A more powerful machine can leave a person with fewer choices. That is the problem this book begins with.
We can admire what an AI system does and still ask who gets to decide how it is used. A school might gain a tool that answers questions quickly. Teachers might lose the freedom to choose how they teach. A small business might gain cheap help with orders, then become unable to move its records to another service. A public office might process applications faster while making it harder for someone to correct a mistake. These are possible outcomes, not accounts of particular cases. They show why a score for technical ability cannot settle a question about public benefit.
The question is wider than whether AI is good or bad. Which task should it do? Who chooses the goal? What happens when it is wrong? Who can change the arrangement after money, records and habits have begun to depend on it?
My position is that people should have practical powers over the systems that shape their lives. They should be able to help choose their purposes, inspect the evidence for their use, challenge harmful decisions, and leave an arrangement when a better option exists. These powers will not look the same in every setting. A patient, an employee and a software developer need different forms of protection and access. But calling a system public, private or open does not tell us whether any of them has those powers.
This is a standard I am proposing for judgment. It is not a scientific result. The research that follows helps us test specific claims about how to meet it.
The goal comes before the score
Consider a simple decision. A public office wants to reduce the time people wait for a service. One plan gives staff an assistant that searches rules and drafts replies. Another uses a model to reject applications it judges unlikely to succeed. A third spends the same budget on clearer forms and more staff at the busiest times.
The right comparison is among workable ways to improve the service. It is not just between a new model and an old model. The office must count errors, time spent correcting them, staff work, cost and unequal treatment. It must also ask whether someone can still obtain help when they cannot use the new system. Faster processing has little value to a person whose correct application keeps returning the wrong answer.
The example is hypothetical. Its point is a decision rule: define the public problem before choosing the technology. Otherwise the available product quietly becomes the definition of progress.
That can happen without a villain. A supplier has a product to sell. A manager has a deadline. A budget contains money for software but little time for training. Each person may act reasonably within a narrow role, while the final choice leaves an important need unanswered. To change that outcome, we need to examine the rules and incentives around the decision, rather than simply demand better intentions.
A useful proposal should say what improvement would count as success and what evidence would make the buyer stop. If the goal is shorter waiting time, measure the full wait, including time spent appealing errors. If the goal is better access, include people who cannot use a standard interface. If the goal is less tedious work, ask workers which tasks have disappeared and which new duties have taken their place.
A target also needs an owner. Someone must be responsible for checking it after purchase. A promise with no person, budget or review date attached is difficult to enforce.
Technical change has a direction
Economists Daron Acemoglu and Pascual Restrepo offer a helpful distinction. Technology can raise output, take over tasks people used to perform, or create new tasks for people. Those changes have different effects on the demand for work. Their task-based model shows why higher productivity does not by itself guarantee that workers benefit. The balance depends in part on what new tasks appear and how production changes. This is a framework, supported by historical analysis, rather than a forecast for today’s generative AI. It does not tell us what will happen to a particular job. Automation and New Tasks
The practical lesson is to ask what the investment is trying to make possible. Does it help an experienced person solve a problem they could not solve before? Does it remove a task nobody wants? Does it replace a human service with a cheaper but worse one? A single label, automation, can cover all three.
These questions do not imply that every existing task must be preserved. Some work is dangerous, exhausting or needlessly repetitive. Keeping it merely because someone is paid to do it would confuse the job with the person’s interests. The challenge is to remove the burden while protecting the person’s chance to live well.
It also matters who can suggest a different use. Someone doing a task may know that the hardest part is checking missing information, while a buyer assumes the main problem is writing the final answer. A tool aimed at the wrong bottleneck can make the visible step faster and leave the whole process little better. That is a hypothesis about work design, to be checked in the setting where the tool will be used.
A trial should therefore compare the whole job, including preparation, review and repair. It should allow an awkward result. The best answer might be to change the workflow, use a narrower tool, or keep a human step. Evidence is useful only if it is allowed to change the purchase.
Ownership contains several questions
Ownership matters because it can give people income, decision rights and a way to influence what gets built. But those benefits do not always travel together.
A person can own a small financial stake without a meaningful say in product design. A researcher can use a model without being able to inspect how it was made. A community can be consulted without having power to reject a proposal. A public agency can own a system while the people affected by it cannot obtain an explanation or a remedy. These are different arrangements, and each calls for a different test.
Elinor Ostrom’s research helps make this distinction. She studied how groups governed shared resources and public services. Instead of treating ownership as one simple right, her account examined bundles of rights and the rules people could use to make decisions, monitor conduct and resolve conflict. She also resisted the idea that one institutional form would work in every setting. Her work does not establish how to govern AI. It gives us a reason to compare what an institution actually lets people do. Beyond Markets and States
For this book, four plain questions will guide that comparison.
Can people choose? This includes choosing a purpose before deployment, taking part in decisions that affect them, and retaining workable alternatives.
Can they inspect? They need enough relevant evidence to judge a claim. Sometimes that means public information. Sometimes it means qualified, independent access to confidential material. Publishing everything is not the only way to permit scrutiny.
Can they challenge? A person needs somewhere to take a disputed result, a responsible decision-maker, and a process that can change the outcome. A place to send feedback may help a product improve without helping the person who needs a decision corrected now.
Can they leave? An alternative must be usable at a cost they can bear. Records, skills, contracts and essential services can all make leaving harder. A formal right to switch has limited value if using it means losing the work required to keep a service running.
These are questions, not a universal ranking system. We will not award points and declare a winner. In some public services a person cannot simply choose another provider, so challenge and remedy become especially important. In research, access to computation may be the immediate limit. In a dangerous application, a restriction may be justified even though it reduces a user’s freedom. The task is to name the tradeoff and examine the reason for it.
Make the promise answerable
The title asks who owns tomorrow. No person or institution can own the future in a literal sense. They can, however, control resources and decisions that make some futures easier to pursue than others. That is the control worth examining.
This requires care with evidence. A count of important models is not a count of market power. A laboratory test is not a guarantee about use in the world. A government program can reach many researchers without proving that it reduced inequality. A successful workplace trial does not show that the saved time became higher pay or a shorter week.
We will keep those distinctions visible. We will use an executed descriptive analysis where the data permit one, and label proposals as proposals. We will also ask what would count against an appealing explanation. If a public investment reaches only groups that already have access, it has not yet shown that it opens new doors. If a disclosure gives a citizen no way to understand or contest a decision, visibility has not become practical power.
The ambition is still large. People should share in the benefits of better tools and retain a real say over how those tools shape work and public life. But an ambitious goal needs details that can be checked. Who decides, on what evidence, with which alternatives, and with what power to correct a mistake? Those questions turn a vision of progress into a plan people can examine.