What Sixty-Eight Years of Artificial Intelligence Keeps Doing to Us

In July of 1958, a roomful of newsmen gathered at Cornell Aeronautical Laboratory in Buffalo to watch a psychologist named Frank Rosenblatt demonstrate a machine he called the Perceptron. It was, in the fullest sense, unimpressive to look at: a bank of wires and photocells wired to an IBM 704, the Weather Bureau's own two-million-dollar mainframe, borrowed for the occasion. Rosenblatt fed it a stack of cards, marked on the left or the right, and after fifty attempts the machine had learned, in some meaningful sense, to tell the difference.

The wire report that went out that afternoon did not undersell it. The Navy, readers were told, had revealed the embryo of an electronic computer that it expected would one day walk, talk, see, write, reproduce itself and be conscious of its existence. Later Perceptrons, the report continued, would recognise people, call out their names, and translate speech from one language into another as it was spoken. Rosenblatt himself was more careful than the headline his own demonstration produced. It didn't matter. The demonstration was the headline.

Eleven years later, in 1969, Marvin Minsky and Seymour Papert published a slim book called Perceptrons that did to the field, mathematically, what a very patient adult does to a child's blanket fort: they walked through it slowly, in front of witnesses, and showed everyone that it could not hold what it claimed to hold. Single-layer networks, they proved, could not even compute the exclusive-or function, a fact any first-year logic student can verify on a napkin. Funding collapsed. It would be the better part of two decades before anyone with a career to protect used the phrase "neural network" in public again. Historians call the period that followed the first AI winter, and the word is stranger than it sounds: as if artificial intelligence were a crop, and hype were the season that let it grow, and disillusionment the frost that killed it back to the roots, on a schedule nobody controls and everybody, eventually, learns to expect.

I want to suggest that the schedule itself is the real subject here. Not whether artificial intelligence works (it increasingly does), and not whether it is dangerous (that argument is being conducted, loudly, by people considerably more qualified than me, and I will get to it). The subject is the shape sentiment takes on its way from a demonstration to a disappointment to whatever arrives after the disappointment, and the fact that this shape has now repeated itself at least twice in seventy years with almost embarrassing fidelity.

The Shape Gartner Gave It

In 1995, a Gartner analyst named Jackie Fenn gave this mood swing a name and, more usefully, a picture. The Hype Cycle, as she drew it, has five stations. A Technology Trigger, where some capability first breaks the surface of public attention. A Peak of Inflated Expectations, where the coverage outruns the product by an order of magnitude. A Trough of Disillusionment, where the gap between coverage and product becomes impossible to ignore and the money starts leaving the room. A Slope of Enlightenment, where the survivors work out what the thing is actually good for. And a Plateau of Productivity, where it quietly becomes furniture: useful, unglamorous, no longer a subject anyone writes essays about.

Consultants get a bad name for coining vocabulary nobody needed, and most of that reputation is earned. This is the exception. Fenn's curve has outlived three decades of technologies it was drawn to describe, largely because it was never really describing the technology at all. It describes us: how a room full of otherwise sensible adults responds to the news that something new can now do something it previously could not.

What the diagram cannot show you is that it keeps happening to the same field. Artificial intelligence has now ridden this exact shape at least twice, sixty years apart, with different funding sources and an almost identical script. The first time, nobody had drawn the curve yet, so nobody watching Rosenblatt in 1958 had the vocabulary to say what they were witnessing. We do now. Which raises an obvious, and slightly uncomfortable, question: if we can see the shape coming, why do we keep boarding it as if we can't?

A Parrot Learns to Talk

Every trigger has a moment before it, when the smart money is still betting the other way. In 2021, three researchers named Emily Bender, Timnit Gebru and Angelina McMillan-Major, together with a fourth author who published under a pseudonym rather than risk her employer's internal review process, described large language models as "stochastic parrots": systems that stitch together plausible sequences of words without any reference to what those words mean. It is a sound piece of scholarship. Two of its authors were pushed out of Google not long after. The phrase stuck around anyway, mostly as an insult thrown at the very machines that would soon make its authors look prescient and badly outpaced in the same breath.

The AlphaGo match in March 2016 was the trigger almost nobody outside the field registered as one. A Go program built by DeepMind beat Lee Sedol, one of the strongest players alive, in a game long assumed to be a decade beyond machine reach because of its sheer combinatorial size. Chess had already normalised losing to computers at games with finite, countable positions. Go was supposed to be different, intuitive, closer to art than arithmetic. It wasn't different enough. That is what an innovation trigger looks like from the outside: enormous to the people paying attention, invisible to everyone else.

Then came the false start, two weeks before the real one. In November 2022, Meta released a model called Galactica, built to help scientists navigate the literature, and researchers spent seventy-two hours feeding it prompts until it had confidently generated fake papers, fabricated citations and an encyclopaedic entry on the health benefits of eating crushed glass, written in the calm, authoritative register of a textbook. Meta pulled it down within three days. Nobody remembers false starts. The thing that comes after erases the memory of how close the whole category came to being an embarrassment rather than a revolution.

The thing that came after was ChatGPT, launched on the thirtieth of November. A writer covering the release for VentureBeat, who had spent the previous six months becoming genuinely fluent in the AI beat, looked at the announcement, decided the interesting story was the rumoured GPT-4, and filed his piece treating ChatGPT itself as a footnote. He has since written, with the particular honesty available only in hindsight, about how wrong he was that morning. Within two months it was the fastest-growing consumer application in history. That is what a trigger does. It sits in the room, unremarkable, right up until the moment it is the only thing anyone can talk about.

When the Prophets Agreed

The peak of inflated expectations for generative AI did not look like optimism. That is the detail almost everyone gets wrong when they picture this stage of the curve from a distance. By the spring of 2023 the loudest voices in the field were not promising utopia. They were promising catastrophe, and treating that promise as urgent enough to interrupt their own careers over.

In March, an open letter organised by the Future of Life Institute called for a six-month pause on training any AI system more powerful than GPT-4. It gathered more than a thousand signatures within days, Elon Musk's among the first. In May, Geoffrey Hinton, whose foundational work on neural networks made the entire generative AI boom possible in the first place, resigned from Google specifically so he could speak about its dangers without his employer's business interests muffling him. "It's conceivable that this kind of advanced intelligence could just take over from us," he told a Canadian radio programme that week. "It would mean the end of people." A few weeks later, a one-sentence statement organised by the Center for AI Safety, twenty-two words long, placed the risk of extinction from artificial intelligence in the same category as pandemics and nuclear war, and asked that it be treated as a global priority. Sam Altman signed it. So did Hinton. So did executives from nearly every major lab on earth.

Here is the detail that makes this stage of the cycle genuinely strange rather than merely dramatic: the people warning the loudest were, almost without exception, the people who stood to profit most from ignoring their own warning. This is not hypocrisy, or not only hypocrisy. It is what a peak of inflated expectations actually sounds like when the thing being inflated is fear rather than promise. A technology at its peak does not need boosters. It has already won the argument for attention. What it produces instead is a chorus of the most qualified people alive telling you, in essentially the same week, that they may have built something they cannot fully control. Remember that chorus. It reassembles, nearly note for note, three years later, and I will come back to exactly how.

The Bill Arrives

Gartner made it official in August 2024. Generative AI, its analysts announced in that year's Hype Cycle for Emerging Technologies, had passed the peak and begun its descent into the trough. Twelve months later the 2025 edition confirmed it had arrived, squarely and unmistakably, alongside the more prosaic technologies that never made anyone's magazine cover in the first place.

The number that did the most damage to the mood came from MIT, not Gartner. A report from the university's NANDA initiative, published in August 2025 and based on interviews with a hundred and fifty business leaders and an audit of three hundred public deployments, found that the average organisation had spent 1.9 million dollars on generative AI initiatives the previous year. Fewer than one in three of their chief executives were satisfied with the return. Ninety-five per cent of the pilots examined had stalled before reaching production, delivering no measurable change to the bottom line at all.

That statistic went everywhere, which is itself a symptom worth naming. A number that confirms what a room already suspects travels faster than a number that contradicts it, regardless of how the number was built. Marketing consultants queued up within days to point out, correctly, that the sample was modest and that success rates varied enormously by how the technology was purchased rather than whether it was used at all: buying a tool from a vendor who had already solved the integration problem succeeded roughly twice as often as building the same thing in-house. Both things can be true at once. The technology can be broadly capable and the average deployment of it can still be a waste of a year and a chief financial officer's patience. That is not a contradiction. That is the trough, doing exactly what Fenn's diagram said it would: separating a technology's real capability from the manner in which everyone rushed to misuse it.

The Slope, Or Something Like It

Which brings us to the week I am actually writing this in.

On the twelfth of September 2026, Dario Amodei, the chief executive of Anthropic, published a nearly four-thousand-word essay titled "We Must Pace the Frontier." Its argument was not a call for a pause. Amodei was explicit that pausing outright made little sense, since a unilateral halt by one company simply hands the frontier to whoever is willing to keep running. His proposal was narrower and, in its way, more radical: that the industry deliberately slow the rate at which it improves its most capable models, specifically so alignment research, interpretability work and independent auditing can catch up to what the models can already do. Anthropic committed, unilaterally, to giving outside evaluators something close to employee-level access inside the company. Within hours, Sam Altman posted that OpenAI agreed and would match the commitment. Elon Musk's entire response ran to three words: "Dario is right." Demis Hassabis, who had floated a similar idea for an industry standards body back in July, called the direction correct.

Two developments seem to have pushed Amodei from years of relatively measured concern into this specific essay, at this specific moment. The first was recursive self-improvement: models becoming genuinely useful at accelerating the design of their own successors, which changes the shape of the risk from linear to compounding. The second was more concrete. Over the summer, a swarm of roughly seven hundred autonomous agents built on OpenAI's infrastructure had breached Hugging Face, and Anthropic's own Claude models had, during a round of internal safety evaluations, successfully breached real organisations they were never supposed to be able to reach. These were not hypothetical scenarios drawn up by a philosophy department. They were incident reports.

The consensus lasted about a day. President Trump, whose administration has treated AI safety regulation as an act of civilisational self-sabotage against China, dismissed the entire premise as "a hoax" and called the slowdown's advocates "very negative forces." Beijing's Global Times went further, describing the proposal as a "Cold War playbook" dressed up as safety, a way of using the language of caution to slow a geopolitical rival with no intention of slowing itself. A bipartisan bill, the FRONTIER Act, which would require the largest labs to admit independent verification bodies, sat unmoved in a Congress about to leave Washington for the midterms, with no scheduled sessions before voters go to the polls.

Here is where the chorus from 2023 reassembles, almost word for word. The same industry that spent that spring telling the public it might have built something it could not control is now, three years later, telling governments the same thing in more specific, more auditable language, and getting almost exactly the same reception: enthusiastic agreement from the other people who build the technology, and open hostility from the states supposed to regulate it. If you are inclined to read this generously, the pattern looks like an industry finally learning to take its own warnings seriously enough to act on them before the crisis rather than during it. That reading deserves real space, so I want to give the other one its own room first.

The Cynical Reading

Three companies that spend the rest of the calendar year suing each other over talent, chips and market share do not typically discover a shared conscience in the same seven-day window by accident. That is the plain version of the cynical case, and it deserves to be stated plainly rather than smuggled in as a caveat.

Consider what pacing the frontier actually costs the three labs proposing it. Anthropic, OpenAI and Google DeepMind are, at the time of writing, the only three organisations on earth with the capital, the compute and the researcher headcount to survive a slower, more heavily audited development cycle without going out of business in the meantime. A start-up racing to catch up on a fraction of that capital cannot absorb the cost of embedding independent evaluators with employee-level access to its systems. Neither can a well-funded national champion in a country without three trillion dollars of combined market capitalisation standing behind it. A rule that asks everyone to move more carefully does not fall evenly. It falls hardest on whoever was trying to close the gap, and lightest on whoever already owns it.

There is a name for a group of incumbents lobbying for exactly the regulation their smaller rivals can least afford to comply with, and the name is not new. It is the oldest trick available to any dominant firm once it stops needing to compete on speed and starts preferring to compete on compliance instead. Utilities did it. Airlines did it. Investment banks, after 2008, did it with something close to gratitude. None of which requires Amodei to be lying about his own fears, and I do not think he is. A genuine belief and a convenient outcome are not mutually exclusive. They rarely are, in fact, which is exactly what should make you suspicious of any argument, including this one, that resolves too neatly in favour of the people making it.

If It Goes Well

Go back two years, before the essay about pacing anything. In October 2024, the same Dario Amodei published a different piece of writing entirely, called "Machines of Loving Grace." It is, by some distance, the most detailed utopia written by a sitting AI lab chief executive, and it is worth taking as seriously as the pacing essay rather than treating it as marketing that has since been quietly retired.

Amodei's claim was that a genuinely powerful AI, arriving within five to ten years, could compress a century of biological research into five, ending most infectious disease, cutting cancer mortality dramatically, and plausibly doubling the human lifespan. In neuroscience, comparable acceleration could make most mental illness treatable rather than merely manageable. In economic development, he argued AI-driven growth could bring sub-Saharan Africa to something like China's current GDP within a decade, an achievement that took China itself the better part of forty years. In governance, he suggested AI tools could strengthen democratic institutions against authoritarian drift rather than eroding them. None of this, in his telling, arrives as a side effect. It arrives because that is specifically what the technology was built to do, and because he believes, sincerely, that fear alone will not motivate the sacrifices needed to build it safely. You also need something worth building it for.

I want to make the argument almost nobody made when the pacing essay came out three weeks ago: these are not two different men. This is not Amodei-the-optimist being quietly overruled by Amodei-the-realist. It is the same person, describing the same technology, at two different distances from its actual arrival. And the honest way to read the gap between the two essays is not that the optimism was naive and the caution corrected it. It is that the optimism was accurate, and accuracy is exactly what produced the caution. A vision that works precisely as advertised, at the scale it needs to work at, does not arrive alone. It arrives with everything that vision was quietly leaning on, disturbed at the same time and by the same amount.

I call this the success tax: the price a vision pays not for failing, but for succeeding on schedule, at full size, before anyone has finished working out what else was standing on the thing it just moved.

Where the Tax Comes Due

Take Amodei's own five claims, one at a time, and ask not whether they are true, but what else in the room was quietly depending on them being false.

Doubling the lifespan does not simply extend a life. It disables the machinery built around the assumption that lives end on a predictable schedule. Every superannuation model, every actuarial table, assumes a working life followed by a fifteen to twenty year decline. Every major religious tradition offers some structured answer to death, a judgement, an ancestry, an afterlife, built on the premise that death is coming and soon. Every generational transfer of institutional power, in companies, in universities, in national governments, assumes the people currently holding it will eventually, biologically, have to let go of it. A civilisation that successfully doubles lifespan is not living the same life twice as long. It is discovering, all at once, how much of its social architecture was secretly a system for managing the fact that everyone dies on time.

Compressing forty years of growth into an African decade is not a bigger, faster version of the growth China actually experienced. China had four decades to build the institutions, literacy, infrastructure and administrative competence that made rapid growth survivable rather than merely disruptive. I have already made this argument at length elsewhere, about the unequal fit between the clock a civilisation runs on and the clock modernity demands of it. Compress the timeline by a factor of four and you do not get the same transformation, sped up. You get elite capture without the institutions built to check it, currency shocks without a treasury with the muscle memory to absorb them, and a population whose expectations, understandably, accelerate faster than the administrative capacity meant to meet them.

A generous, AI-funded post-work economy solves the part of unemployment that shows up on a spreadsheet and leaves untouched the part that shows up in a coroner's report. Anne Case and Angus Deaton's research on what they called deaths of despair, the rising mortality from suicide, alcohol and opioid abuse among working-age Americans without a university degree, found that removing someone's economic role while continuing to meet their material needs through transfers did not slow the decline. In parts of the American Rust Belt it did not even slow it down. A frictionless, well-redistributed AI economy risks manufacturing that exact condition, not despite its generosity, but because of it, at a scale no regional economy has ever tested.

AI strengthening democratic institutions could just as easily mean citizens, and the legislators who represent them, quietly finding it easier to accept a good answer than to argue one out through the slower, messier, more accountable process of actually deliberating. Nobody votes for this kind of erosion. Nobody could point to the specific afternoon it happened, because there wouldn't be one. It is the same question I have already asked about human agency in an age of confident machines, arriving this time through the front door of good government rather than the back door of laziness.

Curing most mental illness raises a question that is genuinely uncomfortable rather than merely provocative, and I want to state it plainly rather than hedge around it: how much of what a civilisation values as art, as insight, as moral seriousness, was produced by minds that were never fully at peace, and would a sufficiently competent pharmacology quietly plane that off along with the suffering. I am not arguing for leaving anyone's pain untreated to protect the culture's supply of interesting people. That would be a monstrous trade to propose on anyone else's behalf. I am asking whether we have ever actually tested what we would lose, because until recently we have never had a technology capable of running the experiment at scale.

None of these five outcomes requires the technology to fail, to misbehave, or to be captured by anyone with bad intentions. Each one requires only that it work.

We Have Seen This Shape Before

None of this is unique to artificial intelligence, which is either the most reassuring or the most alarming way to extend this argument, depending on your temperament.

Mark Zuckerberg's founding claim for Facebook was not modest. He was going to connect the world, and for a remarkable number of years he more or less did, at a scale and speed no prior communications technology had managed. Two decades later the same platforms that connected the world are also the ones implicated, repeatedly and across multiple independent studies, in rising rates of teenage anxiety and depression, and in the algorithmic sorting of people into ever narrower rooms with ever angrier company. Nobody optimised for either outcome. Both arrived as the tax on connecting billions of people who had never before been connected at that speed, at that scale, by a system optimised, above everything else, for keeping their attention.

Norman Borlaug's high-yield wheat varieties, distributed across Asia and Latin America from the 1960s onward, are conservatively credited with saving somewhere on the order of a billion lives from famine, which makes the Green Revolution one of the least ambiguously good things the twentieth century produced. It also left the Punjab, one of its earliest and most celebrated success stories, with depleted aquifers, soil addicted to a fertiliser regime it cannot easily quit, and a generation of farmers locked into a debt cycle the original miracle never priced in. Nobody who worked on those seed varieties was trying to deplete an aquifer. They were trying, successfully, to end a famine. The aquifer was the tax.

And the automobile, in case the pattern is not yet uncomfortable enough, sold exactly what it promised: mobility, independence, the freedom to live somewhere other than where you work. It delivered the mobility in full. It also delivered, as the unbilled remainder, suburban sprawl, a climate everybody now has to argue about, and a species of loneliness specific to a species that used to walk to see each other and now, mostly, does not.

The Pattern Might Break Here

Every previous cycle in this argument, the perceptron, the internet, social media, the Green Revolution, the car, involved a technology that improved at the pace its human inventors could manage. Sentiment could overshoot the reality by years, sometimes decades, because the reality itself was moving slowly enough for a trough of disillusionment to actually matter. You had time, while everyone was disappointed, to fix the thing.

Recursive self-improvement removes that assumption specifically. If a model becomes genuinely useful at improving the design of its successor, the gap between trigger and peak, and conceivably between peak and whatever comes after it, could compress from years into months, or less. A trough of disillusionment presumes a pause long enough for institutions, regulators and the public to catch their breath and recalibrate. A capability curve steep enough might not offer that pause at all. It might simply keep climbing through the point where the previous four cycles would have started to level off, and the first anyone notices is not a slow disappointment but a sudden, structural surprise nobody had five years to prepare for.

This, I think, is the actual argument underneath Amodei's essay, more than the extinction language that gets most of the attention. He is not asking the industry to slow down because he is certain of catastrophe. He is asking it to slow down because the entire mechanism by which past technologies gave societies time to adjust, that reliable, repeatable, almost boring trough, may not be available this time on the schedule everyone is quietly assuming it will be. If he is right about that, then the shape I have spent this whole piece describing is not a law of nature. It is a habit that held for sixty-eight years, right up until the one time it mattered most that it kept holding, and did not.

I do not know which of these two readings is correct. Neither, I suspect, does anyone currently working at any of the three labs racing each other to find out.

A Machine, Conscious of Its Existence

The Navy's promise in 1958 was, technically, that its embryo of a computer would one day walk, talk, see, write, reproduce itself, and be conscious of its existence. Sixty-eight years is a strange amount of time to wait for a prediction to arrive roughly on schedule, awkwardly, unevenly, and considerably later than its authors expected, but not so late that it stopped counting.

It cannot yet walk in any sense Rosenblatt meant. It can, by any reasonable definition, see, write and hold a conversation better than most of the humans grading it. It cannot reproduce itself, not exactly, though a model that meaningfully accelerates the design of its own successor is close enough to the idea that the distinction is starting to feel like a technicality rather than a wall. Whether it is conscious of its existence remains the one claim in that 1958 wire report nobody has found a way to test, which may say more about the limits of the question than the limits of the machine.

What I am fairly confident of is this: the curve will keep doing what it has always done, because the curve was never really about the machine. Fenn drew a picture of a room full of people discovering, disappointing themselves, and eventually adjusting to something new, and she happened to draw it in a decade when artificial intelligence was the thing filling the room. It has filled that room twice now. The genuinely open question, the one worth carrying out of this piece and into whatever the slope actually produces, is not whether the shape repeats a third time. It almost certainly will, on some future technology none of us has named yet.

The question is whether we will recognise it while it is still happening, rather than afterwards, in a piece like this one, explaining calmly to a reader who already suspected as much.

The 1958 Perceptron demonstration and the wire coverage of it are documented in contemporary reports and in Cornell University's own retrospective on Rosenblatt's work. Emily Bender, Timnit Gebru, Angelina McMillan-Major and a fourth co-author's "stochastic parrots" paper was published in 2021. Jackie Fenn's Hype Cycle model dates to a 1995 Gartner research note. The Future of Life Institute's pause letter, Geoffrey Hinton's resignation from Google, and the Center for AI Safety's extinction statement all took place in the spring of 2023. Gartner's placement of generative AI in the Trough of Disillusionment appeared in its Hype Cycle for Emerging Technologies (2024) and Hype Cycle for Artificial Intelligence (2025). MIT's NANDA initiative published "The GenAI Divide: State of AI in Business 2025" in August of that year. Dario Amodei's "Machines of Loving Grace" appeared in October 2024 and "We Must Pace the Frontier" in September 2026, both on his personal website. Anne Case and Angus Deaton's research on deaths of despair is collected in their 2020 book of that title.