Back from holiday. What changed in AI over summer 2026 (and what didn't)

The phone spent three weeks in a drawer. I got back to my desk and realised I knew nothing about what had happened in AI since June. That was the good sign. It meant the holiday worked.

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.