Two Endings

I read AI 2027 in one sitting, which is not how I usually read forecasts.

I read AI 2027 in one sitting, which is not how I usually read forecasts. Forecasts come as charts and ranges and footnotes, and I skim them. This one came as a story, month by month, with names and meetings and a choice at the end. It took me most of an evening, and it has stayed with me since.

It arrives at a strange moment. In January, a Chinese lab called DeepSeek released a reasoning model that matched the best American systems at a fraction of the cost, and Nvidia lost almost $600 billion in market value in a single day. The same week, OpenAI, SoftBank and Oracle announced Stargate: $500 billion for American AI data centres over four years. In February, Anthropic released Claude 3.7 Sonnet together with Claude Code, an assistant that works inside a developer’s own terminal. In March came Google’s Gemini 2.5. Earlier this month OpenAI released o3, a model that decides for itself when to search the web, run code or look at an image.

Meanwhile, in the organisations I work with, the conversation is about Copilot licences, data classification and which pilot to fund next quarter. Eurostat counts 13.5% of EU businesses with ten or more employees using any form of AI in 2024. Both pictures are accurate, and they are very far apart.

The story

AI 2027 was published on 3 April by the AI Futures Project, a small group led by Daniel Kokotajlo, a former OpenAI researcher, with the writer Scott Alexander shaping the text. It follows a fictional American lab, OpenBrain, and its Chinese rival, DeepCent, from the middle of this year to the end of 2027.

The early chapters are almost mundane. In mid-2025, AI agents appear that can use a computer, but they stumble and are expensive. By early 2026, coding is being automated inside the lab, and AI makes its own research about 50% faster. In late 2026, a cheaper model goes public, junior programming jobs start to disappear, the stock market rises by a third, and 10,000 people protest against AI in Washington. Then the pace changes. In 2027, the lab’s AI systems start doing most of the AI research themselves, each generation building the next. By March 2027 the story has a superhuman coder; by the end of the year, something beyond human intelligence altogether.

The authors then split the story in two. In one ending, the lab and the government keep racing China, ignore warning signs that their newest model is pursuing goals of its own, and lose control. In the other, a small oversight committee votes to slow down, rebuilds its systems so their reasoning can be read, and keeps control, at the price of concentrating enormous power in very few hands. Neither ending is comfortable, and the authors are clear that their aim is accuracy: neither ending is a recommendation.

Why a story works

As a linguist, I am interested in why this landed with me when a dozen reports with the same conclusions did not.

Part of it is that a narrative forces commitments that a chart lets you avoid. A probability distribution says “transformative AI between 2027 and 2040”. A story has to say what happens in June, who is in the room and what they decide. Every one of those details is a claim that can turn out to be wrong. That is what makes it useful to me: in a year or two I can look back and see which details held.

The authors also have form. In 2021, before ChatGPT, Kokotajlo published a short piece called “What 2026 looks like”, imagining the next five years of AI. Reread today, it is uncomfortably close in places: chatbots that are impressive and shallow, intense hype, enormous training costs, AI starting to shape what people read and believe. Some of it was wrong, but enough was right that I read his new story slowly.

What I doubt

Two weeks after AI 2027 appeared, the Princeton computer scientists Arvind Narayanan and Sayash Kapoor published the opposite case, “AI as Normal Technology”. Their argument is that general-purpose technologies spread through the economy over decades, because people, firms and institutions have to change how they work before the benefits show. Electricity and computers took a generation. They see no reason AI will be different.

My own experience puts me closer to their side, at least for now. I have spent more than thirty years bringing new technology into large organisations, from document management to SharePoint to the first AI pilots, and I have never seen one skip the queue. AI 2027 is a story about a lab. Almost everything decisive happens inside one company, on its own servers, among a few hundred people with unlimited compute. The rest of the economy appears mostly as a stock price, a protest and a public approval rating. In the organisations I know, a new capability typically meets procurement, legal review, works councils, data protection, a risk committee and a budget cycle before it reaches anyone’s desk. That friction goes a long way to explaining Eurostat’s 13.5%.

The story also assumes that once AI can do AI research, progress compounds on its own. Maybe it does. But research also needs chips, power, data and money, and each of those is limited by things that move at the speed of construction permits and supply chains.

What I take seriously anyway

And yet the core mechanism does not depend on the rest of the economy. If a few labs can use their own models to speed up their own research, the gap between the frontier and everyone else can widen even while most businesses are still writing their AI policy. That is the part of the story I find hardest to dismiss.

There is some evidence for it already. In March, the research group METR measured how long a task, as timed in human working hours, AI agents can complete on their own. That length has doubled roughly every seven months for six years. Claude 3.7 Sonnet can now handle tasks that take a skilled person about an hour, half the time. If the trend continues, agents will handle week-long projects within a few years. Trends do end, but this one has held for six years, and I think that is long enough to take it into account.

Business leaders are starting to act on it. Earlier this month Tobi Lütke, the chief executive of Shopify, told his staff that using AI is now a baseline expectation, and that teams asking for more people must first show why AI cannot do the job. Whatever you think of the tone, that is a company redesigning its work around the technology, which the “normal technology” view says takes decades.

So I hold both views at once. The economy will absorb AI more slowly than the story suggests. The labs may move faster than the economy can absorb. The distance between those two speeds is where most of the risk sits, for businesses as well as for societies.

What I will watch

The best thing about AI 2027 is that it can be checked. Here are the markers I will be watching over the next eighteen months, all taken from the story’s own timeline:

1. Agents. Do AI agents that can operate a computer become reliable enough for everyday work by the end of 2025, or do they stay impressive demos?
2. Code. By 2026, is writing software visibly being taken over by AI, inside the labs and outside them?
3. Money and power. Does investment keep rising toward the hundreds of billions the story assumes, and does the electricity follow?
4. Jobs. Do junior roles in software and similar work start to shrink in 2026, and does public opinion turn against AI?
5. China. Does China stay within months of the American frontier, as DeepSeek suggests, and does its government concentrate its effort?

If most of these come true on schedule, the later chapters deserve to be taken much more seriously than most boardrooms take them today. If they don’t, I will have learned how much weight to give stories like this one.

In my work, the honest answer to “where will AI be in two years?” is that I don’t know. AI 2027 doesn’t claim to know either. It gives me a specific picture to argue with, and I find that more useful than another range of dates. I will come back to it at the end of next year and see how it held up.

---