The Future Has a Waiting Room

In Francis Bacon’s New Atlantis, the men who study nature have a remarkable privilege. They decide which of their discoveries the rest of the world will be allowed to see.

Their institution, Salomon’s House, gathers knowledge, conducts experiments and searches for ways to improve human life. Near the end of the account, its spokesman explains that members meet to choose which inventions to publish and which to conceal. They swear to keep some discoveries secret. Some are disclosed to the state. Others are withheld even from it.¹

The work is full of confidence in what knowledge can accomplish. Yet its promise contains a closed door. The people who possess the discoveries also decide how much everyone else should know.

That is an uncomfortable place to begin a book about progress. It is also the right place.

My position is clear: broad AI pauses are reckless. They make too many different kinds of work answer to a single stop signal, while the conditions we hope to improve continue doing damage. We should accelerate useful research and applications, and put far more effort into the testing, safeguards and public institutions that make progress trustworthy. Safety work should be a field of urgent invention in its own right. A society capable of building powerful models should not accept helplessness about how to test and govern them.

We have a responsibility to pursue the better future that greater intelligence may help us build. Disease, wasted effort and exclusion are not morally neutral conditions simply because we inherited them. If we can gain more power to reduce suffering, we should do the work of gaining it and make its benefits widely available. The danger of a darker future is a reason to steer with purpose.

A society can give great honor to invention and still leave most people with almost no say in the life invention makes possible. It can celebrate a new power while treating the distribution of that power as a small matter to be sorted out later. It can assure people that its work is for their benefit without giving them much chance to define the benefit, challenge the cost or refuse the bargain.

The familiar argument about artificial intelligence seems to offer two buttons. Press one to go faster. Press the other to stop.

Look more closely. Who has the console?

The person who cannot wait

Imagine a woman named Mara sitting at a public computer. She has an hour before the library closes. A letter in front of her asks for information she has already supplied, but in a different form and under a heading she does not understand. Her bus leaves shortly after closing. Missing it means another long wait.

A tool that helps her understand the letter, gather her own records and prepare an accurate reply could give her something precious. It could give her back the evening. Repeated across many such evenings, that help could change the shape of her life.

Now imagine that the office which sent the letter also uses an AI system. Its software flags Mara’s reply as suspicious. She cannot discover why. The worker on the phone can read the screen but cannot change the decision. Another letter arrives.

These are two uses of the same broad class of technology. One expands Mara’s room to act. The other may close it. The fact that both bear the name AI tells us very little about whether either should be deployed tomorrow.

Mara can want faster access to the first and stronger limits on the second without contradicting herself. Her position is more precise than the demand to speed up everything or halt everything. She is asking for help she can inspect and power she can challenge.

This book takes that request seriously.

The title, Don’t Pause My Future, expresses impatience with avoidable suffering and blocked possibility. People who lack a good school, a reliable service, an accessible workplace or enough time should not be asked to wait quietly while others debate their future from a position of comfort. But the title also belongs to the person whose future is interrupted by an unreliable decision, a lost livelihood or a system imposed without a practical way to object.

The word my matters. People have lives of their own. A larger economy, a faster machine and a more valuable company cannot stand in for every one of those lives.

So does the word future. It includes more than the next product release. It includes the ability to learn, to earn, to recover from a mistake, to disagree and to change direction. A tool that makes one task faster while making a person permanently dependent on an unaccountable provider may offer a gain worth having. It may also ask a price worth resisting. We should demand the benefit without surrendering the right to challenge the bargain.

The people carrying today’s suffering deserve urgency. The people who may be harmed by tomorrow’s systems deserve competent protection. Our obligation is to build the capacity to meet both claims, and to stop treating safety work as an afterthought whose neglect later becomes an argument against the whole enterprise.

Progress has an address

A machine becomes better at a test. A company announces the result. Investors revise their expectations. A newspaper calls it a leap forward.

What has changed for Mara?

Possibly nothing yet. The test may concern a task she never needs. The service may cost too much. It may fail in her language. It may require records she cannot retrieve. The office may refuse to accept a reply prepared with it. Or it may work beautifully and still leave her waiting because the queue is at the other end.

None of this makes the technical result worthless. It tells us that a technical result is one part of a larger chain.

First there is an ability: something the system can do under certain conditions. Then there is a use: a place where that ability fits an actual task. Next comes a result for someone who uses it. Finally there is the question of how widely that result is shared and what it costs other people.

Those stages can move at different speeds.

Suppose software reduces the time needed to prepare a report. The employer could use the saving to serve more clients. It could give workers a shorter day. It could reduce prices. It could raise profits. It could ask each worker to prepare more reports in the same day. Several of these choices might occur together.

The software does not settle the choice. Ownership, competition, contracts, public policy and bargaining power help settle it. When people speak as if productivity gains automatically become leisure or higher pay, they skip the room in which the distribution is decided.

It is possible to believe strongly in invention and still insist on entering that room.

Consider another difference. A service can become cheaper for its customer because work has become easier. It can also become cheaper because work has been shifted to someone who is not paid for it. If an automated office saves staff time but leaves residents correcting repeated errors, the savings recorded by the office will miss the burden placed on those residents. A genuine evaluation would follow the work across the boundary.

Mara’s hour counts even when it is absent from the budget.

A useful measure of progress therefore asks more than how much output increased. It asks whose effort fell, whose position improved and whose choices grew. It asks whether the gains survived outside a demonstration. It asks what happened to the person who could not use the system, the worker required to supervise it and the stranger affected by its output.

This is a demanding view of progress. It should be. The claims made for AI are large enough to invite it.

An ordinary product can be judged by whether its buyers like it. A technology proposed as a force that will remake science, work and public life owes us a broader account. If its supporters want to speak for humanity, they must be willing to look beyond the most profitable customers.

That does not mean every new tool must solve every social problem before anyone may use it. It means the size of the claim should match the size of the evidence. A good writing assistant can be praised as a good writing assistant. We do not have to turn it into a complete theory of human flourishing to justify its existence.

Waiting is work too

Delay often looks quiet from a distance. Nothing new has been installed. The old system remains. The person in charge can say that no additional risk was taken.

The person living with the old system may see things differently.

A difficult form still consumes the evening. A research question remains unanswered. A person who could work with the right support still cannot obtain it. These costs do not stop because a committee has not reached agreement. Existing harms continue on their own timetable.

We should be careful here. A possible future benefit cannot simply be counted as a benefit already lost. If a proposed tool has never shown that it works, there may be little reason to believe that deploying it now would solve the problem. The evidence has to support the comparison. But uncertainty about the benefit cannot justify pretending that the current situation is harmless.

There are two questions, and both deserve an answer. What could go wrong if we proceed? What continues to go wrong if we do not?

That position leaves room for specific delays with a defensible purpose. A system that makes hard-to-reverse decisions about people’s lives may require much stronger evidence than a private tool that helps someone draft a paragraph. The answers may favor a limited trial. They may favor offering help in one setting while forbidding a more dangerous use of the same underlying model.

That range of choices is the beginning of responsible ambition.

Imagine that Mara’s library is considering a tool that explains public forms. It could begin with a small, clearly marked service, avoid storing her personal records, show the official source beside its answer and provide a human route when it is unsure. The staff could check whether users actually complete their tasks and keep a record of serious errors. If the trial goes poorly, the service could be withdrawn while the library studies what happened.

Now compare an office that silently uses the same model to decide which residents should lose a benefit. It may be drawing on different data, exercising different power and producing harm that is much harder to repair. Even if the model’s underlying test score is unchanged, the public decision has changed greatly.

A sensible society should be able to distinguish these cases without declaring itself either for or against intelligence.

The person who needs help deserves a program that gets useful work done. Slowing everything is a poor substitute for learning which things can be made trustworthy and then making them available.

This is an argument about broad pauses. A temporary restriction on a named high-risk use, or on a defined training activity, must be judged by its actual scope. It does not automatically stop every existing tool or every scientific project. The honest case against a proposal follows the work it would delay; it cannot assign every unrealized cure to a rule that would never have touched the relevant research.

There will be moments when a restriction covering a whole category of dangerous use is justified. A capability may be difficult to contain, an exposure difficult to reverse, or the evidence too weak to support a safe boundary. That is a different decision from freezing the development of general capabilities across many uses. Restrictions can also protect those already powerful enough to obtain useful tools through private channels. Neither outcome should be assumed in advance. The details decide whether a restriction protects the public or mainly keeps others out.

A rule should name what it is restricting, why the restriction is needed and what would justify changing it. A release should name what has been shown, what remains unknown and who will respond when it fails. The public should not have to choose between indefinite secrecy and an experiment from which it cannot leave.

The right to a useful question

Bacon’s imagined researchers have knowledge that others lack. There is an obvious reason to listen to them. A person who studies a machine may understand its limits better than a person who has never seen it.

But technical knowledge does not answer every question about the machine.

How much risk should a stranger bear so that I can obtain a benefit? How should a gain be divided between the people who funded, built, maintained and used the tool? Which needs should public money serve first? When can a person refuse a system without being shut out of a necessary service?

These questions require facts. They also require judgments about what people owe each other. Expertise can clarify the consequences of a choice without giving its possessor sole authority to make that choice for everyone.

The reverse mistake would be to treat every technical claim as a matter of popular preference. A vote cannot make a failing test pass. A room full of sincere people cannot settle an engineering question by wanting a particular answer. Public participation should help choose the aims, the acceptable conditions and the people to whom decision-makers must answer. It should be supported by evidence good enough to expose wishful thinking, including popular wishful thinking.

We need institutions that can do both kinds of work. They must learn what is happening and decide what to do in a way that those affected can contest.

A person does not need to know how to train a model to ask why an official record cannot be corrected. A worker does not need a doctorate to ask who receives the saving from a faster process. A patient does not need to design a clinical study to ask whether the claim concerns a laboratory finding, a diagnostic result or a demonstrated improvement in health.

Those are intelligent questions. A public culture that treats them as obstacles to progress has confused progress with permission.

In Virtue of the Gods, I wrote: “It is ultimately about feeding the hungry, healing the sick, and building up the capabilities of the poor.”² That sentence states a purpose. It supplies no evidence that a particular technology will achieve it. The distance between the purpose and the result is precisely what this book must examine.

There is dignity in wanting more from invention than an impressive demonstration. There is also dignity in wanting the demonstration to succeed. A person can be moved by a scientific advance without signing over the right to question its use. Hope becomes stronger when it can survive a careful account of what has and has not been shown.

The temptation is to treat every qualification as a weakening of the cause. I think the opposite. A cause that cannot tolerate the true limits of its evidence is already planning to ask too much of the people who believe it.

Whose clock

Two people can look at the same year and see different amounts of time.

For one, a delay is an inconvenience. For another, it consumes a chance that will not return. At the same time, a hasty decision can leave a person with damage that an apology, a refund or a later software update cannot repair.

We cannot solve that conflict by choosing the person with the more moving story. Nor can we solve it by adding every imagined benefit to one column and every imagined disaster to another until the largest number wins. We must ask how the events could happen, how likely the evidence makes them, what choices remain and who is being left out of the account.

Sometimes the honest answer will be that we do not yet know enough. My response is to intensify the work that can resolve the uncertainty, rather than let uncertainty become a permanent command to wait. What study would help? Which use could proceed while another waits? What support can be offered to people who bear the delay? What evidence would cause us to stop?

The clock continues while these questions are asked. It also continues while they are ignored.

I favor continued AI development, serious investment in science and faster access to applications whose benefits stand up to testing. We should build the capacity to solve more problems, and we should work hard to bring that help within reach. A general mood of fear is a poor reason to keep useful power scarce. A specific danger, supported by evidence, is a reason to change what we build or how we release it.

My argument is that AI progress should be judged by the lives and powers it changes, and that those changes should be open to public challenge. Stronger tools can help us achieve things we have good reason to want. They can also increase the reach of a bad decision. The same increase in ability can strengthen a person, an institution or an abuse of authority. Its direction is not settled by the word intelligence.

This leaves us with more work than a slogan would. It leaves builders having to explain their claims, critics having to specify their fears and governments having to account for both the action they permit and the help they withhold. It leaves ordinary people with a claim to more than reassurance.

Salomon’s House offers the world its inventions when its members think fit. A shared future should ask more of us than gratitude at the door.

We should be able to enter, understand enough to question, and help decide what happens next.

The gate

We found no horned guard waiting in the night,
No pit of flame beyond the iron door;
A screen kept every patient name in white.

The clerk could call the wealthy to the floor.
The rest held folded forms instead of bread,
And heard the clean machine request one more.

No trumpet marked the minutes that had fled.
A woman held a place beneath her name;
The child asleep beside her must be fed.

Above the gate, the letters praised the flame
That drove the engines buried under stone.
Below, no printed line described her claim.

The lights were bright. The power was not her own.
She felt its warmth pass upward through the air,
And knew how crowds can leave a soul alone.

A man behind her whispered something like a prayer:
Let there be room for us within the day.
Let all this cleverness become repair.

We had not crossed an ocean to turn away.
We had not carried children to this seat
To learn that wealth alone could make us stay.

The future ran like current through the street.
We asked who set the price, who held the hand
That made the light withdraw from tired feet.

No answer came. We rose and made a stand,
Not high above the city in the sky,
But here, where every promise touches land.

And someone asked the only worthy why:
If all this power can open up the gate,
What raises us when only few rise high?

We did not come so far to learn to wait.

Sources and notes