In short

Worked: three questions instead of a list of rules. People still remember them a week later; they never remember policy clauses. Printing the page and putting it by the coffee machine worked too. Fewer questions came in, and the ones that did were specific.

Didn't work: explaining any of this through GDPR articles. Start with the legislation and the room disconnects within a minute. Banning things without offering an alternative did not work either. Block personal accounts without providing a company one and people simply stop talking about it out loud.

Let me start with myself. A couple of years ago I pasted a contract into a chat window because I needed to read it fast. Not a client's, my own, but it had another person's name and numbers in it. The decision took three seconds and I only thought about it a week later. Nothing bad happened. But I now know exactly what it looks like from the inside: nobody sits there planning to leak data, they are just trying to finish something.

That same question now comes to me more often than any other. The wording is almost always the same, and it almost always starts with "this is probably a stupid question". It is not stupid. And the answer people usually get is a bad one: "don't paste anything confidential." It sounds sensible and it helps nobody.

Date

15/08/2026

What was checked

Whether conversations are used for training by default: ChatGPT, Claude, Microsoft Copilot and Gemini, on personal and on work accounts. Claude and Gemini retention periods. The change to individual GitHub Copilot plans from April 2026. The course of the preservation order in the case brought by The New York Times. No prices: not one pricing page could be fetched cleanly.

Why "nothing confidential" does not work

A rule like that hands the decision to a person who is in a hurry. A contract looks confidential, fine. But what about a colleague's email with one paragraph about a client problem? A CV you need to compare against five others? A spreadsheet row with surnames in it? Every time, you decide again, you decide in two seconds, and you usually decide the same way: this is probably fine.

So I give people three questions instead of a prohibition. Questions are something you can remember. Policy clauses are not.

Three questions before you paste

1. Is there a person in this text who can be identified? Name, personal code, phone, address, photo, health data, salary, performance review. Indirect identification counts too: "our only accountant in the Panevėžys office" is a specific human being. This is where GDPR applies, and what it says is not "don't use AI." It says personal data can only go somewhere there is a legal basis and a contract for. When you paste a CV into your personal account, you are sending another person's data to a company your employer has no agreement with.
2. Did I sign something saying I would not show this? Contracts, NDAs, unpublished results, budgets, pricing, negotiation terms, pending decisions about people. GDPR is beside the point here. What you signed is the point.
3. Does this unlock something else? Passwords, access tokens, connection strings, internal links. A snippet of code looks harmless right up until there is a key in it. Since April 2026 individual GitHub Copilot plans train on your data by default, with an opt-out, while business and enterprise seats stayed excluded. That is a different product from the Microsoft 365 Copilot assistant on a work account, and the two are very easy to confuse.

If the answer to any of the three is yes, you cannot paste that text as it stands. That does not mean you cannot use AI. It means you need a step in between. And it is worth being precise about where the actual problem sits: not with AI, but with data leaving for a company nobody signed anything with.

The second item on that list is the one people find least clear, so here is what it looks like in practice. This is how I ask when I have to answer an awkward client email, without putting the email itself anywhere.

Help me write a reply.
Situation: [the client is unhappy that the work is two weeks late].
What I want: [own the delay, offer a new date, do not apologise three times].
Tone: [businesslike, no decoration]. Up to [100] words.
Give me three different versions. I will fill in the facts myself.

The facts, dates and names go in afterwards, in my own inbox. It takes half a minute and it settles the whole first question.

What a company account actually changes

Most people assume the difference is price and speed. The real difference is somewhere else entirely, and it explains everything above.

On personal accounts, conversations are used to improve the models by default. That is how ChatGPT works on the free and personal paid plans, how personal Claude accounts work, how Microsoft Copilot works for a signed-in consumer, and how Gemini works with activity saving switched on. On business and enterprise accounts it is the other way around: ChatGPT Business, Enterprise and Edu, Claude for Work, Microsoft 365 Copilot on a work account and Google Workspace are not used for model training by default. On top of that you get admin settings, retention controls and, most importantly, a data processing agreement. That agreement is exactly what is missing when an employee works from a personal account.

Retention differs more than people expect. With Claude, if you allow your conversations to be used for training they are kept for five years, and thirty days if you opt out. On a personal Gemini account, activity history is kept for eighteen months by default, human-reviewed conversations for up to three years, and even with the setting off, chats are held for another 72 hours.

Two caveats I do not enjoy giving, but they matter. First: turning training off is not the same as not sending the data. The text still travels to the provider's servers and sits there for a while, so you reduce the risk without removing it. Second: "deleted" does not always mean deleted. In 2025 a US court in the case brought by The New York Times ordered OpenAI to preserve conversation logs, including ones users had deleted. The broad order was lifted in the autumn of that year and narrower retention continued. The case will pass, the principle will not: a court can freeze deletion, and your settings will not change that.

These settings change often. This page was written in August 2026. Do not recite it from memory a year from now. Open the settings and look.

A policy nobody reads is not a policy

Now the uncomfortable part, and here my opinion is not subtle. A fourteen-page AI policy in a shared drive, which everyone approved by clicking "I have read this", protects nobody. It protects the person who wrote it, and only until something actually happens.

Three things work: one page instead of fourteen, one person you can message and ask without consequences, and a permitted route that is more convenient than the forbidden one. The third is the important one. If logging into the company account takes longer than opening personal ChatGPT, people will use personal ChatGPT. Every time. Not out of bad faith, but because the work has to be done by five.

And one more thing leaders usually do not enjoy hearing. The people pasting contracts into a personal account are generally the fastest people on the team. They are not a risk. They are the signal that a tool is missing, and it is better to read that signal early.

What to do on Monday: print the page, put it somewhere visible, and tell the team the name of one person they can message without consequences. One page and one name. That is the whole policy you need for the first week.

The fair objection: three questions are too crude. A lawyer will tell you straight away that there are at least five more distinctions between "there is a person in the text" and "you may send it", and they will be right. But I would rather have a rule someone remembers while in a hurry than one that is more correct on paper. A more precise policy stays unread, and then it protects nobody.

Bans do not stop people who are trying to get the work done. Only a more convenient route does.

If that was a lot for one sitting, the same rule also exists as a story: The contract that left the building, a comic in ten panels.

p.s. I did delete that contract from my chat history later. After reading how deletion actually works, I understood it was more for my own peace of mind than for security.

The honest answer is boring: probably not, but it depends on what you actually do all day. The difference between these four tools today is smaller than the difference between a person who opens one of them daily and a person who opens one once a month. Real differences do exist, just not the ones in the advertising. The one that matters most is not how clever the model is. It is location: whether the tool already sits where you work. In your email. In your files. In your meeting.

Date

15/08/2026

What was checked

Where each tool actually lives (Microsoft 365, Google Workspace, the editor, the browser) and what happens to conversations by default on personal versus work accounts. Not checked and therefore not stated: prices, model versions, free-tier limits as numbers.

What is deliberately not in this table

I do not list model versions. They change every few weeks, and if I put them in, this page would be wrong before Christmas. Tools are named by product. I do not list prices either, because they change more often than this page and vary enough between countries and plans that any number would mislead. Check the price on the vendor's own page on the day you decide.

I am not judging by benchmark scores but by where the tool lives in normal work. That moves more slowly and it is more useful.

This is a September 2026 snapshot. It will go out of date and I will update it.

A simple pattern falls out of that table. If your company lives in the Microsoft stack, Copilot wins, not because it is smarter but because it is already inside Outlook and Teams. If you live in the Google stack, Gemini wins for the same reason. ChatGPT and Claude win when the work is not tied to a file or a meeting: writing, thinking, long texts, building your own small tools.

One clarification on research. All four now have a deeper research mode that returns a report with links. And all four still occasionally attach a claim to the wrong source. I got caught twice, so now I open the links even when I am in a hurry.

If you click none of the links, it is not research. It is a nicely formatted guess.

What the free tiers actually give you

More than a year ago, and more than you expect. For ordinary office work, meaning emails, summaries, tidying up text and searching, the free tiers are enough today. All four let you upload a document, search the web and get an answer with links. A couple of years ago you had to pay for that. Now it is the baseline.

It stops in three places. The limit arrives at the worst possible moment, not while you are playing with it but on Friday when something has to go out. Long documents go worse: a hundred-page contract on a free tier often gets handled superficially, or refused outright. And the newest agent features are almost always kept for paid plans.

But the thing that matters most barely gets discussed. On free and personal accounts, conversations are used to improve the models by default, and on work accounts they are not. If you handle real data, that matters more than any feature in the table.

When paying is worth it

It is worth it if you hit the limits more than once a week, if you work with long documents regularly, or if you want to hand over multi-step work rather than get an answer to one question.

It is not worth it if you open an AI tool three times a week for short texts. In that case the free tier plus whatever your company gave you covers everything, and a second subscription is a quiet tax on the feeling that you might be missing out. I paid that tax myself and I am not embarrassed about it, only annoyed about the time.

One more thing, and it tends to disappoint: the best result almost never comes from switching tools. It comes from explaining what you want more precisely and adding an example. When I compared the four, I gave all of them the same real task of mine and the same request.

Rewrite this text for [audience: a client who is in a hurry].
Up to [120] words, businesslike, no decoration.
Here is an example of how my writing sounds: [I paste one of my own paragraphs].
Here is the text to rewrite: [text].
At the end, tell me what you cut and why.

There were differences between the four answers, but they were smaller than the difference between this request and what I used to type: "rewrite this, but nicer".

And if your company gave you nothing

This is still common, especially in smaller organisations. Work with a free tier, but hold one rule: no real client, colleague or contract data. A free account is a personal account, and on a personal account your conversations go into model training by default. For drafts, structure and wording that is entirely enough.

What to do on Monday: for one week, write down how long one specific recurring task takes with the tool and without it. Two numbers on one page, nothing more.

And ask for a tool, just not the way people usually ask. "Everyone else has one" never works. One sentence with a number does: how long a specific job takes each week, and how long it takes when you do it with the tool. A week of notes is enough. To a manager that stops being a trend and becomes a number they can hold against a subscription price. I say that from the side of the table where those requests get approved or refused.

If the answer is still no, ask what the decision is missing: money, a policy, or clarity on who owns it. Usually it is the second or the third. At least then you know what actually needs fixing.

What will go out of date first

The verdicts on agent features will age fastest. That area moves monthly, and what looks slow and unreliable in September may be normal by December. Free-tier limits will age second, because they get changed often and quietly.

The slowest thing to change is where each tool lives. Copilot will still be in the Microsoft stack, Gemini in the Google stack. That is exactly why I built the table this way instead of around benchmark scores.

The fair objection: this table is organised around where a tool lives, not around which model is better. If model quality genuinely matters to your work, my angle will look convenient and shallow. I agree that such work exists: long code, hard analysis, legal text. But in ordinary office work the quality gap is drowned out by whether the tool opens inside the app you are already working in.

The best tool is the one you open every day. Everything else is haggling over a few percent.

p.s. I did cancel the second subscription in the end. The first month without it passed without me noticing anything, which was the clearest answer I could have got.

The bad sign came an hour later. I opened the news and within ten minutes found three articles describing the same regulation differently, one of them still writing in the future tense about something that has been in force for a while. So I stopped reading news and opened an empty file instead. I wrote down what changed over June, July and August for someone doing an ordinary office job in Lithuania. Not model launches. Not stage demos. It took an hour, it came to seven items, and next to each one I wrote what to do about it on Monday. Without that part an item is just news.

I am writing this in mid August. Where I am not certain, I say so.

Date

15/08/2026

What was checked

The date the transparency rules started to apply and the grace period, the publication and entry into force of the digital omnibus, the change in the wording of Article 4, the date national authorities got their powers, RRT's role, where Lithuania's own law stands in parliament, the Eurostat figure and two population surveys. Prices and model versions are deliberately absent: they would age faster than the page.

Six things to do on Monday. Tick off the ones you have done.

1. Since 2 August, AI-generated content has to be labelled

On 2 August the transparency rules in the EU AI Act started to apply. Most people missed it, because August.

They are simple enough. If someone is talking to a chatbot, it has to be obvious to them that it is a bot and not a member of staff. If you publish AI-generated text, images or audio, it has to be marked. The strictest treatment goes to video and audio where a real person appears to do or say something they never did.

Systems that were already on the market got a four month grace period, until 2 December.

What to do on Monday: find out two things. Whether your website has a chatbot, and whether it says so. And who on your team owns the labelling of AI content you publish externally. The answer to the second one is almost always "nobody", and knowing that is half the work.

2. The big deadline moved, and nobody made much noise about it

For a long time it looked like 2 August would be the turning point. That was the date the rules for high-risk AI systems were meant to start applying: recruitment, education, credit scoring and so on. It did not happen.

On 24 July the so-called digital omnibus, Regulation (EU) 2026/1744, was published in the Official Journal. It came into force on 27 July, a few days before the date it was amending. The requirements for high-risk systems listed in Annex III of the Act moved to 2 December 2027. For those embedded in regulated products, to 2 August 2028.

This is law in force now, not a proposal. In the spring most coverage wrote about it in the future tense, and that is exactly the confusion I walked back into.

What to do on Monday: nothing urgent. But if someone spent the summer selling you a compliance tool on the argument that "it becomes mandatory in August", ask what exactly they meant. The deadline moved. The need to get organised did not.

3. The AI literacy duty stayed, only the wording softened

Article 4 of the AI Act has applied since 2 February 2025. In short: an employer has to make sure that staff using AI know enough about what they are using.

The same omnibus softened this. Before, you had to ensure literacy. Now it is enough to support and develop it. That sounds like a small change and legally it is not: an obligation to produce a result became an obligation to make an effort.

At the same time, national supervisory authorities got real enforcement powers on 3 August. In Lithuania that is RRT, the Communications Regulatory Authority, which has been the market surveillance authority and single point of contact since 1 April 2025.

Plainly: nobody is going to show up to inspect whether your team has been trained. But if questions ever come up about another part of the Act, an untrained team becomes an aggravating factor.

What to do on Monday: one file, one hour. Who on the team uses which AI tool, and for what. Without that list, any training is guesswork.

4. Lithuania's own AI law has not passed yet

This one gets misread more than anything else, so it gets its own point.

Lithuania designated its authority on time and was not among the countries that ran late. But the separate national law that sets out the full implementation framework only reached the Seimas in June and is still in committee. As far as I know in mid August it has not passed, and the debate should continue in the autumn session.

What to do on Monday: do not wait for it. The transparency rules apply directly, with or without a national law.

5. The tools started doing things, not just talking

This is the biggest practical change of the summer, and no regulation had anything to do with it.

A year ago an AI tool mostly answered. Now it opens a browser, collects the data, fills in the table and hands back a result. Through the first half of 2026 every major vendor pushed in exactly this direction.

Now the honest part, because I tried it and took it on the chin. In July I handed an agent a dull spreadsheet job and walked away for half an hour. The request went roughly like this.

Here is [export.csv], about [600] rows.
Do three things:
1. Normalise the date format to [yyyy-mm-dd].
2. Number the rows by date.
3. Move the amounts into a separate column [Amount].
Return the tidied file and a separate list of exactly what you changed.

I came back, looked at it, and everything seemed fine. It was only days later that I noticed two columns had swapped their values, and there was no way to see it without reading row by row. The list I had asked for did not save me either: it described what the agent believed it had done. For short work it is usually faster to do it yourself, and longer work still needs supervision. But the direction is clear, and it changes the question itself. We used to ask what to type into the box. Now we ask which job we can hand over.

What to do on Monday: pick one boring, repetitive task that eats more than an hour a week. One. Try handing it over, and write down how much time you saved or lost. The second number is also a result. I wrote mine down.

6. The numbers surprised me

I assumed Lithuanian companies were moving slowly. I was wrong.

That is nearly 12.5 percentage points more in a year, the third fastest rise in the whole EU behind Denmark and Finland, against an EU average of about 20%. The data is Eurostat's, and the Ministry of the Economy and Innovation repeated it this summer. There is nothing newer yet; the next measurement is published at the end of the year.

On the personal side the picture is different. In July 2025 a Swedbank-commissioned Spinter survey of 1,009 people found that 49% had tried an AI tool at least once. In the autumn, an LRT-commissioned Baltijos tyrimai survey of 1,019 people found 12% using one daily.

Two different surveys, so you cannot compare them directly. But the gap between "tried it" and "use it daily" shows up anyway, and everything I think about this fits inside that gap. Everyone has the tools now. Almost nobody has the habit, and no new version is going to build that for us.

What to do on Monday: check yourself. Over the last five working days, how many times did you open an AI tool for work rather than out of curiosity.

7. What did not change

The models got better over the summer. Prices moved in places. Nobody I know lost their job over it.

And the thing that matters most did not change either. Most employees still do not know what they must not paste into that chat box. Not because they would not understand, but because nobody told them clearly. That one gets its own post, because it does not fit in a bullet.

The fair objection: seven items in an hour is not a survey, it is one person's field of view. If you work in a factory or a hospital, your summer looked nothing like this, and the first four items will be too general for you. Fair enough. But seven of your own items are worth more than a hundred of someone else's, and anyone can redo this list for themselves in the same hour.

The rules changed over a single summer. Habits do not move that fast, and habits are what decide whether any of this pays off.

The same story without the text: The Monday after the holiday, a comic in ten panels.

p.s. I still keep that spreadsheet with the swapped columns in a separate folder. Good reminder of how confidently a machine gets things wrong.

Today the same report takes four minutes with AI. And that is where the real story starts. When work gets close to free, everyone suddenly sees that nobody needed it. What bothered me was not the machine. It was those two weeks per quarter.

What the x-ray is

AI does not break the bone. It shows the fracture that was there yesterday, and the year before that.

The Danish anthropologist Dennis Nørmark and the philosopher Anders Fogh Jensen called this pseudowork. It is activity that looks like work, fills the calendar, produces files, and changes nothing in the outcome. It used to be easy to hide, because it cost time, and time looked like proof.

Three layers show up in the picture. First: documents nobody reads, the reports and summaries and roll-ups that travel from one folder to the next. Second: meetings that exist so somebody feels informed. A machine now writes their content in a minute, and then you can see how much of it there ever was. Third, and the least comfortable: intermediate steps. Rewrites, reformatting, moving things from one template into another. They exist not because they were needed, but because the alternative was hard for somebody.

The test is simple. Take any recurring piece of work and ask what would happen if this once it simply did not get done. If the answer is "somebody would call", the work is real. If the answer is "nobody would notice", you have your diagnosis.

Why the picture is unpleasant to look at

An x-ray does not heal. It only shows, and it shows without any tact.

Pseudowork is usually not one person's fault. It is what is left of an old system, and sometimes it is somebody's entire job description, handed to them on the day they were hired. So the answer is almost never "let that person go". The answer is to hand them a real problem, and real problems are not in short supply.

There is one more easy trap. AI is excellent at producing pseudowork faster. If you used to write one unnecessary document a week, you can now write ten, and they will look better. The x-ray helps only if somebody actually looks at the picture.

Where I use this framework

It opens my talk at HR Week Lithuania in September. Not the tools, but the question of what the tools light up.

When people ask me where to start with AI, I do not look for a new tool. I ask them to open last week's calendar and mark everything that nobody would have noticed if it simply had not been done. The list is usually an uncomfortable length. Mine was longer than I expected.

In my own work I keep one rule: if AI does some task very well and very fast, that is not a win, it is a question about why that task was in my week at all.

One thing I did not understand right away, and it cost me one ruined conversation. You cannot show somebody the picture before you have something for them to do next. Otherwise it is not a diagnosis, it is an accusation, and the conversation ends right there. It took exactly one time for me to learn that, and I still regret it.

AI did not steal that job. It just showed there was no job there.

Fourteen experiments are not a strategy. That is a hobby, only expensed.

I see the same thing inside companies. Marketing has its own tool, customer support is testing another, and in finance one person has built something useful and told nobody. Ask what changed and the room goes quiet.

Two boxes that look alike

Almost everything we do with AI fits into two boxes. They behave nothing alike, which is why it is worth separating them before buying anything.

Automation fits where a task repeats and has a correct answer. Booking invoices, sorting documents, a first translation draft, moving data from one form into another. The goal is to take the human out of the chain, and you measure it in time and in errors that come back.

Augmentation fits where somebody has to decide. What price to offer. Who to hire. Which project to stop. Here AI decides nothing, it only widens the picture: more options, a counterargument, reading you would never get through yourself. The goal is a better decision, not a faster one.

The difference is easy to spot. If two smart people can look at the same output and disagree, it is a decision. If there is nothing to disagree about, it is a repetitive task.

The expensive mistakes come from mixing the boxes. Automate a decision and you get confident nonsense that nobody checks any more. Augment a repetitive task and you get an expensive toy that offers three options every morning where one button was needed. I made the second mistake myself and paid a subscription for it for three months.

Turning experiments into an operating model

This takes three things, not ten.

1. Two lists. What repeats and what gets decided. Written by the people who do the work, not by the leadership team.
2. An owner for every item. Not a team, a person with a name. No owner, no item.
3. No more than five numbers for the whole company. Two on automation: hours freed per month and errors coming back. Two on decisions: days to decide and how many decisions get reversed later. One on usage: how many people use it every week without being told to.

Five numbers look like too few. But nobody reads twenty numbers, and everybody remembers five.

A warning about that last number: the moment usage becomes mandatory it stops telling the truth. People open the tool so it gets logged, then close it.

How I use this in my own week

When a conversation about a new tool starts, my first question is which box it goes in. If the answer is "both", we do not yet understand what we are buying, and the conversation is better postponed by a week.

I do the same with myself. Repetitive things get handed over and I do not go back to them. Where a decision is needed, I use AI as an opponent: give me three reasons my plan will not work. Once in a while one of those three is good enough that the plan has to change. I still make the call myself, because responsibility cannot be handed to a machine, however tempting that would be.

Automation frees up time. Augmentation decides whether that time was worth anything.

I have nothing to give them, and I say so out loud. Not out of modesty, but because the question is the wrong one. Safety today does not come from the field, it comes from how fast you can move to the next one. It took me a few changed fields to work that out, and none of those moves looked like a plan at the time.

Two honest ends of the same argument

On one side there are ten thousand hours. Malcolm Gladwell made that number famous in "Outliers", drawing on Anders Ericsson's research into deliberate practice. Ericsson himself later said the popular version flattened his work too much, but the idea stuck: depth comes from many hours in one narrow field.

On the other side there is David Epstein's "Range". His argument is simple. In clear, rule-bound domains like chess or golf, early narrow specialisation works beautifully. Where the rules are vague and keep moving, the people who win are the ones who tried many things and can carry a solution from one field into another. Working with AI is the second kind of environment, because the rules change every few months.

And here is the unpleasant part. AI is strongest exactly where a competence is narrow, well defined and repeatable. That is the job description of the classic narrow specialist. Not because that person is weak, but because their work is the clearest target on the board. I am not saying specialists stop being needed. I am saying specialisation on its own no longer makes anyone safe. Depth used to be both a tool and a shield. It is still a tool.

What a ranger is

Ranger is a word I made up for my talks: someone who moves across wide terrain instead of digging one deep well. What that person can do comes down to four things.

Pick the problem: not wait for a task, but see where it hurts and say what is worth solving first. Connect fields: bring a solution from logistics into HR, from games into training. Direct the machine: say what the result has to be, check what comes back, and say no when the answer is beautiful and wrong. And relearn without drama, because a new tool is not a threat, just another thing in the backpack.

Now the correction, without which all of that sounds far too pretty.

A ranger with no depth at all is just a person with opinions.

You need at least one field where you have worked long enough to feel it instantly when something is off. That feeling is the only real protection against a machine's well-phrased mistakes. Without it, AI looks right every time, and it is not.

Where I use this framework

When I talk to people about careers, I suggest measuring not years in one position but how many different problems they carried from start to finish. The number is often a lot smaller than the CV suggests.

For my own learning I keep a rule: one field where I dig deep, and several where I stay a beginner without any embarrassment. That is where most ideas come from. And I do not tell my own kids they need to learn to code. I tell them to learn to notice when something is broken, and to not be shy about saying it out loud. My daughter does this daily already, mostly about me.

A narrow specialist waits for the task. A ranger arrives with the problem.

Then three more hours went by. The app stopped working after my first change. The model kept confidently proposing a fix that fixed nothing, I said no and asked again, and again. The fourth time I finally read the error message myself, and it was solved in two minutes. That is the real experience: one hour of magic and three hours of stubbornness. I am not a professional engineer, but this is exactly how I build my own tools.

What it is, without the polish

Vibe coding in plain words: you describe what you want, AI writes the code, you run it, see what is broken and keep talking. The loop runs until it works.

Three places where it is genuinely good. Prototypes: in one evening you find out whether an idea is worth a real developer's time, and most ideas are not, so learning that cheaply is worth a lot. Internal tools: a small app for one person or one team that nobody would ever buy. And learning: for a junior it is the best way to understand logic before syntax. The machine writes the syntax now. The logic still has to be in your head.

Three lines I do not cross

1. No sensitive data. Personal data, health information, customer lists, financial records. None of it goes into a tool built this way. Ever.
2. No keys or passwords in code I have not read myself. The model will happily write a key straight into a file and then suggest pushing it to a public repository. This is the one place where a mistake costs real money the same day.
3. No regulated or core business logic. Payroll, customer invoicing, anything an auditor checks. That needs a person who answers for it by name and can explain every line.

Everything inside that chalk circle is a fair playground. Everything outside it stops being a vibe coding question and becomes a question about who carries the responsibility.

The part nobody says on stage

The first eighty percent arrives in twenty minutes and feels like magic. The last twenty percent is repeating yourself, reading the output, and refusing to accept the model's answer. That is where the weekend goes, and that is what never makes it into a demo, because it is boring.

The people it does not work for almost always quit at the point where they needed to ask a third time.

And you still have to read. Not every line, but enough to see where the data is kept and what the app talks to over the network. If you do not know that, you are not controlling the limits, because the limits live exactly there.

Where I use this framework

All my Mac tools are built this way. They run on my machine, the data never leaves it, and nobody else uses them. Before asking for budget for a real system, I spend an evening making an ugly prototype, because the conversation that follows is nothing like a conversation about a slide. And when a young person says they want to learn to code, I tell them to start with something they actually need. Not a course. A problem.

p.s. that first Saturday app is still running. I use it maybe once a week, and I do not regret those four hours at all.

What actually changed over the summer, with the dates and the sources: What changed in AI over the summer of 2026.

Why those three questions are the ones, what separates a personal account from a company one, and what happens to deleted chats: What not to paste into ChatGPT.

The whole thing started earlier, and not very nicely on my part. When the first widely known version of ChatGPT appeared, I used it to write a piece about diamonds and investing in them, sent it to Viktorija and asked her to edit it. I did not say who wrote it. She spent a good half day on that text, and when she was done she asked who on earth had written it. I told her. She was furious. But she is a good person, she forgave me quickly, and soon we were sitting down and working out how to put the same question not to one person but to a few hundred.

Tool

Twelve separate ChatGPT assistants, one per writer, plus one plain free ChatGPT with no preparation at all

Date run

Results published 02/12/2024

Task

25 texts on the theme "Why do I write?", up to 200 words each. Readers had to mark each one as written by a person or by AI

Time

Two weeks on a single instruction, seven to eight A4 pages. About 20 seconds per generated text

What was measured

How many of the 25 texts each voter marked correctly, and what share of voters assigned each text to a human. 233 people voted

How it was done

1. I spent two weeks writing a single instruction. It came to seven or eight A4 pages.
2. From it I built twelve AI assistants, one per writer taking part.
3. The assistants were fed only publicly available texts by those people: five samples each that they sent us, plus whatever we found ourselves on social media, in magazines and in newspapers.
4. Twelve professionals (writers, journalists, communications, advertising and PR specialists) each wrote up to 200 words on the theme "Why do I write?". The assistants never saw those texts, they were kept separately.
5. Each assistant generated its own version in about 20 seconds. No human touched them afterwards, even the mistakes were left in.
6. One more text came from plain free ChatGPT with no preparation at all. That made 25 texts in total: 13 from machines and 12 from people.

Then we loaded them into a separate platform and sent them out to be judged. 233 people voted, from experienced editors and publishing house staff to people with no connection to writing at all: IT specialists, engineers, hairdressers, builders, insurance brokers. One task. Read it and mark who wrote it, a person or AI.

The result

The text that the largest number of people marked as human-written came from a machine. It was the assistant copying the style of a writer who covers personal finance.

76.8 %The share of readers who decided that the most human of all 25 texts was written by a person. It was written by an AI assistant.

Second and third place went to actual humans: one at 74.2 %, the other at 72.8 %. In total, five AI texts were assigned to a human by more than half of the readers.

But first place ended up surprising me the least. The average surprised me more. Out of 25 texts, people got 13.66 right on average, an accuracy of 54.64 %. The worst score was 5 out of 25. The best was 22, and exactly one person managed that. Nobody marked all of them correctly. The number looks harmless until you remember that 13 of the texts came from a machine, so marking everything as AI without reading anything would already have scored 13.

And one more small thing, still my favourite. Some of the voters admitted they had ignored our request and quietly checked the texts with AI detection tools. Those tools called all 25 texts human-written. All of them.

What the experiment does not prove

The honest part, because back then I talked about it too little.

This was not a scientific study. 25 texts, one theme, up to 200 words each, one language, one December in 2024. The people who came to vote were the ones who wanted to, and someone like that is already interested in the subject, so the sample is not representative. The numbers I quote here are from the main round with 233 participants. A few days later, by the time I talked about the experiment on television, more than 700 people had filled in the open form, so you will find that figure quoted in places too. It is not a contradiction, just two different moments.

Most importantly, the experiment did not prove that machines write better. It only showed that in a short text people can no longer tell who wrote it. That is a completely different thing.

And one more detail that came across too quietly at the time. The result was not free. Two weeks of work and eight pages of instruction sit behind it. The text written by plain ChatGPT with no preparation was the one people recognised best of all: 77.7 % correctly marked it as machine-written.

What I think now

Almost two years have passed, and those eight pages of instruction are no longer needed. What took me two weeks in 2024 you can now do in an evening, and the result will be better. Today's version fits in a few lines.

You are going to write in my style. Here are [five] of my texts: [pasted].
First, write down what you notice: sentence length, which words I avoid,
how I open and how I close.
Then write a [200] word text on [topic].
Do not correct my habits and do not decorate anything.
If something in the style is unclear, ask instead of guessing.

Which means my conclusion at the time, that we still need a person who can train the machine well, aged faster than I expected.

"Can AI write like a human" was an interesting question in 2024. It has been answered since, and it stopped being interesting.

A different question is still interesting: what do I need that for. I do not want a machine that writes instead of me. Not out of principle, but for a very practical reason: when the machine writes the first draft, I no longer have anything to say, I only have things to fix. The difference is enormous. Work like that produces text that is tidy and empty.

This site is my answer. I write the first draft myself, because that is where all the content is. I bring AI in where the work is boring: reordering the structure, checking facts and dates, telling me which sentence does not work, translating into English. There is nothing romantic about it, it is just the split of labour that gives me the best result.

What I would not use this for

One thing I would not do, then or now. I would not build an assistant that copies a living person's style without telling them. In the experiment all twelve took part voluntarily and sent us their own texts, and that is the only reason any of it was decent. Without that consent, what is left is a convenient way to impersonate someone.

And I would not use it where the text is a signature: an opinion, an apology, a message of sympathy. Not on principle, but because a text like that has to be your own or it is worth nothing.

What to do on Monday: take one of your own texts from this week and underline where your thinking is and where it is only form. The form can go to the machine. The thinking is worth keeping.

One question from 2024 is still sitting there unanswered. Professionals will be fine, I am relaxed about that. But what about the people who are only learning to write and can see that a machine does the same thing better and in twenty seconds. Where do they find the motivation to go through the unpleasant stage when their own writing is still bad. I had no answer then and I have none now, and it does not only apply to writers.

The fair objection: 233 volunteers, one theme and up to 200 words is far too thin a base for any conclusion about writing. That is true, which is why I never call this a study. But one thing survives even a sample like that: not one of the 233 marked all 25 correctly, and the best score was reached by exactly one person.

p.s. Viktorija could not sleep for two nights after the results came in. I slept perfectly well that week, and I now think that says something bad about me rather than about her.