Deep Dive Nº 06 · Regulation

The EU wrote the rulebook for artificial intelligence

Europe's AI Act is arriving in stages, and in 2026 the EU quietly pushed its heaviest rules back to 2027. Britain is not bound by any of it, yet the tools on your laptop are. Here is what the law actually says, and where it reaches you.

For a decade, the rule on artificial intelligence was that there was no rule. Products shipped, models scaled, and the law watched from the touchline. That era is closing. The EU Artificial Intelligence Act, the world's first comprehensive law on AI, entered into force on 1 August 2024 and is arriving in phases, each one switching on a fresh set of obligations. In 2026 those phases were reshuffled: a reform package called the Digital Omnibus pushed the heaviest rules further out, to 2027 and beyond. You are a British student, so the EU does not govern you. Read on anyway, because the companies that make your favourite AI tools very much are governed by it, and they tend to build one product for everyone.

The Act's core idea is refreshingly simple to grasp. Rather than ask what a system is, it asks how risky its use is, and scales the rules to match. A spam filter and a system that decides who gets a student loan are both "AI", but only one of them can quietly wreck a life, so only one of them carries a heavy compliance burden. Everything flows from that single sorting decision.

The law does not regulate AI. It regulates what you do with it, and the higher the stakes, the heavier the rules.

How the law sorts the machinesFour tiers, one question

Every system in scope is dropped into one of four risk tiers. The tier decides the obligations, from an outright ban at the top to no new rules at all at the bottom. It is worth holding this table in your head, because almost every headline about the Act is really an argument about which tier something belongs in.

The four risk tiers · EU AI Act
TierExamplesWhat the rules require
UnacceptableSocial scoring, untargeted face-scraping, emotion detection at work or schoolBanned outright across the EU
HighBiometric ID, medical devices, some hiring and exam or admissions toolsCertified checks, human oversight, logging, documentation
LimitedChatbots, deepfakes, AI-generated images and textMust be disclosed and labelled as AI
MinimalSpam filters, recommendation feeds, game AINo new legal obligations

The vast majority of AI you touch daily sits in that bottom tier and is left alone. The Act spends its energy on the two tiers above it, where a wrong output is not an annoyance but a harm.

The lines nobody may crossWhat is banned outright

The prohibitions are the sharpest part of the law, and they have been in force since 2 February 2025. These are uses Europe has decided are simply incompatible with fundamental rights, and no amount of paperwork makes them legal. Among the practices now forbidden:

That last ban is the newest and, for students, one of the most relevant: the "nudifier" apps that fabricate explicit images of a classmate from an ordinary photo are now squarely prohibited at the model level. The line on emotion recognition in schools is the other one to notice. A university cannot lawfully deploy a system that watches your face in a seminar or an online exam to judge whether you are engaged, bored, or stressed. That is now off the table inside the EU.

When disclosure becomes the ruleThe right to know it is a machine

This is the tier that touches your rights most directly, and its transparency duties are currently scheduled to apply from 2 August 2026. The principle is simple: if you are dealing with a machine, or with something a machine made, you are entitled to know. In practice that means three things.

First, chatbots must identify themselves. A system you converse with has to make clear you are talking to AI, not a person, unless it is obvious. Second, AI-generated content must be marked, ideally in a machine-readable way, so platforms and browsers can flag it downstream. Third, and most pointedly, deepfakes must be labelled. Synthetic image, audio, or video that depicts real people or events has to be disclosed as artificially generated.

Why this matters to you

The deepfake rule is the closest the Act comes to a personal shield. Image-based abuse and fabricated "revenge" clips hit students and young people hardest. A labelling duty does not stop a bad actor, but it hands platforms and courts a clear legal hook, and it makes "I did not know it was fake" a much weaker excuse.

The models under the modelsRules for the engines themselves

ChatGPT, Claude, Gemini and their peers are not built for one job, so the Act gives them their own category: general-purpose AI, or GPAI. Obligations on the companies that make these foundation models have applied since 2 August 2025. Every GPAI provider has to keep technical documentation, publish a summary of the data used to train the model, and put in place a policy to respect EU copyright law.

The largest models carry more. A model trained with more than 1025 floating-point operations is presumed to pose "systemic risk", which triggers extra duties: adversarial testing, incident reporting, and cybersecurity safeguards. That compute threshold is the law's rough proxy for "frontier", and it is why the biggest labs are watched more closely than a small open-source project.

The rules arrive slower than writtenThe Digital Omnibus reshuffles the timeline

The single most important development of 2026 is not a rule switching on. It is a rule switching off, or at least sliding into the future. For much of last year, 2 August 2026 was pencilled in as the day the high-risk regime would bite: the certified checks and oversight duties on systems such as hiring, admissions, and medical tools. That is no longer the case. A reform package called the Digital Omnibus on AI, proposed in November 2025 and given final sign-off by the Council on 29 June 2026, is now law, and it pushes those deadlines back.

Under the finalised Omnibus, the main high-risk obligations for stand-alone systems move from August 2026 to 2 December 2027, and for AI built into already-regulated products to 2 August 2028. Brussels framed it as simplification and breathing room; critics called it a rollback under industry pressure. Either way, the headline effect is clear: the Act is arriving more slowly than it was first written.

The useful way to hold all of this is as a phased timeline, not a single switch: the bans came first, the rules for the model-makers followed, and the heaviest obligations come last, with 2026 nudging that last step further out.

The phased timeline · as amended in 2026
DateWhat applies
1 Aug 2024The Act enters into force
2 Feb 2025Prohibited practices become enforceable
2 Aug 2025General-purpose AI model obligations apply
2 Aug 2026Transparency duties scheduled to apply
2 Dec 2027High-risk rules (stand-alone systems), deferred by the Omnibus
2 Aug 2028High-risk rules for AI inside regulated products

None of this softens the enforcement teeth. Once a duty is in force, breaching it is expensive, and the fines are pegged to a company's global turnover so that the numbers stay frightening even for the largest firms.

7%
Of global annual turnover, or €35m, whichever is higher: the ceiling for banned practices
3%
Of global turnover, or €15m, the cap for GPAI and most other breaches
2 Dec
2027: the deferred date the high-risk rules are now due to bite
10²⁵
FLOPs, the compute threshold above which a model counts as systemic risk

So if you see a confident claim that "the whole Act lands in August 2026", treat it as out of date. The prohibitions and the model-maker rules are already live. The rest is real, but it is now a slower, staggered rollout stretching into 2028.

Not in the EU, still in scopeWhy a British student should care

The UK has taken a pointedly different road. There is no British AI Act, and as of mid-2026 no equivalent Bill before Parliament. The government's stance is "pro-innovation": rather than one horizontal law, it asks existing regulators such as the ICO, Ofcom, and the FCA to apply five broad principles (safety, transparency, fairness, accountability, and contestability) within their own patches. Lighter touch, by design.

Two things follow for you. The first is the Brussels effect: it is expensive to build a labelled, disclosed, documented product for Europe and a separate unlabelled one for everyone else, so the strict standard tends to leak outwards and become the global default. When your AI tools start announcing themselves and tagging their outputs, thank the EU. The second is that the Act is now the reference text the whole world argues with, including a future UK government. Understanding it is understanding where the rules for your working life are heading, whichever side of the Channel writes them.

The takeawayThe mental model to keep

Strip away the acronyms and the Act is one clean idea executed at scale: sort every use of AI by how much damage it can do, ban the worst, licence the dangerous, label the deceptive, and leave the harmless alone. The banned list and the labelling rules are live now. The heavy obligations on high-stakes systems are coming, a little later than first planned. And even outside the EU, the tools you use every day are quietly being reshaped to fit.

✦ End of issue ✦

Get the next one in your inbox

One AI tool, one prompt, three quick wins, every week. Free, and under 3 minutes to read.

Subscribe Free →