On AI Fatigue

Three commencement speakers were booed this May for saying “AI” out loud. Eric Schmidt at the University of Arizona. Gloria Caulfield at the University of Central Florida. Scott Borchetta at Middle Tennessee State. Schmidt drew the loudest reaction when he reached the line about AI and the future of work. He paused, said “I can hear you,” and tried to continue. The booing only came back louder.

In the European version, no one boos. Across the rooms I worked in this spring, in Warsaw and Helsinki and Tallinn, in policy sessions and youth dialogues and the hallway conversations between them, men and women retreat into their screens when AI comes up. Their eyes drift toward the plasterwork. They sigh, “AI, again”.

Faced with this, one almost wishes for hostility; hostility, at least, implies a recognition of stakes. What occupies these rooms instead is a heavy, terminal boredom.

Fatigue is the right word for it, and fatigue is something else entirely from opposition. Opposition would be a clarifying gift. Opposition is what Schmidt received in Tucson, and what every honest argument deserves; an audience that boos has at least done you the courtesy of listening.

Fatigue demands nothing. Doom demands belief. Criticism demands engagement. Fatigue asks only for the energy to look away, and barely that. It is the cheapest of political emotions and the most contagious. A room of doomers will at least argue. A room of critics can be persuaded. But the exhausted have already departed; only their bodies remain seated at the table.

The exhausted have never read Amodei’s “Machines of Loving Grace” or any other thought piece concerned with AI, for that matter.

Superintelligence, compute scaling laws, the question of who builds the datacenters; none of this troubles them. They could not tell you what GPT-5.5 does that GPT-4 did not. They have heard the acronym “AI” too many times this year and have arrived at the private verdict that the topic itself is spent. The question of what the machine can actually do has been quietly retired from their heads.

Some portion of this exhaustion is manufactured. Fast Company this month catalogued Facebook pages with names like “California Life” and “Rural Ohio Life” pumping AI-generated images of crops and coastlines stamped with anti-datacenter slogans; OpenAI last year took down four China-linked operations for using ChatGPT to seed coordinated commentary across TikTok, X, Reddit, and Facebook. The joke writes itself.

Hostility has a half-life. The boos in Tucson will be a clip for a week, an essay for a fortnight, an awkward memory by summer. Schmidt will stand on another stage in another state. The graduates who shouted him down will draft their first resumes, perhaps using AI, for companies built on the technology they rejected. Opposition consumes itself. It needs an audience, an enemy, a reason to keep its voice up; eventually it loses one of the three.

Fatigue runs the other way. It accumulates. It surfaces nowhere because there is nothing to surface; no story attaches itself to the absence of an action. When a delegate scrolls past an agenda item without raising her hand, nothing happens. Nothing ever happens.

Policy gets made this way, through the slow withdrawal of attention from one room and its slow accretion in another. The Maritime Safety committee fills because the phrase “Maritime Safety” contains no enemy. Unobjectionable topics absorb attention. The AI working group empties because the stakes feel abstract, the vocabulary feels foreign, and the audience has decided it has heard “enough”.

The American boos will fade. The European yawns will set the regulatory environment for the decade.

Most of the public discussion still frames AI as a moral question. Is AI safe? Is AI ethical? Is AI fair? Will AI take our jobs? Does AI steal from artists? Is AI biased? Will AI become conscious? These are genuine questions; several of them have genuine answers; all of them will be decided downstream of a different question entirely.

That question is whether your country has electricity. Whether your country has chips. Whether your country has the researchers and engineers to build models, and the policy environment that lets them build. Whether your country owns its compute or rents it. Whether your country exports its models or imports them. Whether your country sits upstream of the AI economy or downstream of it.

AI is, above all, an infrastructure question. It belongs in the same family as the railroad in 1850, the electrical grid in 1920, the highway system in 1955, the internet backbone in 1995. The countries that built those infrastructures shaped the century around them. The countries that argued instead about whether the railroad was good for the soul shipped iron ore to the ones that didn’t argue.

Europe has cast itself as the country arguing about the soul. The continent was not competitive in AI before the AI Act, and the Act will not make it competitive after. Europe arrived at this position through structural choices that the regulation does not touch: shallow capital markets, fragmented compute, a labor regime hostile to research velocity, an academic pipeline trained to leave. The Act, when it lands, will regulate models Europe was never going to build. The rules were drafted by people who have never trained one. It is the kind of cheap, late legislation that political systems produce when they want to be seen acting on a phenomenon they do not understand. Mistral, the one team genuinely trying to build a frontier lab on the continent, will continue working underneath it anyway.

Behind the fatigue, behind the moral framing, behind the eye-rolls in the committee rooms, sits an ideology rarely named aloud. Its name is degrowth.

Degrowth says: we have enough. The planet is finite, growth is destructive, energy consumption is suspect, and any new technology demanding more electricity and more water and more chips and more labs is marching in the wrong direction. Datacenters are bad because datacenters are large. AI is bad because AI consumes. The path to a virtuous future therefore runs through doing less, building less, computing less.

The first failure is empirical. American datacenters consume less water than American almond farms, by orders of magnitude. The electricity required to train a frontier model is a rounding error against the electricity required to keep Berlin warm through one cold winter. The grid problem is a building problem, and humanity has known how to solve building problems since the nineteenth century.

The second failure is arithmetic. Degrowth treats the world as zero-sum: a litre of water spent on a server is a litre denied to a field, a kilowatt for training is a kilowatt taken from a home. The framing collapses on inspection. We can keep our cities warm and train frontier models in the same winter, and we already do. The serious version of the critique is not about litres or kilowatts at all. It is about what we owe the next generation in a constrained world. Fair question. The answer is that we owe them the largest possible expansion of capability, because every problem they will inherit, from disease to climate to scarcity itself, is solved downstream of more intelligence. To choose less is to choose to hand them a smaller answer to a larger question.

The third failure is philosophical, and it is the worst of the three. Degrowth assumes that doing less will lead to more good. Growth, however, is the precondition of any livable future. Every problem degrowthers claim to care about, from climate to inequality to disease, gets solved downstream of more capability, more energy, more intelligence brought to bear on it. AI is the largest lever for capability humans have ever built. To refuse the lever because the lever consumes electricity is to confuse a poverty of ambition with a triumph of ecology. It is a fetish for stagnation.

A civilization that talks itself out of growth preserves nothing. It chooses to be smaller next year than it was this one. The cost shows up as cancer deaths that did not have to happen, as scientific discoveries that arrived two decades late, as young people who never lived to see what their lives could have become with the technology their ancestors refused to build.

Beyond degrowth lies the harder question: what is AI for?

The honest answer begins with admitting that humans were never built for what most of us do all day. Hunching over a spreadsheet. Moving a rectangle on a slide deck. Editing paragraph breaks in a document that will never be read. None of it is what humans were made to do. We were built to eat well, to sleep well, to make love, to walk the woods, to argue with friends and reconcile with them, to make and raise children, to look at things that move us and to make things that move other people. The catalogue of legitimate human activity is long and ancient, and the Monday-morning to-do list contains almost none of it.

A serious humanism would notice this and ask what could free us from the work we were never built for. The honest answer is capability. AI is, among other things, the first technology since the steam engine with the scope to deliver that liberation at scale. To call it a threat to human flourishing is to define human flourishing as the freedom to remain at the spreadsheet.

The smallest discipline is the refusal to nod. In every committee room across Europe this spring, the dismissive line has gone unchallenged. Someone calls AI overhyped, and the table nods. Someone calls the AI working group a distraction from real issues, and the table nods. Someone proposes redirecting attention toward energy security or climate or housing, and the table nods, never noticing that AI is the lever that helps solve all three.

The industrial picture is plainer still. Europe needs compute. Europe needs labs. Europe needs the electricity to run both. Mistral will not carry the continent. Its CEO Arthur Mensch told the French National Assembly this month that Europe has roughly two years to build sovereign AI infrastructure or accept permanent dependence; he called the alternative becoming a “vassal state” of American technology.

The German Mittelstand will need to integrate AI before its competitors do, and the policy environment will have to make that integration possible rather than punishable. Compute corridors. A nuclear restart. The lifting of the data and labor frictions that keep European AI labs an order of magnitude smaller than their American and Chinese peers. None of this is mysterious. The mystery is why the rooms responsible for it have decided the subject is beneath them.

The intellectual diet matters too. The loudest voices in the AI discourse have never shipped a product, trained a model, or sat with the technology long enough for it to change their mind about anything. Skip them. Read the engineers shipping products, the labs publishing benchmarks, the Karpathys and Suttons who put the field on the map. Opinions formed by use survive contact with reality. Opinions formed at second hand do not.

The future is not yet finished. Schmidt said it in Tucson on his way off the stage, with the boos still hanging in the air.

The future is not yet finished. The people finishing it are doing so at a speed and inside a culture that has no analog in any room I have sat in this spring. They are in San Francisco. They are in Beijing. They are running experiments while the committee is on lunch break. By the time the committee reconvenes, the experiment will have produced a result.

The yawners do not decide the future. They decide only how much of the future Europe gets to participate in.

Whatever the working groups, committees and policymakers of other sorts decide, the model will exist. It will be trained somewhere, hosted somewhere, owned by someone. The interesting question is whether any of those somewheres are in Europe. On current trajectory, none of them will be, and the continent will live downstream of decisions made elsewhere.

The next time the eye-roll arrives, wherever you happen to be, refuse to roll back. Consider what has already happened with tools any person can open in a browser. This spring, a 23-year-old with no advanced mathematics training cracked a 60-year-old Erdős problem with ChatGPT in eighty minutes. A man with no medical degree used the same ChatGPT to design an mRNA vaccine that shrank his dog’s cancer by 75 percent. Every day, ordinary people sit down with these models and produce work they could not have produced alone: software engineers ship code beyond their solo reach, writers find sentences they would not have found, designers and musicians and small-business owners build things that would have required, a year ago, a team they could not afford.

Now extrapolate. Five years. Ten.
The future is unimaginably large. It is also closer than we think.