In the summer of 1845, the potato crop in Ireland looked healthy. A few weeks later it lay black in the ground. A pathogen, known today as Phytophthora infestans, had reached the island. It struck other European countries too. Nowhere did it do what it did in Ireland.

The reason lay in the field itself. The poorest rural families lived almost entirely on a single variety, the Lumper. It grew in poor soil. It yielded a great deal. Potatoes are propagated from tubers, so every plant was a copy of the next. What infected one infected all. About a million people died, and more than a million left the country.

Nobody had planted the Lumper because it was bad. It was the most sensible choice a poor tenant farmer could make. That is the lesson. A single variety prevails because it is the best. Its weakness shows only on the day when everyone shares the same weakness.

Uniform Thinking

We are now planting such a field, this time for thoughts. Billions of people put their questions to a handful of language models. They have contracts explained, applications written, test results interpreted, arguments sorted. Depending on how you count, around four fifths of all visits go to two providers, OpenAI and Google.

These systems are good. That is why they are used, just as the Lumper was planted. Yet they resemble one another more than their names suggest. They learn from largely the same texts. Their makers tune them by similar standards for politeness, balance and caution. Ask three of them the same question and you often receive three versions of one answer.

With the search engine this mattered less. It showed ten sources, and the person chose among them. A language model delivers the finished thought. It selects which objections appear. It sets the tone in which a matter is discussed. Whoever does that every day for billions of people shapes the thinking of an era, whether they intend to or not.

A Prison Without Guards

The worry that follows comes in two pictures. In the first, the most powerful machine belongs to a single person who decides through it what billions read. In the second, the machine itself decides what people should think. Both pictures lead to a prison of the mind. Both recall the surveillance state George Orwell described.

Orwell, however, described coercion. His state watches, punishes, rewrites the past. People can rise against such a regime because they recognize it. The prison that threatens us needs none of that. Nobody forces us to ask the machine. We do it because it is convenient. The answer arrives faster than our own thought. Over time, our own thought becomes rarer.

Experiments support this suspicion. Groups that develop ideas with a language model arrive at more similar results than groups without one. Texts written with machine help converge in vocabulary and content. The individual often writes better than before. The community still loses breadth. The research is not unanimous, because under certain incentives the effect weakens. As a tendency it is well documented.

This changes the order of the danger. The single owner and the machine that decides alone stand at the end of a development. It begins much earlier, with a dependence that feels like a service. Only when a society has handed over its thinking does a ruler find the lever waiting. Anyone who wants to prevent the prison must therefore start with the field.

What Svalbard Keeps

Agriculture learned from Ireland. On Svalbard a vault lies inside a mountain, deep in the permafrost. It holds more than 1.4 million seed samples from almost every country on Earth. Hardly anyone still grows many of these varieties at scale. They yield less than the high-performance crops. They are kept because nobody knows which trait will one day be missing.

The vault has already been needed. When the war in Syria made a gene bank near Aleppo inaccessible, the researchers took their samples back out of the ice. They rebuilt the collection in Lebanon and Morocco.

Thinking has no such provision. There are language models from many countries, in many languages, with different emphases. Most are smaller than the market leaders and score worse in benchmark tests. The market sorts them out, as it sorted out the low-yield varieties. We should preserve them for the same reason we preserve seeds. A model that thinks in a small language knows distinctions that do not exist in English. A model trained on the legal texts of another continent weighs things differently. These differences are the reserve.

Who Owns the Diversity

Svalbard holds a second lesson. Every sample in the vault remains the property of the institution that deposited it. The operator keeps, it does not own. Diversity alone is therefore not enough. What matters is whose hands it lies in.

Today the leading language models lie in the hands of a few companies. That is no accusation, since they built and paid for them. But it means that a society's basic equipment for thinking is private property. Earlier generations faced a similar situation when books were expensive. Their answer was the public library. Knowledge should not depend on the size of a wallet, so the community built a house for it.

The library for language models is already taking shape. In Switzerland, ETH Zurich and EPFL, together with the national supercomputing centre, have released a model called Apertus. Its design, training data and code are open. The nonprofit Current AI is working on a free alternative with commitments of 400 million dollars. In Toronto, the Hypha cooperative runs a small model of its own, governed jointly by its members.

Each of these paths has a limit. The Swiss model is publicly funded. Behind Current AI stand the French state, large foundations, DeepMind and Salesforce. The cooperatives are tiny. What is missing is the fourth form. Wikipedia has shown that a work of global importance can live on donations. It belongs to no one who could sell it. A language model of that kind does not yet exist: carried by many small contributions, independent of state and corporation, owned by its users. Building it would be expensive. It is no longer impossible.

The Machine in Your Own Home

That it is no longer impossible is due to a development that gets little attention. Open models, whose blueprint anyone can download, now trail the best closed systems by only a few months. Smaller versions run on an ordinary laptop. The largest still need computers no household can afford.

The difference from a rented service is fundamental. A model on your own device does not change overnight. Nobody can decide from afar what it stays silent about. Nobody reads along. The private bookshelf once protected against censorship, because a printed book could not be recalled. A model in your own home does the same for the conversation with the machine.

Most people will not take this path. It takes effort. The rented service remains more convenient. That does not count against it. It is enough that enough people can take it. A society stays free as long as the exit stands open.

Where the Machines Disagree

Diversity in shared hands and in private hands creates a possibility that does not exist in a field with only one variety. Anyone given a serious diagnosis seeks a second opinion. Nobody regards that as distrust. It is the ordinary care owed to questions that carry weight.

The same habit can be applied to machines. An important question goes to several systems that were built independently of one another. Where they agree, the answer stands on firmer ground. Where they diverge, the real work begins. The disagreement marks the spot where a human must judge. Technically this is easy to build. A program asks the question several times and lays the differences side by side.

The precondition is real independence. Five offshoots of the same model do not contradict each other. They confirm one another, just as five plants of the same variety fall ill together. A second opinion has value only if it comes from a different field.

The Source

Beneath all fields lies the same soil. Language models learn from texts written by people. That stock is changing. A growing share of what appears online now comes from machines. The next generation of models learns from it too.

A study in the journal Nature has shown what happens then. A model trained mostly on machine-generated text loses the rare things first. The unusual phrasings vanish, then the dissenting thoughts. After a few generations an impoverished average remains. The researchers call this model collapse.

Physics knows a related case. After the first atomic tests, newly made steel carried traces of radiation. For sensitive instruments it was long unusable. Technicians therefore salvaged steel from ships that had sunk before 1945. Other methods exist today. The idea still holds. Some things can only be obtained from the time before a change.

The texts people wrote before language models arrived are such a stock. It no longer grows. Letters, diaries, newspaper archives, court records, dialect recordings, the books of small publishers: all of it is human thought with no machine admixture. It deserves to be secured and kept accessible. Future models will have to be measured against it.

What We Sow

Uniformity does not disappear by itself. Irish tenant farmers planted the Lumper again in the spring of 1845. Until then it had given them the best harvest every year. The leading language models will win in the same way, as long as nobody sows anything else. Waiting for their decline would be a mistake.

None of what needs doing requires a ban. It begins with the decision to preserve a small model even though it loses the comparison. It survives only if it belongs to someone who does not have to sell it. Whoever owns it can run it at home. Whoever runs several can question them against each other. What they are worth depends in the end on the texts they learned from.

Ireland's poor had no second variety in the ground in 1845. We still have the choice. So take the question that matters most to you and put it to a machine you have never used. Watch for the place where it answers differently from the one you know. That is where your own judgment begins.

| The Planet Futures Organization, the first multiplanetary NGO.

Homepage: https://planet-futures.org