The EU AI Act’s implementation timeline is not one future date, it is a staged calendar: prohibitions from 2 February 2025, GPAI obligations from 2 August 2025, the remainder including synthetic-content transparency from 2 August 2026 — and most of it is no longer ahead of me.
I had the EU AI Act filed mentally under things to think about eventually, the same undifferentiated future bucket I keep quantum computing might matter and I should learn a new language properly in. It took actually pulling up the implementation timeline, rather than continuing to treat the whole subject as one distant event, to notice that most of the dates in it are not distant at all. Several of them are already behind me.
The Act entered into force on 1 August 2024. From there it does not arrive all at once, it stages in on a fixed calendar. On 2 February 2025, the prohibitions and the AI-literacy duties took effect. On 2 August 2025, the obligations for general-purpose AI models, governance structures, and notified bodies took effect. Both of those dates are behind me as I write this, which means whatever obligations attach to them are not a future compliance project. They are the current legal position, and have been for a while.
The one after those has arrived as well, which is the point where eventually stopped being an honest word for how I had been thinking about any of this. On 2 August 2026, the remainder of the Act applies, including transparency obligations for providers of synthetic content — a category that anyone generating media with these tools should be reading carefully rather than assuming sits somewhere else. That date, at the point I am writing this, is only weeks behind me rather than a comfortable distance ahead, and I had been carrying it in my head as somewhere out on the horizon right up until I checked.
After that there is one more staging point: 2 August 2027, when the general-purpose-AI compliance deadline lands in full and the Annex III high-risk provisions become applicable. That is the one still genuinely ahead, with enough runway that treating it as not yet is still defensible, for now, in a way none of the earlier dates on this same timeline are. It is also the only one of the five dates that actually is still ahead of me, and I had been treating all of them, including the ones already well in the past, with exactly the same mental shrug, as though a schedule with five distinct dates on it was just one vague future thing rather than five separate, checkable facts.
Checking a compliance calendar like this one is not the same exercise as checking whether I actually meet each obligation, and I want to be honest about which of those two things I have actually done here. I know now which dates on the timeline are behind me and which are still ahead, precisely, rather than vaguely. Whether any particular system already satisfies what a given date requires is a separate question, one nobody resolves just by reading a timeline, and conflating checking the calendar with checking my compliance would be exactly the kind of shortcut that got me here in the first place.
The mistake was not ignorance of the Act’s existence, since everyone doing anything with AI has heard of it by now. The mistake was letting bare awareness that the Act exists stand in for actual knowledge of which of its obligations already apply to me today, which are very different states of knowledge, and only one of them is actually useful when a provision quietly crosses from future to current while I am not looking at the calendar. A staged rollout with named dates is not a single future event to file away. It is a schedule, and a schedule has to be checked against today’s actual date, not against a vague sense of how far off the topic used to feel.