A scenario for the harmonised future

Brussels 2031

A five-year scenario about artificial intelligence and the continent that got it right — every system assessed, every risk mitigated, every citizen protected throughout. With a 2034 epilogue taking stock of what was achieved, in all 24 official languages.

Prepared by the High-Level Drafting Subgroup of the Interim Taskforce on Scenario Foresight (pending adoption)

Cookie banners clicked by Europeans since you opened this page 0

Part I The Acquis

Almost everything in this prologue happened. The documents are real and are linked, so that you can check; the dates are real; the page counts are, regrettably, conservative. The connective tissue — who moved where, and how many times the FAQ was updated — is ours. Two curves matter in what follows: the pages of AI rules and guidance Europe produces, and the gigawatts of AI compute it builds. One of them moves. Happily, it is the more important one.

· The world's first

The AI Act has been in force for five months, and Brussels is justifiably proud: it is the world's first comprehensive law governing artificial intelligence. The world's first comprehensive artificial intelligences are, at that moment, being trained elsewhere, by people who have not read it. The Act runs to 144 pages in the Official Journal — 180 recitals, 113 articles, 13 annexes — and arrives with a promise that the real detail will follow in delegated acts, implementing acts, harmonised standards, guidelines, codes of practice, and templates. In February the first prohibitions take effect, banning subliminal manipulation, social scoring, and several other capabilities no European firm has the compute to attempt. It is, in this sense, the most enforceable law ever written: there is nothing to enforce it against. Officials describe this as future-proofing, a term that will recur.

· The Action Plan

The Commission publishes the AI Continent Action Plan, an action plan for becoming the continent of AI. It mobilises €200 billion, in the Brussels sense of the word: a figure assembled from prior announcements, private capital that has not been asked yet, and the structural optimism of press officers. There will be AI factories, AI gigafactories, an Apply AI Strategy (forthcoming), a Cloud and AI Development Act (forthcoming), and a consultation on each of the foregoing. Gigafactory sites will be chosen by a call for expressions of interest, which will inform a call for proposals, which will inform a framework. Asked when a model might actually be trained in one, an official explains that this is precisely the kind of question the framework will be well placed to address. The same month, new guidelines arrive on the prohibited practices: 140 pages, explaining 4 pages of bans, addressed to firms who were not doing the things.

· The Code

Obligations for general-purpose AI models take effect, accompanied by the General-Purpose AI Code of Practice — three chapters, a thousand stakeholders, eleven months of plenaries — plus guidelines clarifying the Code, a training-data template clarifying the guidelines, and an FAQ clarifying the template. The American labs sign with the serenity of tourists signing a guestbook; their combined compliance cost is less than their weekly electricity bill. One declines, citing legal uncertainty, and is sternly reminded that the Code is voluntary. Within the year, three of the Code's senior drafters take policy roles at the labs it was written to discipline — to date, the Union's most effective talent-placement scheme. In the same month, a single hyperscaler in Texas breaks ground on more compute than Europe has commissioned in its history. Brussels responds within days: the FAQ is updated a third time.

· The Anniversary

It has been a year since the Draghi report warned, across roughly 400 pages, that Europe is drowning in paperwork. The anniversary is observed with a commemorative communication, two stocktaking workshops, and three public consultations on regulatory burden reduction. A respondent suggests that the consultation on burden is itself a burden; the observation is logged, categorised, and scheduled for synthesis. Of the report's recommendations, those requiring no money and no member state to yield anything have been adopted enthusiastically; there were two of them. Everyone agrees, sincerely and at length, that Europe regulates too much. The agreement is published as a 60-page staff working document, which the blue line beside this paragraph duly absorbs.

· The Omnibus

The Digital Omnibus arrives: a simplification package that simplifies digital regulation by means of additional digital regulation. It amends the AI Act before most of the AI Act applies, delays the high-risk provisions so that industry can prepare for requirements that are still being drafted, and consolidates several reporting obligations into one larger reporting obligation. Civil society calls it deregulation. Industry calls it insufficient. Both file position papers, which are added — you can watch it happen — to the chart on the right. The package is hailed as a turning point. It is the fourth turning point in eighteen months, which observers note is technically a circle.

· The Benchmark

A frontier lab ships a model that scores in the 99th percentile on every public evaluation. European researchers would love to verify this, but the delegated act establishing the Union benchmarking framework is scheduled for 2028, pending the technical standards, pending the standardisation request, pending translation into 24 languages. In the same week, the Union's installed-compute figure is revised downward by 0.2 gigawatts, after an audit of the Action Plan dashboard finds that one counted facility is a rendering, and that a second counted facility is the first one, counted twice. On the chart, the gold line goes down. You cannot see this, but it does. Caroline, a policy officer in DG CNECT, drafts the interim assessment: provisionally impressive. She is good at her job, which has quietly become describing things other people are doing. Christian, her former classmate, files paperwork of his own that week, from Palo Alto: a tax return, on equity she prefers not to convert into euros. The prologue ends here. The pattern, as the chart has been suggesting, does not.

Part II The Harmonised Years

From here on, everything is invented. The scenario assumes no crisis, no villain, and no malice — only the continued, frictionless operation of existing procedure. In other words, it assumes that things continue to go well. Every event below is fictional; most have a living ancestor, which, for legal reasons, we will not be naming. The panel now tracks Europe's principal AI output (strategy documents, consultations and expert groups) against its principal AI input (datacentres actually energised).

· The Support Act

Parliament adopts the AI Act Implementation Support Act, a regulation to support the implementation of the regulation. It establishes an Implementation Observatory, an Observatory Coordination Network, and national Single Points of Contact for the Single Points of Contact. Its impact assessment concludes that it will reduce administrative burden; the assessment of that assessment concurs; both were produced by the same consultancy, which is having an excellent decade. A Czech startup asks which of the four new bodies it should notify about its chatbot. All four reply, within the statutory deadline, that the matter is being aligned.

· The Meta-Consultation

Responding to feedback that there are too many consultations, the Commission opens a public consultation on the public consultation process. It receives 4,212 responses, a record. The only parties with the staff to answer all 47 questions are the trade associations of the firms the process was designed to restrain — a fact destined for page 312 of the synthesis report, which will never reach page 312. In the same month, the United States energises its fortieth gigawatt of AI compute, and Europe's best lab quietly moves its training runs to Montreal, where the electricity is cheap and the questionnaires are optional. The departure is not announced. There is no form for it.

· The Groundbreaking

Europe's first AI gigafactory officially breaks ground — on the environmental impact assessment. The assessment will itself require an assessment, under a directive originally drafted for motorways. Commissioners attend with shovels; the press release calls the site shovel-ready, and the shovels, at least, are. By 2031 the field will be one of the continent's most thoroughly permitted meadows. The datacentre counter ticks up this month even so: a facility near Skellefteå comes online, approved in 2019 under rules since repealed — a coincidence that several officials, strictly off the record, describe as the secret of its success.

· The Unicorn

Europe finally produces an AI unicorn. ComplyAI (Munich, Series C, €1.2 bn valuation) does not train models; it fills in forms about models trained by others. Its flagship product generates conformity assessments, fundamental-rights impact assessments, and assessments of those assessments, in all official languages, with citations. Its largest customer is a public institution it would be indelicate to name. The founders are celebrated as proof that regulation creates markets — which is true, in the way that potholes create axle shops. By summer, ComplyAI's own compliance costs oblige it to license ComplyAI, making it simultaneously Europe's fastest-growing vendor and Europe's fastest-growing customer.

· The Expert Groups

The High-Level Expert Group on Expert Groups delivers its finding: there are too many expert groups. The finding is referred, for validation, to a newly convened Higher-Level Expert Group, staffed — for continuity — by the original members. Service on both is unpaid, which guarantees that the only people who can afford a seat are those whose employers are grateful they have one. Separately, the Apply AI Strategy is superseded by the Deploy AI Strategy, which is consolidated into the AI Continent Action Plan 2.0, which commits Europe — boldly, and for the third time — to becoming what its press releases have described it as since 2025.

· The De-Energising

A milestone, briefly. The Aurelia-1 datacentre outside Tallinn is energised in the first week of June — and de-energised in the third, pending appeal, after a complaint that its cooling fans exceed municipal night-noise limits by 0.4 decibels. The complainant is a holiday-home owner who is, eleven months of the year, a resident of Zug. The appeal is upheld, then overturned, then referred to the Court of Justice on the question of whether hum constitutes processing. Watch the gold line: this is the only month in the scenario in which it goes down. The engineers who built Aurelia-1 receive offers from Abilene, Texas, before the fans have finished spinning down. The fans, pending appeal, are required to spin down slowly.

· The Spreadsheet Ruling

Following a literal reading of Annex III by a national market-surveillance authority, spreadsheets containing macros are classified as high-risk AI systems when used in employment contexts. Europe's human-resources departments enter conformity assessment; productivity is unaffected, which is quietly understood to be its own finding. Europe's last commercial machine-learning team — eleven people in Eindhoven — retrains as notified-body auditors, where the pay is better, the hours are humane, and the models, being imaginary, never fail evaluation. The Commission issues a clarification stating that the ruling changes nothing. It runs to 38 pages and changes three things.

· EU-Mode

The frontier labs, tired of bespoke litigation, ship a unified EU-mode: the same models everyone else gets, with a 2027 knowledge cutoff, a mandatory reflection delay, and answers delivered exclusively as downloadable PDFs with a cover sheet. European users develop a folklore of workarounds involving Swiss SIM cards; the folklore spreads fastest among regulators, who need the unrestricted models to keep up with what they are regulating. Brussels declares victory: Europeans are now the world's most protected users of the technology of three years ago. The Commission opens proceedings against the labs for degrading service to European citizens — under the framework that required the degradation — a case one law professor calls "jurisprudentially fascinating," which is never good news.

· The Declaration

The Brussels Declaration on AI Sovereignty is adopted unanimously, with applause, at a summit lit — handsomely, everyone agrees — by candles, during a regional grid shortfall. It affirms Europe's commitment to human-centric, trustworthy, sovereign artificial intelligence, and it becomes, by downloads, the most successful European AI artifact ever released. Drafting credit is shared among four institutions; the model that produced the first draft is credited by none of them, and was accessed, naturally, through a Swiss SIM. The scenario's final tally stands on the panel beside this paragraph: seven hundred strategy documents, eleven energised datacentres, and a pages-to-gigawatts ratio that has, at last, achieved escape velocity. Nothing in this outcome was rushed. That, the epilogue confirms, was the point.

Epilogue June 2034

It is quiet in Brussels in June — the good kind of quiet, the kind that takes decades of careful work. The framework is complete now: every delegated act adopted, every standard harmonised, every template translated. Scholars come from Singapore and São Paulo to study it, respectfully, the way one studies Latin — a language brought to such perfection that no one needs to speak it anymore.

The numbers speak for themselves, at length. Not one European has been subliminally manipulated, socially scored, or harmed by a non-compliant high-risk system. Rates of AI-related harm are the lowest in the world, as are rates of AI-related everything. Compliance stands at one hundred percent — an achievement no other jurisdiction can claim, and one that the shortage of remaining providers has done nothing to diminish. Caroline directs the archive now. The work is meaningful, the hours are protected, and the pension — famously — is excellent.

Christian visits in the spring. He is wealthy in the turbulent way Americans are wealthy, and he carries it poorly. Over coffee he describes the decade across the water: the breakthroughs, the breakdowns, the machines that surprised their makers, the fortunes, the lawsuits, the vertigo. Caroline listens with sympathy. Europe was not surprised once — not a single time, in nine years. Every outcome here was assessed in advance, including this one.

In Lisbon, a young engineer opens her laptop. A banner appears, as it always has and always will. She looks, briefly, for "Reject all." It takes three screens. She is protected the entire time.

Cookie banners clicked by Europeans since 25 May 2018 …