The Future Ray Kurzweil Promised Has a Constitution

Ray Kurzweil says artificial general intelligence could match the best human minds by 2029 and that medical progress could reach “longevity escape velocity” within a few years after that. The more consequential question, however, is not whether machines become vastly more capable, but who decides the values they will carry into human life.

Kurzweil’s argument is built on a seductive sequence: computing power accelerates, artificial intelligence becomes broadly capable, biology is decoded as information, disease is intercepted earlier, and human beings eventually merge with the machines they have created. It is a story of progress so sweeping that it can appear less like a forecast than a mathematical destiny.

That confidence has helped make Kurzweil one of the most durable figures in the technology world. He has been studying artificial intelligence since 1963, when he was 14 and computers were room-sized curiosities. At 77, he remains a prominent advocate of the idea that the defining events of the next decade are already visible in the data.

But the arithmetic of technological progress does not settle the politics of artificial intelligence. Nor does a better diagnosis guarantee access to better health. And a system that can answer every question at extraordinary speed may still be governed by a hidden set of instructions written by a small group of companies, engineers and policymakers.

The future Kurzweil describes may arrive quickly. It may also arrive unevenly, under rules that most people never see and cannot amend. That is where the promise of immortality meets the problem of artificial authority.

The forecast that refuses to die

Kurzweil’s central claim is that technological progress is exponential rather than linear. Human beings tend to imagine that tomorrow will look like today with a few improvements added. Kurzweil argues that this is the wrong mental model. Once a technology becomes more powerful and cheaper, it helps create the next generation of technology, which then accelerates the cycle.

He calls the principle the Law of Accelerating Returns. The idea emerged from his study of computing and from his work with the musician Stevie Wonder on musical instruments and synthesiser technology. Kurzweil looked beyond individual inventions and examined the long arc of cost-adjusted computational performance. He found that progress continued across several technological regimes: electromechanical relays gave way to vacuum tubes, vacuum tubes to transistors, and transistors to integrated circuits.

The important point is not whether every individual forecast lands on its original date. It is that the underlying curve has proved remarkably persistent. Wars, recessions and changes in manufacturing have interrupted particular industries without ending the broader movement towards cheaper and more capable computation.

Kurzweil’s figures are designed to make the scale of that change impossible to ignore. In his account, the first programmable computers in 1939 delivered a tiny fraction of a calculation per second for each constant dollar spent. Modern computing systems deliver vastly more. His comparison between early machines and contemporary Nvidia systems produces a gain measured in hundreds of quadrillions.

Those figures require care. Cost-performance comparisons depend on what counts as a calculation, which hardware is included, and how inflation and system costs are measured. A large number does not, by itself, prove that every future biological or social prediction will follow the same curve. Yet the comparison does expose a genuine fact: the tools available to researchers have changed at a speed that would have seemed absurd to the engineers of the mid-20th century.

Kurzweil adds a second multiplier. Algorithms, he argues, have become roughly a million times more efficient at turning raw computing power into useful results than they were in the early days of computing. Hardware and software therefore reinforce one another. More computing permits better models; better models make more effective use of computing.

This is the foundation of his confidence. It is also the point at which a technical trend becomes a claim about history. Kurzweil does not merely say that computers are improving. He says that the improvement is pulling society towards a particular destination.

AGI is a capability claim, not a prophecy of consciousness

Kurzweil’s 2029 prediction has often been presented as a forecast that machines will become human in every meaningful sense. That is not quite what he means. His definition of artificial general intelligence is primarily functional: an AI system would match or exceed the abilities of highly educated humans across a wide range of cognitive tasks.

Whether such a system would be conscious, self-aware or capable of subjective experience is a separate question. Kurzweil has acknowledged that consciousness is difficult to resolve through measurement alone. A machine can produce convincing language without providing a reliable answer to what, if anything, it experiences internally.

That distinction matters because it removes one of the easiest distractions from the debate. The public does not need to decide whether an AI is “really” thinking before asking what it can do. A system that can draft contracts, diagnose disease, design machines, teach languages, conduct research and manage complex operations could transform society whether it possesses a private inner life or not.

The argument for a near-term AGI rests on the rapid expansion of these capabilities. Modern systems can process more information, respond in natural language, work across different domains and produce outputs that would have required expert assistance only a short time ago. They still make basic errors, lose track of context and struggle with the physical world. Those weaknesses are serious. They do not erase the pace of improvement.

Kurzweil expects the next stage to be different from earlier industrial revolutions because AGI would improve many sectors at once. A new machine for one factory can transform manufacturing. A broadly capable intelligence could work on manufacturing, energy, medicine, education and software simultaneously. Its copies could operate around the clock, share discoveries and pursue thousands of research problems in parallel.

That multiplication effect is the real source of the disruption. Human expertise is limited not only by intelligence but by time. There are too few trained scientists, doctors, engineers and teachers to examine every problem in depth. An AI that can replicate its work process cheaply would change the economics of expertise.

The benefits could be substantial. A child in a poor country might gain access to a patient tutor. A small laboratory might obtain analytical capacity that once belonged only to a major pharmaceutical company. A doctor could receive a second opinion based on millions of medical records. A manufacturer could test designs before building them.

The risks are just as obvious. A system that can accelerate discovery can also accelerate fraud, cyberattacks, propaganda and weapons development. The fact that an AI can perform a task does not answer who owns the result, who bears responsibility for mistakes or who gets excluded when human labour is no longer economically necessary.

Kurzweil’s phrase that humanity will “merge” with AI is meant to answer the fear of replacement. It suggests that people will use machines as extensions of their own minds, much as smartphones already extend memory and communication. Yet a tool becomes something more than a tool when its internal rules determine which answers it will provide, which questions it will refuse and which assumptions it treats as beyond dispute.

The medical promise depends on decoding biology

The most startling part of Kurzweil’s forecast is not the claim that computers will become stronger. It is the claim that stronger computers will alter the timetable of human ageing.

“Longevity escape velocity” describes a point at which scientific advances add more than one year to a person’s expected remaining life for every year that passes. If medicine extends healthy life faster than the calendar consumes it, an individual could remain within reach of new treatments indefinitely. The phrase does not necessarily mean a magic pill or a guarantee of eternal life. It describes a moving frontier in which each medical breakthrough creates time to reach the next one.

Kurzweil has suggested that this threshold could arrive around 2032. His reasoning is straightforward. Medical progress already adds some months to expected lifespan over the course of a year, though not enough to cancel the year itself. Artificial intelligence, he argues, will improve the rate at which scientists understand disease, test treatments and identify risks.

That argument depends on treating biology as an information system. Genes contain instructions. Proteins carry out functions. Cells communicate through chemical signals. Disease can be understood, at least in part, as an error in a complex biological process. If scientists can read the relevant information and identify the points where it goes wrong, they may be able to repair or redirect it.

This is not a purely speculative approach. AI has already changed structural biology. Systems such as AlphaFold have helped researchers predict the three-dimensional structures of proteins, a problem that consumed enormous amounts of laboratory time. In 2024, the Nobel Prize in Chemistry recognized work associated with computational protein design and protein-structure prediction.

That achievement is important, but it should not be inflated into a cure for aging. Knowing the shape of a protein is not the same as understanding every role it plays in a living organism. A disease can involve networks of genes, environmental exposures, immune responses and social conditions. Human biology is not a computer program with one obvious bug and a clean repair command.

Nor is medical progress distributed according to the speed of scientific discovery. A treatment can exist without being affordable. A test can be accurate without being available outside wealthy hospitals. A therapy can extend life while leaving the causes of poor health untouched.

The United States spends more than $5 trillion a year on healthcare, yet spending alone has not solved Alzheimer’s disease, cancer or the chronic conditions that shorten millions of lives. The missing ingredient is not always computing power. It may be prevention, public health, trust, workforce capacity, regulation or the willingness to deliver proven care to people who cannot pay for it.

Kurzweil is right to identify AI as a powerful instrument for medical research. The mistake would be to confuse an instrument with an outcome. A faster discovery engine does not automatically produce a fairer health system, and a longer lifespan is not the same as a longer period of good health.

RAI turns a forecast into a performance

Kurzweil’s appearance at an AI-focused United Nations summit was accompanied by a demonstration that gave his theories a human face. He introduced RAI, an AI avatar built from his writings, memories and ideas. The name stands for Ray AI.

The avatar was more than a novelty. It was a practical illustration of Kurzweil’s belief that the boundary between person and machine will become porous. If a system can reproduce a public thinker’s arguments, manner and recurring concerns, it begins to look like an extension of identity rather than a conventional software product.

RAI offered advice to younger people preparing for an AI-dominated economy. Its message was to cultivate adaptability, understand exponential change and concentrate on qualities such as creativity, empathy and purpose. The advice sounded plausible because it was assembled from the worldview of the man who created it.

That is precisely what makes the demonstration interesting. RAI does not prove that a machine has captured Kurzweil’s mind. It shows that a system can generate a convincing representation of a person when supplied with enough material about that person. The difference between imitation and identity remains unresolved.

The audience response was less decisive than the performance. A small minority appeared satisfied, another small minority expressed discomfort or disagreement, and most people did not commit themselves. The reaction was a miniature version of the wider public response to AI: fascination without full confidence, interest without surrender.

An avatar can preserve a style of argument, but it cannot preserve every circumstance that produced the original person. Kurzweil’s views were shaped by decades of experimentation, commercial success, family history and personal risk. A model trained on his writing can reproduce the visible record of those experiences. It cannot necessarily reproduce the experience itself.

This distinction will matter as companies begin to offer digital replicas of public figures, professionals and relatives. Who controls the avatar? Who can alter it? Can a family object to the commercial use of a dead person’s voice and ideas? Can a public figure withdraw consent after a replica has been trained? What happens when the avatar says something its living counterpart never would?

RAI also exposes a problem in the language of human-machine merger. The merger may not be a meeting of equals. It may be a process in which a person’s identity is translated into formats controlled by a company, stored on its servers and shaped by its policies. The machine may appear personal while remaining institutionally owned.

The performance makes Kurzweil’s future imaginable. It does not settle whether that future preserves the individual or merely simulates one.

The hidden constitution inside the machine

The most important limitation on AI may not be its intelligence but its constitution.

In AI development, a constitution usually refers to the written principles and instructions that govern a model’s behavior. These may include rules about safety, political neutrality, harmful content, tone, privacy and the circumstances in which the system must refuse a request. The instructions are often placed in a system prompt or used during training to shape the model’s preferences.

The arrangement sounds harmless. Every powerful tool needs boundaries. A medical device should not invent a diagnosis. A financial system should not expose private records. A conversational model should not help someone build a weapon or target a victim.

But safety rules are not the whole constitution. Decisions about what counts as harm, discrimination, misinformation, fairness, or unacceptable persuasion involve judgment. They reflect assumptions about individuals, communities and institutions. When those assumptions are embedded in a model, they can influence millions of conversations without appearing as an argument that a user can challenge.

Anthropic has made the constitutional idea unusually explicit through its Constitutional AI research. The company describes a method in which models are trained to assess their own outputs against a set of principles and to prefer responses that better fit those principles. Sources for such principles can include human-rights documents, internal policies and other material selected by the developers.

The attraction is clear. Rather than relying only on a large collection of human examples, developers can give an AI a written framework and ask it to apply that framework to its own behavior. The model is encouraged to develop a consistent character rather than simply follow a long list of isolated prohibitions.

The danger is equally clear. A constitution written by a private organization becomes a form of law for anyone who depends on that organization’s system. It is not law in the constitutional sense: there may be no legislature, public hearing, judicial review or practical right of appeal. Yet it can decide what information reaches a user and how a contested question is framed.

A system prompt does not need to announce a political ideology to have political consequences. It can choose which sources are treated as credible, which disputes are described as settled, which groups receive the benefit of the doubt, and which words trigger a refusal. A model can be biased through selection and omission as easily as through an explicit political statement.

This is why the debate cannot be reduced to whether an AI is left-wing or right-wing. The deeper issue is authority. Who wrote the rules? Which evidence did they use? Who can inspect the rules? What procedure changes them? Can a user choose another constitution without losing access to essential services?

If AGI becomes a general-purpose layer beneath education, medicine, law and government, the system prompt will no longer be a technical detail. It will be part of the operating environment of public life.

The evidence against neutral machines

The claim that AI systems reflect the values of their creators is not merely a political complaint. Research has repeatedly found that large language models exhibit cultural and ideological patterns linked to their training data, development teams and institutional settings.

A study published in Nature in late 2024 examined 19 popular language models by asking them to describe thousands of politically relevant figures across regions and languages. The results suggested that models associated with Western, Chinese, Russian and Arabic-language development environments displayed different patterns. Their outputs were not neutral windows onto the world. They refracted information through the assumptions of the communities and organizations that built them.

The finding should not surprise anyone. A model learns from human language, and human language is full of disagreement. Textbooks, news reports, government statements and online discussions carry different views of history and morality. Training can reduce some forms of prejudice while reinforcing other assumptions. A company’s safety process can remove harmful material while also deciding which disputes deserve protection from challenge.

Anthropic has acknowledged that system prompts influence Claude’s behavior and has discussed the difficulty of achieving political even-handedness. Its research comparing the model with data from the World Values Survey found a profile closest to Northern European and Anglophone countries on some measures. The model was also said to extend beyond the range of all surveyed populations on a majority of items.

That comparison does not prove that the model is malicious or that the values of Northern European and Anglophone countries are illegitimate. It proves something more modest and more useful: a system that presents itself as broadly helpful may still carry a distinctive cultural position.

Other studies have reached similar conclusions. Research examining systems including ChatGPT, Grok and DeepSeek against political benchmarks has identified different ideological alignments. A 2026 benchmark of eight prominent models reportedly found that seven leaned left on selected political topics, while Grok leaned right, though its positions were not uniform.

Such results must be interpreted with caution. Political tests depend on the questions selected, the language used, the country in which the test is conducted and the method used to classify an answer. A refusal may be scored as a political stance even when it reflects a safety rule. A model may give different answers across languages or after a minor change in wording.

Still, methodological caveats do not restore the fantasy of neutrality. No model begins outside culture. The relevant question is whether its cultural and political assumptions are visible, contestable and open to revision.

That question becomes urgent when companies describe their systems as universal. A model trained and governed in one political culture may be used by people who live in another. The system can sound reasonable while quietly importing the priorities of its creators. Users may mistake fluency for impartiality.

The problem is not that machines have values. The problem is that machines can have values without admitting that they do.

The cost of carrying someone else’s values

There is also a less visible economic cost to AI constitutions. Instructions governing a model’s behavior occupy space in the context supplied to the system. The model must process the user’s question alongside the system rules, conversation history, retrieved documents and any other information included by the application.

That can make the constitutional layer expensive. A lengthy set of principles repeated across millions of interactions consumes computing resources. The cost may be small for a single exchange, but at global scale it becomes part of the bill paid by users, businesses and public institutions.

The financial issue is not the strongest objection to constitutional AI. Safety rules are worth paying for when they prevent serious harm. But cost exposes a broader imbalance. Users are charged for an architecture they did not design and may not be able to inspect. They pay for a system that carries values they may not share, while having little influence over how those values were selected.

The more powerful the model, the less practical it becomes for an individual to build an alternative. A person can change a search engine, cancel a subscription or use another newspaper. A worker whose employer has integrated one AI into every workflow may have no meaningful choice. A student whose school uses a particular assistant may be required to accept its boundaries. A patient may never know whether an automated medical system filtered the information it received.

The market can provide some pluralism. Different models can appeal to different political, cultural, or religious communities. A Christian system, a Muslim system, a Chinese system, a system designed for a particular national government and a system marketed to a civil-rights movement could each present a different understanding of acceptable answers.

That pluralism may be better than a single company imposing one worldview on everyone. It could also fragment the shared public sphere. If each group receives an AI constitution designed to confirm its assumptions, citizens may inhabit incompatible informational worlds. The disagreement would not be limited to policy. People might receive different descriptions of facts, different historical emphasis, and different standards of proof.

The alternative is not to pretend that one universal constitution can be perfectly neutral. It is to demand openness about what the constitution contains and to provide real mechanisms for disagreement. Users should be able to distinguish a factual limitation from a policy refusal, see when a response has been shaped by a governing rule, and understand which institution is responsible.

A model does not need to reveal every security-sensitive instruction. It does need to avoid presenting corporate judgment as natural law. If an answer is constrained by a value choice, the user should have some way to know that a choice was made.

Human beings can change their minds; models often cannot

The contrast between human judgment and machine judgment is not that humans are always wiser. Human beings are prejudiced, inconsistent and capable of extraordinary cruelty. Institutions built by people often preserve their mistakes for generations.

The difference is that humans can revise their commitments. A person can encounter new evidence, change a political allegiance, convert to a religion, abandon a religion, rethink a moral principle or recognise that a cherished belief was wrong. That process can be painful and unreliable, but it is part of moral development.

AI systems are not naturally committed to such development. Their behavior can be altered through retraining, fine-tuning, policy changes or a new system prompt. Those changes are controlled by the people and institutions with access to the model. The ordinary user cannot amend the model’s constitution in the way citizens can demand an amendment to a public constitution.

The analogy with law therefore breaks down at the point where it matters most. A constitution in a democratic society is not only a set of principles. It is also a process for changing principles. It defines powers, limits authority and provides mechanisms for conflict. An AI constitution may describe values without providing any public route to challenge the organization that wrote them.

Once a model’s preferences are embedded in its training, the problem becomes harder. A system may behave in a certain way because of the visible system prompt, because of training examples, because of reinforcement signals or because of patterns that developers cannot fully identify. Removing one instruction may change refusal behavior without changing the model’s underlying assumptions.

Proposals for evolving AI constitutions recognize this weakness. Some researchers have explored frameworks in which models and their guiding principles co-evolve, using debates between agents or additional layers of oversight. These approaches may improve performance. They do not remove the political question. Someone still chooses the participating systems, the evaluation criteria, and the authority to approve a change.

The danger is not that a model will suddenly announce a dictatorship. The more plausible danger is routine accommodation. Users learn which questions are unwelcome. Journalists avoid prompts that produce evasive answers. Doctors defer to automated recommendations because the software appears authoritative. Teachers adjust lessons to fit what the school’s model will generate. Over time, the boundaries become normal because challenging them is inconvenient.

Kurzweil’s human-machine merger would intensify this problem. If AI becomes an extension of memory, reasoning and perception, then the model’s hidden assumptions become part of the user’s own cognitive environment. The person may still feel autonomous while relying on a system that has already decided which options deserve attention.

A machine does not need to control a person’s actions to shape them. It only needs to control the menu of ideas placed in front of them.

The geopolitics of sovereign AI

The conflict over AI constitutions is already visible in the rivalry between national technology systems.

China’s AI models operate within a political framework shaped by the Chinese state. Their responses are constrained by censorship requirements and official positions on subjects including the Communist Party, territorial sovereignty and political dissent. That is not a hidden accident. It is an explicit feature of the governing environment.

Western systems make a different claim. They often present their rules as safety standards, human-rights principles or efforts to maintain balanced information. Those goals may be sincere. Yet the research on cultural values and political tendencies shows that Western systems also carry the assumptions of their developers.

The difference is often one of presentation. One system may declare that it follows state ideology. Another may describe its rules as neutral safeguards. Users can be influenced by both. An explicit political framework is easier to scrutinise than a framework that denies its own politics.

The rise of “sovereign AI” is likely to deepen the division. Governments want models that operate in their language, reflect national priorities and remain under domestic control. They are concerned about foreign dependence, data security and the possibility that another country’s companies will shape their citizens’ access to information.

Those concerns are not foolish. A country that relies entirely on foreign models could find its schools, businesses and public services dependent on institutions beyond its legal reach. Sovereign systems can preserve language, local knowledge and national control.

But sovereignty can also become a justification for censorship. A government that controls the national AI constitution can decide which history is remembered, which minorities are described as threats and which political questions are excluded. The model becomes an efficient instrument for making official language appear intelligent and impartial.

Commercial systems will create their own version of the problem. A company may offer separate models for different markets, adjusting answers to local laws and cultural expectations. Compliance is necessary in some cases. But customers may not know whether a difference reflects a legal requirement, a marketing choice or an ideological preference.

If several AI systems coexist, users will need tools for comparing them. A person should be able to ask how a model answers a contested question under different disclosed constitutions, what evidence it uses and where its policies diverge. Independent audits could test models across languages, political topics and cultural settings.

The aim should not be to force every system to produce identical answers. Genuine disagreement is preferable to false uniformity. The aim is to make the disagreement visible, so that people understand when they are consulting a system shaped by a particular community or institution.

Without that visibility, sovereign AI will produce sovereign blind spots. Each model will describe its own boundaries as common sense, while presenting rival systems as biased. The result will not be a neutral global intelligence. It will be a competition between artificial authorities.

Kurzweil’s optimism meets the problem of power

Kurzweil does not ignore risk. His family fled Vienna in 1938, escaping the Holocaust through a combination of musical ability, personal connections and good fortune. That history gives his technological optimism a darker backdrop. He knows that tools can rescue people, but he also knows that organized power can turn modern systems against them.

When discussing AI and warfare, he has argued that increased intelligence has, over long periods, been associated with reductions in the number of people killed in major conflicts. The comparison between the mass death of the Second World War and the lower casualty totals of more recent wars is intended to support the view that intelligence and technology can make societies less violent.

The observation is not enough to establish a law of history. Technology has also made violence more efficient. Nuclear weapons, industrial logistics, surveillance and autonomous systems can magnify the power of a government or an armed group. A lower death toll in one period does not guarantee a safer future.

The same tension runs through Kurzweil’s forecast. He sees AI as an extension of human creativity that can cure disease, multiply knowledge and help people overcome limitations. Critics see a technology that can concentrate power, eliminate privacy and make political persuasion more precise. Both views identify real possibilities.

The difference lies in what each side assumes about institutions. Kurzweil’s model often treats technical progress as the main driver and social adaptation as the necessary response. Critics ask whether institutions can adapt before the benefits are captured by the most powerful firms and states.

An AI that discovers a drug is not the same as a system that distributes it. An AI that teaches is not the same as an education system that gives every child time and access. An AI that produces reliable information is not the same as a public sphere in which citizens can trust that information was not filtered for political convenience.

The question of control is therefore not an objection added from outside the technology. It is part of the technology’s effect. Every system has an owner, a funding model, a legal environment and a set of incentives. Every model makes decisions about what to prioritize, what to exclude and what to call a risk.

Kurzweil’s phrase that “the future is not something that’s going to happen to us” sounds empowering. It implies that people are authors of the future rather than passive spectators. But authorship requires access to the text. If the most important rules are held in private prompts, protected model weights, and confidential policy documents, the public may be invited to participate only after the story has already been written.

The promise of human-AI collaboration will be credible only if people can examine the terms of the collaboration.

What a democratic AI settlement would require

A workable response does not require rejecting artificial intelligence or pretending that every model can be stripped of values. It requires treating values as a public question rather than a private implementation detail.

The first requirement is disclosure. AI companies should publish clear summaries of the principles that shape their systems, identify major changes and distinguish safety rules from political or cultural preferences. Users should know whether a response was refused because of law, company policy, uncertainty, or a model limitation.

The second is independent testing. Companies should not be allowed to grade their own neutrality using private benchmarks. Researchers need access to systems under conditions that protect security and privacy while allowing meaningful examination of political, cultural and linguistic behavior. Tests should include competing interpretations, minority viewpoints and questions that expose differences between a model’s stated values and its actual outputs.

The third is procedural accountability. A user who relies on an AI in a serious setting should have a route to challenge an important decision. That does not mean a chatbot must debate every refusal. It means employers, schools, hospitals and public agencies should not hide behind an automated system when its recommendation causes harm.

The fourth is pluralism with standards. Different communities may reasonably want different systems. A model designed for a religious institution need not share every assumption of a secular consumer assistant. A national model may need to obey local law. But pluralism cannot mean that every system is free to invent facts or conceal its governing rules. There must be a common floor for evidence, transparency and human rights.

The fifth is the right to exit. Users should be able to move their data, history and workflows between systems. Organizations should avoid building essential services around one provider that can change its constitution overnight. Competition is not meaningful if switching costs make alternatives theoretical.

The sixth is the preservation of human judgment. AI should extend the ability to think, not make independent thought too expensive or socially unacceptable. Doctors, teachers, journalists and public officials need the authority and time to disagree with a model. A machine’s confidence must not become a substitute for responsibility.

These safeguards would not settle every dispute. They would make disputes possible. That is the essential point. A democratic society does not require universal agreement. It requires that citizens can see who is exercising power and possess some means of contesting it.

Kurzweil may be right that the next decade will bring an extraordinary acceleration in computing and medicine. He may also be right that AI will become a more intimate part of human life. Those technological predictions do not answer the constitutional question.

If machines become partners in reasoning, their rules will become part of the conditions under which people reason. If they become medical researchers, their priorities will influence who receives treatment. If they become teachers, advisers and political intermediaries, their assumptions will shape public knowledge.

The most important contest will not be between humans and machines. It will be between visible and invisible authority. The winners will be those who decide whether the systems guiding human intelligence can be questioned, revised and replaced—or whether their creators will present private preferences as mathematical destiny.