A new language model called Talkie was trained only on material published before the end of 1930.
When asked to imagine the world of 2026, it predicted that one billion people would live in Europe. It described London and New York as connected by steamships that crossed the Atlantic in ten days. It imagined the cultured classes still spending winter in Paris and summer in London — as if the future were mainly a longer version of European social seasons.
The predictions sound absurd.
But that is not what gives the experiment its force.
Of course a model trained on information from 1930 would struggle to predict the modern world. It knew nothing of nuclear weapons, commercial aviation, antibiotics, computers, the internet, or the political events that would reshape the rest of the century.
The more revealing point is that the people living in 1930 were trapped inside the same information.
Talkie is not merely a demonstration of the limits of artificial intelligence. It is a reconstruction of the limits of an entire historical present.
It shows us what intelligence looks like when it is confined to the facts, assumptions, vocabulary, and unanswered questions of its own time.
And that should make us uncomfortable.
Intelligence is not the same as foresight
It would be easy to dismiss Talkie's predictions by saying that the model is not intelligent enough.
Perhaps a more capable model would have done better. It might have detected scientific trends, connected ideas that humans had not yet connected, or made better estimates about technological progress.
But this misses the deeper point.
The deeper claim is not that today's models outperform Talkie. Even a much more capable version would face the same temporal boundary. The constraint is not performance. It is the information available inside a historical present.
No matter how intelligent the model became, it would still be reasoning from inside 1930.
It could infer things that were not explicitly written. It could combine existing ideas in original ways. It might even anticipate discoveries that were already latent in the science of the time.
But it could not reliably account for information that did not yet exist, decisions that had not yet been made, or concepts that nobody had developed.
The same was true of the smartest humans alive in 1930.
And it is true of us now.
When a modern model describes 2050 or 2122, it does not sound quaint or obviously constrained. It can discuss artificial intelligence, biotechnology, climate change, demographics, robotics, energy, and geopolitics. Its vocabulary feels broad enough to contain the future.
But it is still reasoning from inside the present.
So are we.
I recently asked a current model to imagine the world in 2122 and an ordinary day inside it. The forecasts are fluent and internally coherent: a slowly shrinking population of about 9.7 billion, cities reshaped by climate migration, electrified transport, AI systems woven into work and medicine, shorter working weeks, and a planet permanently warmer by a little more than two degrees. None of it sounds ridiculous from here.
That is exactly the problem.
Someone in 1930 might have found steamships and European seasons equally reasonable. The question worth sitting with is which parts of our 2122 pictures will someday look as wrong as a ten-day Atlantic crossing — and what categories those pictures cannot contain at all.
We share the model's blind spots
When a forecast from 1930 gets the future wrong, we can see why.
We know about the technologies it missed. We know which political assumptions failed. We know which scientific beliefs were incomplete. We can point to the empty spaces in its worldview because history has already filled them.
We cannot do the same with our own worldview.
Some of the things we currently believe are wrong. Others are incomplete. Some questions that appear central today may eventually become irrelevant. Some problems may be solved through ideas that we do not yet possess.
We know this must be true.
We simply do not know which beliefs are wrong, which questions are badly framed, or where the missing concepts belong.
That is the strange position of every present.
The past looks obviously incomplete because we know what came next. The present feels much more complete because its missing pieces remain invisible.
Historical distance makes that asymmetry easy to see.
Its mistakes are not evidence that people in 1930 were foolish. They were using the best map available to them. The problem was that they could not see where the map ended.
Neither can we.
Some of what we know is not knowledge
The limitation is not only that important information is missing.
Some of the information already inside our model of reality is probably wrong.
A medical model frozen at different moments in history would make this clear. It might confidently recommend treatments later found to be ineffective. It might repeat nutritional advice that would eventually be qualified or reversed. It could correctly describe the consensus of its period while still giving the wrong answer.
The same model would contain several kinds of beliefs.
Some would remain true.
Some would be incomplete.
Some would be treated as facts but later shown to be false.
The dangerous category is the last one. Missing information sometimes appears as uncertainty. Incorrect information often appears as knowledge.
That applies to people as much as it applies to models.
We do not experience our current beliefs as a mixture of truth and future error. We experience them simply as what is known.
Only later does history separate the two.
The future does not extend the present in a straight line
Most predictions begin by extending the trends already visible.
That is often useful over short periods. The immediate future usually inherits much of the structure of the present.
But the farther we look, the more fragile the exercise becomes.
A person in 1930 might reasonably predict faster ships, larger cities, more automobiles, and greater electrification. Those trends were real.
The problem is that the future does not only continue trends. It introduces new forces.
Commercial aviation did not merely improve the steamship. It changed the relevance of the steamship.
The computer did not merely accelerate existing clerical work. It transformed what work could be.
New discoveries do not always answer the questions a previous generation was asking. Sometimes they replace those questions entirely.
This is why long-range forecasts so often resemble the present with better machinery.
We imagine more powerful versions of the things we already understand. We have a much harder time imagining the things that will make our current categories obsolete.
The farther the prediction travels, the less evidence constrains it.
But confidence does not always decline at the same speed.
Neither does fluency.
This is a lesson about human confidence
People often speak about the future with extraordinary certainty.
They do it in discussions about climate, healthcare, geopolitics, technology, economics, and society. The future is described not as one possible outcome, but as something nearly settled.
The prediction may be optimistic or catastrophic. The problem is not its emotional direction. It is the certainty.
The experiment reminds us that a coherent forecast can still be built inside an incomplete worldview. A prediction can use all available evidence, follow a plausible chain of reasoning, and still miss the forces that will matter most.
This does not mean that all forecasts are useless or that nothing can be known.
It means that we should separate what appears likely from what is inevitable.
We should ask which assumptions a prediction depends on. We should consider how far into the future the evidence can reasonably reach. We should leave room for discoveries, decisions, and events that are not yet represented in our model of the world.
Most importantly, we should remember that our ignorance does not always feel like ignorance.
Sometimes it feels like certainty.
We are also historical people
When we look at the past, we grant people the humility of context.
They could not have known about the transistor. They could not have predicted the internet. They did not yet possess the evidence that would overturn a medical belief or reshape an industry.
But we rarely extend the same humility to ourselves.
We quietly assume that we now see the major forces. We imagine that the important categories are already visible and that the future will mostly be a struggle over how those forces develop.
Every generation may have believed some version of this.
The experiment gives us a way to experience the mistake from the outside.
It shows us an intelligent system producing a confident picture of the future from a world we know was radically incomplete.
The uncomfortable lesson is not that the system was trapped in 1930.
It is that we are trapped in 2026.
History has not yet shown us what we are missing.