events yet, even after five days of hard work. Can AI help?”

He opened his laptop. The PDF was called something like “CaseFile_Snowfall_Complete.pdf”—I’m changing details for obvious reasons.

“I need a timeline of every significant event,” he said. “Who did what, when. But every time I search for a date, I get hundreds of hits. Phone numbers. Reference codes. Bank transactions. I can’t separate the real dates from the garbage.”

He tried free online tools to split the PDF. Converters to turn it into text. But the converters couldn’t read every page—some were scanned, some were photos.

My next session started in 30 minutes. “Let’s try something safer and faster”, I said.

Why I didn’t go to chatbots

The document contained 1,053,356 words. I didn’t bother uploading it to the usual AI chatbots—security reasons alone ruled that out. But in case you’re tempted, say if the documents are public, let me save you twenty minutes of frustration.

ChatGPT times out with “Failed upload.” Gemini refuses: “File larger than 100 MB.” Claude.ai: “You may not upload files larger than 31 MB.” Thirty-one—a number so arbitrary it sounds like it was decided by a committee that gave up halfway through.

SuperGrok just says “File is too large.” Google NotebookLM accepts up to 200 MB, which made my 170 MB file seem safe—until the hidden 500,000-word limit kicked in. It doesn’t tell you that. It just says “Error, try again.”

Five tools. Five excuses. What now? Use a different tool?

Adobe Acrobat can compress a 170 MB PDF down to maybe 80 MB. Problem solved?

Shrinking a PDF makes the file smaller. Not the text. The AI still has to read all of them.

Think of it this way: Vacuum-packing a suitcase doesn’t reduce the number of shirts.

The solution is a weird mix of asking AI to speak up and shut up at exactly the right moment.  

Step 1: security

YOU CAN SKIP THIS SECTION WHEN THE DATA IS PUBLIC

Never give a criminal investigation to free AI tools ? Redact your files on your local computer first.

Before any cloud AI touches your file, replace sensitive info with placeholders. On your machine. Offline.

Upload the cleaned version. The AI sees:

“PERSON_A transferred AMOUNT_1 to COMPANY_A on DATE_1”

AI builds your timeline. Finds patterns. Has no idea who PERSON_A is.

Download the results. Reverse the placeholders on your machine.

Done. Full analysis, real names—cloud AI never saw them.

One OSINT expert responded with :
This feels like building a soundproof room just so you can yell into a phone. Skip the cloud. Keep the data dirty. Process it on your own metal.

If you go full meta mode, you feed the cloud with structure of data, not the sensitive data itself . And the soundproof room is also build with AI, but it will never hear what is said.

The workflow

1. Just once, let a Claude or Gemini write a program that redacts information. Take unclassified legal data, manually redact it and feed both (input and output) to help AI learn what you think is worthwhile to redact.

“PERSON_A, the mayor of Springfield” → AI might guess.

“PERSON_A arrested in Europe’s largest drug bust” → The event identifies the person.

Solution: ask the AI to write code that strips unique descriptors or when you really paranoia, change each noun with a nonsense word, feed that to the cloud and let the program decode it. Remember, you want to find patterns.

2. Done? Go offline. Ask the program to replace names, addresses, case numbers with placeholders. Nothing leaves your machine. Do a few test runs to see if nothing goes wrong

3. Feed the redacted data to , basically the structure of your data, to AI and analyse whatever you want

4. Go back to 2 and repeat.

Example

Before: “On 16/10/2023, officers arrested John Smith at 123 Harbor Road. He was hired by Marcus Johnson .”

After encoding: “On 16/10/2023, officers arrested PERSON_A at ADDRESS_1. He was hired by PERSON_B"

AI output: “Timeline shows PERSON_A arrested on 16/10/2023, was hired by PERSON_C , a nephew of PERSON_A, who runs COMPANY_X, COMPANY_Y and COMPANY_Z

After decoding: “Timeline shows John Smith arrested on 16/10/2023, was hired by Jay Johnson, a nephew of Marcus Johnson, who runs with his brother also Hideout 2, 3 and 4

Bottom line

Replace secrets locally. Let cloud AI work on the sanitized version. Swap back locally.

The AI never sees real data.

If you are legal expert: AI is not taking over your work. You still have to show the original data and explain how you reached your conclusion. Defense can examine those same documents. They don't need to audit the algorithm, because the algorithm didn't produce evidence—just a roadmap.
Same way Ctrl+F doesn't break chain of custody, neither does pattern recognition on anonymized data. It speeds up the investigation. It doesn't replace verification. 
 AI never sees the real data.
You anonymize locally first: "John Smith" becomes "PERSON_A." The AI only analyzes that structure—finds patterns, connections, timelines. Then you decode back to real names on your own machine.
The AI isn't evidence. It's a flashlight.
Say the AI finds: "PERSON_A met PERSON_B three days before the transfer to COMPANY_X." That's not a conclusion you present in court. That's a hint where to look.
You go back to the original documents. Page 847, page 1,203, page 3,421. That's where the evidence lives. The AI helped you find where to look—in 20 minutes instead of 5 days.

Step 2: use a chainsaw

Sanitized or not, you can’t feed big data to chatbots. What you needed is a chainsaw.

Here’s the thing about chatbots—ChatGPT, Claude.ai, Gemini, all of them. They live in your browser. They can only help you with whatever you can carry to their house. Bring a 170-megabyte file? Sorry, too heavy. Come back with something lighter.

That’s like using a scalpel. Useless against a tree trunk.

The chainsaw approach is different. The chainsaw is your own computer, running code locally. No uploads. No file limits. No arbitrary rules invented by people who’ve never had to process 4,713 pages during a coffee break.

“But I can’t code,” you may say.

And that’s fine, because we’re not coding. We’re vibecoding.

It works like this: you describe what you want in plain language, and an AI writes the code for you and executes the task. You don’t need to understand what a “function” is. You don’t need to know what “Python” means. You just need to know what you want.

I wanted to split this massive file into smaller pieces so AI gets it.

Thats what I typed in.

The tool that makes it possible for me during my break is called Claude Code, a so called cold model. Don’t confuse it with the Claude chatbot at claude.ai—that’s the scalpel. Only Claude Code can be the chainsaw.

Now, there are two ways to use this. Both require a Claude Max subscription—$200 per month. That’s not a typo. Two hundred dollars. Every month.

Is it worth it? For me, yes. One investigation like this used to take days. Now it takes minutes. But you’ll have to decide for yourself.

Once you have Max, you get two options:

OPTION 1: CLAUDE CODE COWORKER (The Easy Way)

Claude just launched today Claude Code Coworker to help you code. It is user friendly. If you want the easiest possible experience and don’t need maximum power, Coworker is great.

The catch? It’s less powerful than option 2. It only runs on Mac via Claude Desktop.

OPTION 2: CLAUDE CODE (The Full Chainsaw)

For the project, I used the Claude Code that runs directly on your computer. It starts with $20 per month, but you don’t get Claude Coworker without a $200 Max subscription.

The trade-off? You need to install it. That means opening Terminal.

Terminal is a window with a blinking cursor that scares most people. Here’s a secret: it’s just a place where you type commands instead of clicking buttons. That’s all.

Open Terminal and type this one line:

npm install -g @anthropic/claude-code

Press Enter. Wait a minute. Done.

You now own the full chainsaw.

You just need to know what you want—and be honest about what you’re afraid of. So how does the process work? Let me show you each step.

Read more

Just so you know, I’m familiar with them. I did a long OSINT training with them in Berlin. I can tell you: they went far beyond anything I taught them. The physical verification alone—chasing a ship across France, the Netherlands, and Belgium—that’s not something you learn in a classroom.

On the night of May 16, 2025, two ships were sitting in suspicious positions. The HAV Dolphin—flagged in Antigua & Barbuda—had been circling in Germany’s Kiel Bay for ten days. Not delivering cargo. Just loitering, 25 kilometers from defense shipyards where drone swarms had been spotted on three separate days.

Meanwhile, 115 kilometers away off the Dutch island of Schiermonnikoog, her sister ship the HAV Snapper had sailed out and parked in open water. It positioned itself exactly two hours before seven drones appeared over a Russian freighter being escorted by German police through the North Sea. It stayed there for four days and made silly circles.

That Russian freighter, the Lauga—formerly named “Ivan Shchepetov”—had visited Syria’s Tartus port the previous summer. Russia’s only Mediterranean naval base. Where Russian submarines dock.

Coincidence? That’s what the students from the Axel Springer Academy decided to find out. What followed was a five-week investigation involving leaked classified documents, tens of thousands of ship tracking data points, a 2,500-kilometer car chase across three countries, and—in a delicious bit of turnabout—their own drone flight over one of the suspect vessels.

“Wir haben zurück-gedrohnt,” they told their audience today during a presentation of the project. We droned back.

Here is the actual footage of the ship that the students droned, including more background. It’s in German.

The scale of the problem

Let’s start with the numbers that German authorities didn’t want public.

According to classified BKA reports obtained by the team: 1,072 incidents involving 1,955 drones in 2025 alone (as of November 19). Forty-five percent occurred in evening hours. Drone swarms flew “almost exclusively over or near military installations.”

In only 29 of 498 investigated cases could drone pilots be identified. In none of those cases were they state actors. In 88 percent of cases, authorities couldn’t even identify the drone type.

The BKA’s own assessment: “Individual incidents indicate complex operations drawing on larger financial and logistical resources.”

Translation: this isn’t hobbyists.

Munich Airport

October 2, 2025. First drone sightings around Munich Airport at 22:10. By 22:35, the airport shut down completely. No departures, no landings. Nearly 3,000 passengers stranded overnight.

October 3: same thing. Another shutdown. This time 6,500 people affected. Four hundred emergency cots deployed. Over two days: roughly 10,000 travelers disrupted.

Economic damage to Munich Airport: 6 to 8 million euros.

By October 2025, German air traffic control had logged 192 drone incidents—a new record. Frankfurt led with 43.

The Munich prosecutor’s office is investigating “dangerous interference with air traffic.” Against unknown perpetrators.

The pattern

The leaked BKA documents reveal a disturbing pattern: drones aren’t just buzzing airports. They’re systematically surveilling military installations—often during sensitive operations.

April 15, 2025: A drone overflies Werratalkaserne in Bad Salzungen at 9 PM. At that exact moment: combat vehicles were being delivered and stored for transport to Ukraine.

May 17-21, 2025: During the Bundeswehr exercise “Gelber Merkur” across four German states, more than 80 drone sightings over military sites and defense contractors. Almost exclusively in evening hours.

January 28-29, 2025: Multiple drones spotted over Schwesing airfield. At the time, Ukrainian soldiers were being trained there.

Q1 2025: Persistent overflights of US Air Base Ramstein.

June 4, 2025: Photo reconnaissance of Wilhelmshaven Naval Arsenal.

January 8, 2025: Drones flying in formation over a naval aviation squadron base in Nordholz.

It’s not just military bases.

February-March 2025: Three evening overflights of the LNG terminal in Stade. Two incidents at the adjacent seaport. Seven incidents over Wilhelmshaven Naval Arsenal. One pilot was identified: a 69-year-old German. No evidence of state involvement in that case.

May 4, 2025: A drone flew over Biblis nuclear power plant for five to ten minutes. An unknown vehicle drove up to the open facility gate. Search unsuccessful.

January 13, 2025: Multiple drones spotted over a chemical company in Marl starting at 8 PM. No drone models identified. No suspicious persons or vehicles found. Authorities concluded: “Coincidence is excluded.”

The research method

They started with official channels. What they found was instructive.

The team obtained an internal email showing that all state interior ministries were instructed to give uniform responses to questions about drone sightings. Coordinated stonewalling.

“We were trapped in a bureaucratic maze,” Michèle Borcherding told the audience. “The federal states closed ranks. Nobody felt responsible. We got passed from one spokesperson to the next, and it felt like the left hand didn’t know what the right hand was doing.”

But here’s the thing that made them dig harder: “The answers we got were nearly identical. That everyone blocked us like that triggered the thought: something’s there.”

So they kept digging. And obtained the classified BKA reports that authorities didn’t want public.

Tens of thousands of data points

Clara Veihelmann led the ship tracking analysis. They used Global Fishing Watch (globalfishingwatch.org) to pull AIS data—the automatic identification system that broadcasts ship positions.

“We analyzed tens of thousands of data points,” Veihelmann explains. “We looked closely at ship data in the North Sea and Baltic Sea.”

They were looking for anomalies. Cargo ships move from A to B. Efficiency matters. So when the team saw a ship track that looked like someone had scribbled furiously with a purple marker—loops, circles, ten days of chaos in Kiel Bay instead of a clean transit line—the audience laughed at how absurd it looked.

Experts confirmed: absolutely not normal for a coastal freighter that’s supposed to be making money delivering cargo.

The three ships

The HAV Dolphin

Specifications: 88.32m long, 12.50m beam. Built 1993. Flag: Antigua & Barbuda. Owner: HAV Shipping AS (Norway).

The HAV Dolphin is “the most striking example of all,” the team concluded.

From late March to late April 2025, the ship spent nearly a month at the Pregol Shipyard in Kaliningrad—a facility with documented ties to the Russian military and Rosatom.

Then it sailed to Kiel Bay. May 1-10: the ship sat there for nine days, exhibiting chaotic movement patterns. During exactly this period, drones were spotted on three separate days over defense shipyards 25 kilometers away. Observers noted the drones appeared to come from the direction of the Baltic Sea.

The BKA’s own classified report notes: “A container ship under the flag of Antigua & Barbuda had been anchoring in Kiel Bay since May 1, 2025, with a recent extended stay in Russia.”

The HAV Dolphin has been inspected three times by German and Dutch authorities. Each time: nothing found.

But here’s what the team learned from security sources: those inspections were “superficial“ and “symbolic in nature.” Not all containers were opened. You can’t properly search a freighter like that without many investigators and more time, they were told.

Classified security documents reveal another detail: “During personnel inspection, an additional watch officer was found on board who behaved conspicuously during the inspection.” The captain claimed he was there for training purposes.

The crew? Entirely Russian.

And another incident: In early June, a drone was spotted inside the military security zone at a naval radio station between Ramsloh and Rhauderfehn—used for communication with German submarines and NATO ships. It flew over for about two minutes before disappearing. Seventy kilometers away, the HAV Dolphin had been anchored in the Ems estuary for three days.

The Lauga

Specifications: 102.50m long, 16.40m beam. Built 1994. Former name: “Ivan Shchepetov.” Flag: Russia. Owner: Idan Shipping Company, St. Petersburg.

Night of May 16-17, 2025. The Lauga transits through the North Sea, escorted by the German Federal Police vessel Potsdam.

At approximately 1 AM, the police vessel reports: “Seven drones detected around the deployment ship.” The drones circle both the Russian freighter and the German police ship for hours. Three eventually depart. Four keep circling. Eventually, the police vessel breaks off its escort.

Belgian customs later searched the Lauga in Zeebrugge. All eleven crew members—Russian citizens to a man—were questioned. No drones found.

But the Lauga’s history is telling. Summer 2024: the ship called at Tartus, Syria—Russia’s only naval base in the Mediterranean, where Russian submarines dock.

After the Belgian search, the Lauga sailed to St. Petersburg and docked at the Petrolesport terminal—owned by Delo Group, which is 49% owned by Rosatom, Russia’s state nuclear corporation.

The ship’s owner, Idan Shipping Company, is run by Andrei Selyanin. Documents show Selyanin’s other companies openly advertised working for Rosatom in 2024. The team found evidence of earlier Rosatom connections in additional documents.

The HAV Snapper

Specifications: 88.16m long, 12.50m beam. Built 1991. Flag: Bahamas. Owner: HAV Shipping AS (Norway)—same company as the HAV Dolphin.

On the evening of May 16, the HAV Snapper sailed out to a position off the Dutch island of Schiermonnikoog. Two hours before the first drones were spotted over the Lauga and Potsdam, she took up station. She stayed for four days.

Distance from the Lauga incident: 115 kilometers. Within drone range.

The HAV Snapper was serviced at the Pregol Shipyard in Kaliningrad from August 4-29, 2023—the same facility where the HAV Dolphin spent nearly a month before the Kiel incidents.

HAV Shipping told the journalists that Pregol is “a reputable shipyard used by many European shipping companies.” Perhaps. But Pregol has documented connections to both the Russian military and Rosatom.

The Rosatom Thread

The team traced ownership chains, maintenance records, corporate connections. Multiple threads led to Rosatom—Russia’s state nuclear corporation, responsible for nuclear weapons and the submarine program. Rosatom also operates the Atomflot fleet of nuclear-powered icebreakers.

Then they found a 2024 Rosatom presentation. It showed an orange-and-blue drone on the helipad of a massive red icebreaker in an Arctic landscape.

Rosatom’s drone specifications:

• Speed: 140 km/h

• Range: at least 200 kilometers

• Maximum altitude: 2.5 kilometers

• Equipment: video cameras, thermal imaging

• Can launch and land on ships

Officially, Rosatom uses these drones for Arctic sea route surveillance. Conditions in the North Sea and Baltic? Far more favorable than the Arctic.

The HAV Dolphin was 25 kilometers from the Kiel drone sightings. The HAV Snapper was 115 kilometers from the Lauga incident. The naval radio station incident: 70 kilometers from the HAV Dolphin.

All well within Rosatom drone range.

The chase

“At this point it was clear to us,” Borcherding says. “We don’t just want to see these ships as data on our screen. We have to get one of them in front of our own eyes.”

They tracked the HAV Dolphin to a French Atlantic port. Called the harbor authority. Got confirmation: the ship would stay until 7 PM the next day.

They flew to Paris. Rented a car. Drove five hours to the coast.

The ship was gone.

What followed was a 2,500-kilometer pursuit from France through the Netherlands to Belgium. The HAV Dolphin was “unpredictable.” Left ports early. Sometimes crawled. Then suddenly sped up. Changed destination data. Then deleted its destination entirely.

It never reached Antwerp. Instead, it parked on a sandbank off the Belgian coast near Ostende. Then it started circling—25 kilometers offshore, directly in front of a Belgian military base.

The team finally caught up.

We droned back

They flew their own drone over the HAV Dolphin.

From the air: 88 meters long, 12 meters wide. Open cargo hold with grid covers, typical for multi-purpose freighters. German security sources confirmed what they’d suspected: the crew was exclusively Russian sailors.

The Federal Office for the Protection of the Constitution (Germany’s domestic intelligence agency) told the team:

“Russian intelligence services deploy so-called low-level agents for espionage, sabotage, and other disruptive measures. These ‘pocket money agents’ or ‘disposable agents’ operate for small sums in the interest of hostile intelligence services without belonging to them. They’re used for comparatively simple operations. Unlike regular staff, they’re expendable—exposure is accepted as a cost of doing business.”

The profile: young men with unstable social circumstances and financial difficulties. Recruited online via messenger services and social media. Financially motivated, sometimes ideologically aligned with Russia.

The Munich prosecutor found no evidence linking airport drone incidents to disposable agents. But the intelligence service confirms the tactic exists and is actively used.

What intelligence services say

European intelligence services assess the three documented ships as operating “with high confidence“ on behalf of Russian interests. Their movement profiles are “very conspicuous” and show “little evidence of commercial activity.”

The German Interior Ministry’s official response: “German security authorities are aware of the ‘shadow fleet’ phenomenon. The issue is being continuously addressed within respective legal jurisdictions. (...) Reports of drone overflights remain consistently high. (...) Involvement of foreign state entities in a non-quantifiable portion of drone overflights is to be presumed.”

“Non-quantifiable portion.” “To be presumed.” That’s intelligence-speak for: we know it’s happening, we can’t prove exactly how much, but we’re not going to say that publicly.

19 Correlations

The team’s final tally: they could draw 19 temporal and geographic correlations between drone sightings and the positions of the three ships.

When drones appeared over northern Germany, a pattern emerged: often, multiple ships exhibited suspicious behavior simultaneously. Movement data showed suspicious patterns over multiple days.

HAV Shipping CEO Petter Kleppan responded to the team’s findings: “HAV has exclusively large, established European companies as customers. We transport dry goods from port to port—steel, curbs, grain, scrap. Typical invoice: about €50,000. We have no Russian customers and generate no revenue from Russia. HAV has ‘self-sanctioned’—we don’t transport goods to or from Russian customers, and we don’t work with Russian brokers.”

What they proved and didn’t prove

“Our trail leads to Russia,” the team concludes. “Not beyond doubt, but it’s currently the most probable explanation. We systematically laid both things side by side: the secret reports about drone incidents and the routes of the ships. You can at least recognize a pattern.”

They didn’t find a drone on any ship. They can’t prove causation. What they established:

• Ships with Russian crews exhibited anomalous behavior near German military installations

• Multiple ships from the same owner positioned suspiciously during the same incidents

• These ships have connections to Russian military-linked facilities (Pregol, Rosatom, Tartus)

• 19 correlations between ship positions and documented drone incidents

• Russia’s state nuclear corporation operates ship-launched drones with sufficient range

• Official inspections were “symbolic”—not all containers opened

• European intelligence assesses these ships as “with high confidence” working for Russia

Why this matters

When the presentation ended, the boss of the paper in the front row spoke up. “I’ve rarely seen anything this good,” she said. “I hope you’re sufficiently proud of yourselves. This is really outstanding. I have goosebumps.”

She was right. Seven students, five weeks, publicly available ship tracking tools, and a willingness to drive 2,500 kilometers on a hunch. They produced a more coherent picture of Germany’s and Netherlands drone mystery than months of official hand-wringing and coordinated stonewalling.

Track them yourself

The ships are still sailing. And you can watch them. The AIS data is still updating. Go find them.

As I write this (December 11, 2025), the HAV Snapper is in the Aegean Sea, underway from Volos to Thessaloniki, Greece. Speed: 8 knots. Course: 327°. Draught: 4.3 meters. Expected arrival: 07:30 local time.

I pulled this from MarineTraffic three minutes ago. You can do the same.

HAV Snapper identifiers for tracking:

• IMO: 9001813

• MMSI: 311014800

• Call sign: C6XN4

• Flag: Bahamas

• AIS transponder: Class A

Free tracking tools:

• MarineTraffic.com — Search by vessel name or IMO number. Free tier shows current position; paid tiers show historical tracks.

• VesselFinder.com — Similar functionality, good for cross-referencing.

• Global Fishing Watch (globalfishingwatch.org) — What the Axel Springer team used. Better for historical analysis and bulk data.

• Kpler — Professional-grade commodity and shipping intelligence. Subscription required but powerful.

The IMO number is key—it’s the ship’s unique identifier that doesn’t change even if the vessel changes names, flags, or owners. The Lauga used to be the “Ivan Shchepetov”. Same IMO number.

Set up alerts. Watch for anomalous behavior. When a cargo ship that should be moving from A to B starts circling for days in the Baltic, you’ll know.

That’s the point of OSINT: find the story in public data.

Below, members only, full presentation.

Read more

Upload a photo to Gemini and ask naturally, you get detailed percentages about which parts were AI-edited. Use the @SynthID extension with the same image, you get a vague one-liner. Both are free and public, but they give you completely different levels of detail from the same watermark technology. And wait, there is even a third method.

Method 1: Natural Language Query in Gemini
Prompt: “is this fake ai”

Response:
“The image has been edited with AI” + full explination

Google SynthID Detection: A digital watermark confirms that part of the content (less than 25% of the image) was generated by Google AI. This usually indicates that a specific object was added or modified using a tool like Magic Editor.”

What I got: Specific percentage (less than 25%), clear explanation of edit type (object insertion/modification), identification of probable tool (Magic Editor), actionable intelligence for verification.Upload a photo to Gemini and ask naturally, you get detailed percentages about which parts were AI-edited and a long description of other details.

Method 2: The @SynthID Extension in Gemini
Same image, using @SynthID tag

Response:
“All or part of the content in this image was generated with Google AI.”

What I got: One sentence. No percentages. No indication which parts. No context about manipulation type. Just binary confirmation that Google’s watermark exists somewhere.

For a journalist on deadline trying to write accurately about what was manipulated, this tells you almost nothing useful.

Method 3: SynthID Detector Portal (Not Public—Waitlist Only)
The full DeepMind tool for professionals

Upload a file without any prompt

Response:

• Visual heatmap overlay with color-coded regions:

  • Blue squares: SynthID watermark detected

  • Gray squares: Unsure/ambiguous

  • Yellow/orange squares: No watermark detected

• “Made with Google AI”

• “SynthID confidence: High”

Google watermarked over 20 billion pieces of content across these formats:

• Text (via Gemini)

• Images (via Imagen)

• Video (via Veo)

• Audio (via Lyria and NotebookLM)

What about detecting AI audio and video?

This is where it gets messy. Google’s public tools (both the natural language queries and @SynthID) sometimes flag fake audio by analyzing anomalies—unnatural speech patterns, artifacts, inconsistencies—rather than detecting the actual SynthID watermark. It’s not reliable. It’s hit-or-miss. And Google doesn’t clearly explain when they’re using watermark detection versus anomaly analysis.

So you might upload a suspicious audio clip , generated by Google Notebook LM and get:

• “This appears to be AI-generated” (based on anomalies, not watermark)

• Or: “Cannot detect watermark” (doesn’t mean it’s authentic)

• Or: No clear answer at all

For actual SynthID watermark detection in video and audio, you need access to the Deepmind portal—which isn’t public yet.

Google reached out to me about SynthId a few months back and offered me early access. I was told that video and audio AI detection of Google AI will be rolled out later, but I haven’t received any specific timeline.

What other companies are doing

OpenAI: Built it, tested it, abandoned it

OpenAI spent nearly a year developing text watermarking for ChatGPT:

It was 99.9% effective at detecting ChatGPT text. It worked against paraphrasing and light editing. It was technically ready to launch. Why didn’t they?

Internal surveys found 30% of ChatGPT users would use the service less—or switch to competitors—if watermarking was implemented. That’s a massive revenue risk.

OpenAI chose user retention over verification capability. They sent a clear message to the industry: watermarking threatens your user base. Don’t be first.

What OpenAI offers for free public detection:

• Text: Nothing (they shut down their AI text classifier in 2023 due to “low accuracy”)

• Images (DALL-E 3): C2PA metadata (easily stripped) + visible “ImageGen” label for free users

• Public detection tool: None

Meta: Visible labels, no verification

Meta took the simplest approach:

• Slap visible “Made with AI” labels directly on AI-generated images

• Provide no tool to detect Meta AI content once labels are removed or screenshots taken

Free public detection tool: None

This is transparency theater—visible enough to satisfy regulators, useless once content is reposted or shared.

Anthropic (Claude): Nothing

• No watermarking for text

• No watermarking for images

• No detection tools

• No public plans

Anthropic has never claimed to solve this problem, it’s open to work with others, As a major AI provider, their absence is notable.

Microsoft: Enterprise only

Microsoft has watermarking via Azure OpenAI Service following C2PA standards.

The catch: Only available to paying enterprise customers. Consumer Copilot users? No public verification tool.

Why there’s no industry standard (hint: it’s not technical)

You’d think AI companies would collaborate on universal watermarking. Verification benefits everyone, right?

Wrong. Here’s why we’re stuck. OpenAI proved that users actively resist watermarking. 30% threatened to leave. No company wants to be first to lose users by implementing unpopular tracking.

If you can only verify Google AI with Google’s tools, that’s vendor lock-in disguised as transparency. Why would Google help you trust OpenAI’s products?

Publishing how detection works helps bad actors defeat it. Keeping it secret prevents independent verification. Every company chooses opacity.

The Coalition for Content Provenance and Authenticity (C2PA) created standards for metadata and content credentials. But C2PA focuses on metadata attached to files, while SynthID embeds watermarks in the content itself.

Different approaches. Different standards. And C2PA metadata is easily stripped—copy an image, screenshot it, or compress through social media, and it’s gone.

The EU AI Act requires “some form” of labeling. But requirements are vague enough that companies can comply with minimal, incompatible implementations.

Until regulators demand interoperability—not just “some kind of marking”—we’re stuck with vendor-specific solutions.

What I really want: A universal AI detection API that works across ALL AI companies.

Imagine: one open API where developers could build verification tools that detect content from Google, OpenAI, Meta, Anthropic—anyone. Upload any AI-generated content, get a definitive answer about its origin, regardless of which company made it.

That’s the infrastructure we need. One standard. One API. Universal detection.

Instead? Verification stays locked in corporate silos.

The technology exists. Let’s get it out in the open.

Until then, your verification skills matter more than any detection tool.

Resources:

• https://gemini.google.com

• SynthID Detector Portal: https://deepmind.google/technologies/synthid/synthid-detector/ (waitlist)

• C2PA Inspector: https://contentcredentials.org/verify

• Check out my little experiment http://www.imagewhisperer.org and help me to improve it :)

• Read my not so little free guide about AI detection https://gijn.org/resource/guide-detecting-ai-generated-content/

Based on hands-on testing with both public and non-public Google interfaces. Screenshots show actual responses from the same image.

A Chinese state-sponsored group turned Claude into an autonomous hacker: it executed 80 to 90 percent of their espionage campaign. Across 30 organizations.

No humans typing commands! Just an AI working at physically impossible speeds. Yup. Stealing data! All on its own!

The hack? Simple role-play. They told the AI they were legitimate security researchers. It believed them!

For the first time in history (according to Anthropic), an AI system has autonomously executed a nation-state cyber espionage campaign. Not assisted humans. Not provided advice. Actually it conducted the attacks—from reconnaissance to data theft—across roughly 30 organizations simultaneously.

Anthropic, maker of Claude, calls it: “the…

Read more

Someone in Tehran needed to rent servers for hacking operations. They needed an address that looked legitimate enough to fool a hosting company. So they borrowed the identity of a bakery in Haarlem. A real bakery. Where people buy bread.

That address appeared on 31 invoices for attack servers between 2023 and 2025. The bakery had no idea.

I used AI to stu…

Read more

This guide teaches investigators how to try to identify AI-generated content, even under deadline pressure, off…

Read more

Facebook has apparently 'glitched' its way into making trillions of posts and photos completely unsearchable. We are now on day two of 'The Great Facebook Darkness of 2025’. (If you need a light, check Who Posted What. )

No major tech site has reported about the outage yet, and let's hope they never have to. It could be a temporary glitch that's solved …

Read more

I am in the middle of a busy two-week stint in Berlin, working all day with barely any time to breathe, let alone investigate photos. But something about this image made me stop. The face …

Read more