In the Netflix documentary Freefall I watched the father of a crash victim try to read the database of aircraft defect reports kept by the FAA, the US aviation regulator. It took him hours. With my new tool it takes seconds.

I entered the tool in Cerebral Valley‘s GLM-5.3 Flash hackathon. The Foundation for Aviation Safety wrote to me. Ed, from the Foundation, after a first look: “It looks impressive. I’m hoping to dig into it further.”

They asked for a walkthrough. That is the test of a tool like this: whether the people who work on the problem every day want to use it.

It is not only for them. Anyone about to board can type a flight number and a date and read what mechanics wrote about that aircraft, and how often it came back for the same thing. A journalist gets leads: the same part written up on several aircraft of one airline on one day, or a box ticked that contradicts the write-up under it. A researcher gets 1.76 million reports readable at once, and can ask them questions in plain words. This 28 page guide is about how you can fix public data for about $1 in just 2 days time.

Relatives of victims were attempting to find answers to their questions partly via a governmental website. Row after row of capital letters. Zone numbers. Single character codes. It’s called Service Difficulty Report and is maintained by the Federal Aviation Administration. When technicians in the United States find an issue with a plane (e.g., a crack in a component; rust under a piece of equipment; or a failed seal), they file it. These reports are then made available to everyone. You don’t need to sign in, pay a fee, or submit a request for records. There are 1,757,828 of these reports since 1995 through today. They relate to 54,634 separate planes.

I was shocked.

The reports are so bureaucratic that almost everything has been replaced by codes, and the codes conflict. A report will never tell you there was an issue with the landing gear. It will say ZONE 700. An emergency landing? That is just an A. But the same letter elsewhere on the form means the report came from an airline. For researchers, the database is a nightmare. I saw the father of one of the victims searching it, trying to make sense of it.

It was something public, but the government presented it as random unlabeled ingredients on a table. Over 1.7 million records were hidden behind bureaucratic codes. I want to make this completely transparent, to help families and researchers.

An airline’s name is made up of letters such as CALA. The technician’s finding is a completely different code. The code used to identify how the technician found the issue initially is a third code.

When I attempted to conduct a search using a legitimate query, I encountered timeout and latency issues. Typical IT issues with software that has gone unmaintained for a considerable period of time. Creating links between two pieces of data were difficult.

But I have questions I want answered. What types of defects were reported yesterday? Which companies reported the most? At what time did they report them? Who were the individuals involved? What location on the airplane was affected? What type of defect was reported?

What action did the crew take in response to the reported defect? Without knowing the “when,” “what,” “where” and “who,” I cannot begin to determine “why.” Why does a single operator send the same airplane out for repairs 60 times in a year? How can that occur?

In the past (read 2025), I would download all the data into a spreadsheet, create a pivot table, rotated it one direction, rotated it another direction, and hoped the spreadsheet did not crash due to its size. In 2026, I type what I desire in my own language and AI creates it for me if I am precise enough. I build about two original, highly hyper-focused research tools a week for myself and my clients.

Imagination is the limit to creating things, not your ability to remember formulas or knowing how to code. If you can express what you desire, you can create it. How does that work? How did I build a full fledged tool?

A black box with white letters

A majority of users who use AI will probably be doing so by way of a chat box in their browser. If they paste a document into the chat box, the AI will read it. If they paste forty documents into the chat box, some will drop off. If they ask the AI to create a complex tool, they will receive code that they must save, install and run themselves.

But if the same AI is running within a terminal (Windows, Mac, no download needed) it can access your file system, download anything that is required for execution, execute it, observe any errors and correct those errors. The latter portion is the key to this process. After completing its task, the AI observes the outcome of its work and attempts to complete the same task again. You act as an intermediary to convey information: do this, do that. If the terminal and AI is installed on a server, you can post the output directly onto the web. So out of a window that looks like it came from 1989 (it does) you get a working tool on your own machine, or a website anyone can visit. The limit is no longer what you can build. Can you describe what you want?

Let’s start building the tool in the Terminal app. For a researcher who cannot code, the barrier was never imagination. It was having to learn Python, to structure a project, to build in fail-safes, to know how not to leak your own material. Some of that still matters. The part where you personally type the lines of code does not.

I used the brand new version of the model, GLM-5.3 Flash. There was already a first version, built with Claude Code, but it only tested one thing: could the database be made accessible at all? The answer was yes.

It answers back in the same green characters. Yes, that is possible. Should I show you how?

It warned me about two of my questions. The who means airlines, and the airlines are hiding behind odd-looking codes across 26 years of history, some of them obsolete and some not. The what means every engineer writing up a repair in a private thicket of abbreviations, which will all have to be looked up. The where is the awkward one, because a report will often describe a repair without ever saying where on the aircraft it happened. The location is implicit. The mechanic knew. The AI found a list of FAA codes, I checked it, and granted it to rebuild the public database in human language.

During a hackathon of GLM-5.3 Flash , I decided to go way further. I found the information still overwhelming. How do I change that?

The first picture that arrived in my head was a schematic aircraft, seen from the side, sitting at the top. Can AI build that for me? And how do I explain what I want to see? I asked:

Read more

I found eight ways we slip up, and they are all the same slip: part of your answer is already in your question, and you don’t notice — but the machine does. It finds the answer sitting in your question and hands it straight back.

It is the photocopier you keep swearing at because there is a smudge on every single page it produces. The smudge is on your original. The machine is doing the one thing it is built to do, which is give you a faithful copy of exactly what you fed it — cleaner, sharper, five hundred words longer, and just as wrong.

Half a million questions, nineteen languages, 46,287 exchanges rebuilt into question-and-reply, and finally 9,000 judged one at a time. I wasn’t looking for the questions that ask, but for the ones that steer the answer. Why does the report ignore the obvious cost problem — that one steers. Whether there is a cost problem was never checked; it rode in as furniture, and the machine built on it.

When a question carries a claim inside like that, the chatbots let it stand in 19 out of 20 replies.

This is the third piece in a series about a large pile of other people’s chatbot conversations. Part one was about what people gave away in them without meaning to — pension numbers, visa dates, a man asking whether this could stay between us, on a page anybody can still open. Part two was about the rabbit holes the machines pulled people into, including thirteen hundred messages of somebody’s imaginary physics.

How do we slip up?

Eight mistakes, each with its own fix below. But the fixes are all one habit applied in eight places, and the habit comes down to five rules.

Read more

A man spent thirteen hundred messages asking ChatGPT to check his theory about black holes. The AI chatbot told him he was ahead of 90 percent of published physics. It was not physics. It was madness. The machine never said so.

I let my computers analyze over 1.8 million AI chatbot conversations for weeks, to answer one question: how deep can a person sink into a rabbit hole before the machine says stop?

This conversation is one of nearly 112,000 I analyzed: 1,813,570 real messages with ChatGPT, Claude, Copilot and Gemini. Some run for weeks. Some end with the user more certain of something untrue than when they started.

I had four further questions about those conversations, and this piece answers all four.

Does the machine pull people deeper, or do people pull themselves? Is this one company’s problem, or does every chatbot do it? Is it rare, or is it the ordinary output? And can you see it happening while it happens to you?

Part one was about how these conversations became public. That story was about what people gave away. This one is about what the machines gave back.

Let’s start with what happens when a rabbit hole has no bottom.

Read more

Google has rolled back the image generation feature it launched in Google Earth on 30 July. The tool let anyone zoom to any coordinates, type a sentence, and get back a photorealistic image built on Google’s own satellite, aerial and 3D imagery of that spot. It went live worldwide, on the web, with no waitlist. It was withdrawn on 31 July.

Google’s state…

Read more

Read more

PART 1 of our investigation.
See also Deep Inside the Rabbit Hole (Part 2)

A husband whose wife had just left him, opening a conversation in a chatbot with: "Can this stay between us?" A retirement letter carrying a technicians full name, his phone number, his home address and his national ID number. A single line asking whether bleach can cure cancer.

All three are public web pages, still available today, that the people in them almost certainly thought were private. This year alone, across six chatbots — ChatGPT, Claude, Gemini, Grok, Copilot and DeepSeek —Digital Digging found 96,477 shared conversations in Archive.org, some of them readable as a copy , most of them available at the source right now. It looks like people thought "share" means send. But it means publish.

How the chats reached a public archive at all is a genuine puzzle. The obvious answer — that people posted the link somewhere open and a crawler swept it up — barely holds: of these thousands, we could find only a handful ever posted in public. So most of them reached the archive with no public trace of having been shared at all. Then we wondered whether these were older chats simply resurfacing now. They are not:

In 2025, the story was all about ChatGPT. Our confession files found the first of the leaked shared chats; the Washington Post, which asked to use our database, read 47,000 of them. They ran two articles on the story. (1 and 2). OpenAI pulled the shares out of Google, and everyone moved on.

The news now is simpler, and worse: it is all of them — not one, but six chatbots, still sitting open today in Archive.org, in July 2026, one link from anyone. And even ChatGPT is back in the list, although at the bottom.

All told, archive.org has logged more than a quarter of a million of these shared chats — 264,078 across the six.

What do people hand these machines? We learned a lot and will publish more about it the coming weeks. It is the ordinary person doing an errand: a letter, a form, a signature, filled in with a real name and address and ID, probably then shared without a second thought. We found the most identifiable information in Gemini, then Claude, then Copilot and then ChatGPT.

Let’s follow one of the conversations all the way down.

Someone in Germany is filling in the last piece of a five-year story.

They have lived in the country since before Brexit. Their visa is good until 2027. They have not left for longer than six months at a stretch — the box that matters. Now they are writing the supplementary letter that turns all of that into a permanent-residence permit, and they want it to sound right, so they open a chatbot and paste in the facts: the address on their registration, the visa date, the five years, the six-month rule they have never broken.

The bot helps. They get their letter.

Then, to keep it — or to show someone — they click the button in the corner. Share.

They think they have made a link. What they have actually made is a public web page that says a foreign national at this address has a visa running out in 2027, filed as part of a residence application. I know, because I read it. So can you.

One page out of many. And it was on Claude, not ChatGPT.

The button that lies

Picture a town square where you live. There is a board on the wall — the one with the lost-cat notices and the room-to-rent cards. Anyone who walks past can read it. Anyone who ever walks past can read it.

Now picture writing your problem on a slip of paper — your pension letter, your visa history, your kid’s swim-school invoice — and instead of handing it to the one clerk you meant to, you staple it to that board. And the button that does the stapling is labelled give this to the clerk.

That is the share button.

The gap between send this to my colleague and pin this to the town square is the story.

Nothing was hacked to make that page. Someone clicked share, meaning to hand a link to one person, and made a public document instead. The leak is not in the server. It is in the word.

How each page then reached a public archive I can only guess — a shared link is unlisted, so it normally has to get loose first: posted, forwarded, or crawled from somewhere open. But loose they got.

A phone book, not a library

You cannot guess these links: a share URL is a long random string, claude.ai/share/ and then a scramble nobody could type on purpose. Loose in the internet they are needles. Archive.org is what turns them into a list.

But a list is all it is. Open one of these shares on the archive and you mostly see nothing — a logo, a cookie banner, a sign-in wall. The captured page is an empty shell, because the conversation was never written into it: it gets painted in afterwards by code that phones a server, and on a frozen copy that phone call is dead. The archive tells you the chat existed, and where.

So the words in this piece came off the live pages, not the archive — and getting them was the actual work. Each of the six machines hides the conversation somewhere different, so each needed a different way in. Four of them gave it up to a browser patient enough to wait for the page to finish drawing itself. Claude was the one that fought: it runs a bot-check that turns away every automated visitor, and the only thing that got past was a real, visible browser with its robot badge stripped off, parked just off the edge of the screen so the guard read it as a person. The address book was the easy half; the notes at the end give the trick platform by platform.

Here is the part that runs against instinct: taking a page out of the archive changes nothing. People picture the archive as the vault — that if its copy is a dead redirect, or if they get it removed, the chat is gone. It is the reverse. The archive is a map, not the vault.

The chat lives on the chatbot, on the original public page, and that is the copy that stays open. Not always the archive keeps a readable version of a single recent share — every one a dead redirect — and yet nearly three in four of those pages are still live at the source right now, one click from anyone. Deleting the archive's copy does not close the page; the page was never in the archive to begin with. And deleting your own chat comes too late: by the time most people would think of it, the page has been public for months, and someone has already read it. Deletion is a gesture. Permanence is the fact.

What people said

Every case below is real, from the 2026 set, on the platform named. Every name, number, and address is masked. I am not linking to the pages: the whole point is to keep the next person out of the archive, not to walk you to the last one.

Read more

Google is rebuilding Search to answer instead of point. It still shows you the sources — it just decided, before you did, how few you’d need. Here’s the fix, and why the old operators matter more now than they ever did.

On May 19 Google announced the biggest change to its search box in twenty-five years: an AI that reads the web and hands you the verdict, plus background agents that search while you sleep. For a casual user that’s convenience.

For a researcher it’s a problem with a name — the AI chooses the sources for you. It still shows them, in a panel on the right, and you can keep digging the way you would in Perplexity. But it decided how many to read and which ones counted, where you used to wade through hundreds and pick for yourself. The fix is not a different search engine; it’s a different stance, and a return to the tools the AI layer is busy hiding: force the raw index with &udm=14, start from the primary source instead of the search box, and chain the operators — site:, filetype:, intitle:, after: — that turn Google back into an index you steer instead of an oracle you trust.

The problem, precisely

As if we didn’t have enough problems. Google’s normal search quietly dropped less-popular keywords. It favours popular content over exact matches. It bends results to your location. It caps you at a few hundred results no matter what. Those biases pushed the niche, the precise, and the inconvenient down where most people never look. That was the old complaint, and the answer was always the same: operators, to overrule the algorithm and make it show you what you actually asked for.

The AI layer fixes partly that problem but with a price. The first is a narrowed source base. The sources don’t vanish — Google lists them in a panel on the right, and you can follow up the way you would in Perplexity. The catch is that the answer often rests on a handful of pages the AI picked, and it made that call instead of you. Where you once chose how wide to cast — first result or three-hundredth — the machine now sets the width by default, and Pew tracked 68,879 real searches to find that just 1% of users ever open the source links to widen it back. The second is a polluted pool to pick from. A peer-reviewed study for the ACM Web Conference 2026 ran the experiment — when 67% of a search pool was AI-generated content, over 80% of what the system surfaced was synthetic, while the answers still read clean and confident. The lights are on. Nobody’s home.

It helps to picture why the web is going along with this. Imagine a theatre where the actors keep performing to an empty house. One spectator is left in the front row: a robot, taking notes, who will later tell the outside world what the play was about. So the actors start performing for the robot — shorter, louder, every nuance cut, because the robot doesn’t care for nuance. That’s what’s happening to the web. Sites no longer write for the person in the seats; they write for the machine in the front row. And the person only ever hears the robot’s summary.

I came across that machine recently. It was an AI dashboard from Microsoft Copilot, but it could have been anyone’s — a meter for publishers. A site used to watch visitors: how many came, how long they stayed, where they clicked. This panel watched two different numbers. Total Citations, how often the AI quoted your pages, and Cited Pages, how many of your pages it judged worth quoting at all. The counter that matters no longer counts people. So sites write to be quoted: compact, confident, authoritative in tone — which are exactly the qualities a researcher is trained to distrust. No hedging, no nuance, no “according to whom.” Then the AI sands off whatever nuance survived. You end up with an answer written to be cited, summarized by a machine that strips out the doubt, shown to you without the name of who said it. Three filters between you and the truth, and at each one a little of it goes missing.

Google used to send you out onto the web. Now it wants to be the web. The narrowed source base you can widen back with technique. The polluted pool you guard against by archiving what you find. Both start with refusing the pre-chewed answer.

Fix 1: strip the AI off the page

The fastest move first. Add &udm=14 to the end of a Google results URL and you get the classic ten blue links with the AI panels gone — no Overview, no “People also ask,” no carousels. Bookmark a version of it, or set a udm=14 search shortcutas a custom engine in your browser so plain, link-only Google is one keystroke away. The Web tab in Google’s own menu does the same job, one click deeper.

This is the smoke-detector rule. A useless smoke detector says “your house might burn down.” A useful one says “there is no extinguisher on this floor.” The AI Overview gives you anxiety dressed as an answer. The blue links give you a place to go. As a researcher you want the place to go.

There’s a quicker, scrappier version doing the rounds: end your query with -ai. Be precise about what it is, though — it’s just the minus operator excluding the word “ai,” not an official switch, and it suppresses the Overview only on some queries. Handy in a pinch; &udm=14 is the one that works every time.

Fix 2: the operators matter more now than they ever did

Read more

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