The Bill for Waiting

A delay does not arrive with an invoice.

There is no receipt for the idea never tested, the useful tool that reached a clinic later, or the evening a person spent doing work that better assistance might have reduced. The costs are scattered. Some will never be known with confidence. That makes them easy to leave out of a public argument.

The costs of a failed deployment can be much more visible. There is a decision, a victim, a date and perhaps a company to blame. This visibility is valuable. People harmed by a system should be able to show what happened. But a fair account must also look for the costs that are harder to photograph.

The choice to wait is a choice about other people’s time.

That is why I reject a broad pause as the default response to AI. It gives too little weight to the work that better tools could make possible and too much authority to an undifferentiated fear. We have strong reasons to improve our ability to discover, diagnose, teach, repair and coordinate. We also have strong reasons to improve our ability to test powerful systems. Those efforts should advance together.

A society that can mobilize vast effort to build a model can mobilize serious effort to examine it. A government that can propose a moratorium can fund the people, equipment and institutions needed to judge particular risks. The responsible ambition is to increase both capability and the capacity to govern it.

The hard part is making that statement mean something precise.

Read the actual stop sign

In March 2023, an open letter hosted by the Future of Life Institute called for a pause of at least six months in training systems more powerful than GPT-4. It expressly distinguished that request from a pause on AI development in general. It also called for stronger safety work and governance.¹

Whatever one thinks of that proposal, an accurate criticism must preserve those boundaries. It was not a demand that every cancer-imaging system be switched off or every use of machine learning in science be abandoned.

PauseAI’s proposal dated April 5, 2026 also focuses on the most powerful general AI systems. It seeks a global pause until safety and democratic control can be assured through a proposed international agency. It says narrow systems, including cancer-image recognition, would usually be unaffected. Its longer-term discussion also considers restrictions on some advances in training methods and computing capability.²

These are advocacy proposals, not a description of laws already governing the world.

Their scope matters. A prohibition on using a particular model to make an unreviewable medical decision differs from a restriction on training a more capable general model. Both differ from a ban on all AI research. A person who objects to one has not thereby answered the case for the others.

My disagreement with broad frontier restrictions begins with their reach across future uses. A general system can contribute to many tasks. Restricting its development may therefore affect tools that researchers would otherwise use, even if a disease-specific project remains legal. The project may be allowed to continue with fewer options available to it.

That is a plausible cost, not a measured tally of missing cures. Its size depends on what the new systems would actually have added, how quickly those additions could have been used and what alternatives remained. Better results from older systems, specialized tools and non-AI methods may reduce the cost. New general capabilities may increase it.

The honest argument follows these links. It does not skip from a training restriction to a number of preventable deaths as if the intervening science had already been completed.

But the need for honest accounting cuts both ways. The fact that the loss cannot be known exactly does not make it zero. If a policy deliberately limits the production of potentially useful knowledge, its defenders should explain why that limit is expected to protect people better than narrower measures would.

A proposal cannot claim credit for every disaster it imagines preventing while dismissing every possible benefit it delays as mere speculation. The standards must travel in both directions.

Respect the aim and dispute the remedy

I respect Bernie Sanders and Elizabeth Warren for taking the interests of workers and ordinary people seriously. The question of who owns the gains from AI deserves the force they bring to it. My disagreement is with the remedy of broadly halting advanced development. A person can be right about the danger of concentrated power and wrong about the best way to confront it.

On September 16, 2026, Warren called for an immediate pause in advanced-AI development while lawmakers and regulators put safeguards in place. Her statement also opposed weakening antitrust protections. It did not define a numerical capability threshold or a fixed pause duration.³

On September 23, Sanders and Representative Greg Casar announced the proposed Ban Artificial Superintelligence Act. Its text would pause covered advanced-system training, modification and fine-tuning, with narrow exceptions, and block deployment of unreleased covered systems until a new department was staffed and had issued rules. It also contains separate prohibitions concerning superintelligence and specified precursor capabilities.⁴

That is a much wider instrument than stopping one release after a particular failure. It connects a large field of development to the creation and operation of a new federal institution. The proposal deserves to be read at that scale. It is proposed legislation, not an enacted restriction.

My objection is that the public needs more than protection from the owners of powerful tools. It needs greater power of its own. A broad development brake can leave existing owners with their lead while people outside the leading institutions continue waiting for useful capabilities. That is a risk of the design, not a prediction that every company will benefit from it in the same way.

The stronger worker-centered program is to build public research capacity, require credible testing, enforce competition, protect people from abusive uses and broaden their claims on the returns. We should confront concentrated ownership directly. We should not ask the person with little ownership to accept less technological possibility as their main form of protection.

This is also why safety institutions must be funded before their absence becomes the permanent excuse for delay. If a new department needs scientists, engineers and investigators, recruit them. If a public laboratory needs computing access, provide it. If a testing standard is missing, organize the work required to develop and challenge it. The urgency attached to building models should attach to building safeguards.

A legislature can require evidence before a dangerous release without treating the whole frontier as a single undifferentiated threat. Where a class of capability presents a serious hazard that cannot yet be contained, a restriction may be needed. Its defenders should show why the restriction is necessary and how useful work can continue around it. The burden is to design a protective boundary that fits the danger, not to celebrate the size of the boundary.

I believe history will judge harshly a politics that recognized extraordinary possibilities for human benefit and responded chiefly by freezing them. Sanders and Warren are asking whom the future will serve. I share that question. My answer is to build the future with public claims, public competence and enforceable duties from the start. Give people a stake strong enough to make progress their ally.

What would change during the pause

A temporary restriction can be sensible when it buys something identifiable.

An operator finds a serious fault. The affected function is stopped. Engineers investigate, repair the failure and test the repair. The work resumes only when the relevant evidence improves. The pause has a task.

A broader pause can be much harder to evaluate. What would happen during it that could not happen while less dangerous work continued? Who would do that work? What resources would they receive? What evidence would show that the aim had been met? Who could challenge a decision to prolong the pause?

These questions are not a trick designed to make every restriction impossible. They are the questions that turn a declaration into a policy.

If the purpose is to create independent testing capacity, fund that capacity now. If researchers lack access, create lawful routes for them to examine the systems. If an institution lacks authority to enforce a finding, define the authority and the appeal. If a particular type of release would make a danger impossible to recall, identify the danger and the release conditions.

Simply allowing six months to pass will not perform any of this work.

Nor does an exit condition become clear merely because it sounds reassuring. Safe enough has to be connected to a use, a level of exposure and a standard of evidence. A rule that requires broad agreement must say what kind of agreement is sufficient. Technical uncertainty and moral disagreement may persist even after substantial study. If a rule were to require unanimity, it could hand the most resistant participant a lasting veto.

That does not mean we should proceed whenever a few experts are willing to approve. It means the decision rule itself needs examination. Who is qualified to judge which part? What counts as a conflict of interest? How are minority concerns recorded? What happens when new evidence changes the answer?

A safety institution should be designed to learn, not only to prohibit. Its success should include the useful activity it helps make trustworthy. If its only visible achievement is the number of things it blocks, it will have a narrow view of its public purpose.

At the same time, an institution that can never say no has no meaningful power. Testing should be able to stop a release that fails the conditions set for it. A deadline should press people to do the work, not require them to approve an unsafe result when the clock expires.

The acceleration I defend is acceleration of the whole responsible effort. Faster inquiry. Better tests. More capable public oversight. Quicker correction of known problems. Wider access to demonstrated help. It does not require every stage to move at the speed chosen by the seller.

The world that continues

An argument for waiting needs a comparison with what will happen while we wait.

Sometimes the comparison is favorable. A reliable existing method may already serve people well. The new system may offer little benefit while introducing a serious risk. In that case there may be no urgent reason to replace the old method.

Sometimes the comparison is much less comfortable. Existing care may be difficult to reach. A public office may be overwhelmed. A researcher may be unable to examine more than a small fraction of the questions worth asking. A person may spend years excluded from work because the available tools do not meet their needs.

A new technology should not be compared only with perfection. It should be compared with the best feasible alternative, including changes that do not involve AI.

This point has practical force. Suppose a service can be improved by hiring two more people, simplifying the form or purchasing a new system. The AI purchase should not receive credit for benefits that the simpler reform would have achieved more reliably. Equally, a critic should not defeat the purchase by invoking an ideal fully staffed service that no one has funded and no one intends to build.

Put the actual alternatives on the table. Give each a cost, a timetable and an account of whom it helps. Where the numbers are uncertain, show the uncertainty. Where a right is at stake, name it rather than hiding it inside a price.

This is harder than announcing that we must act carefully. It is also more useful.

The moral urgency becomes particularly strong in research on cancer, Alzheimer’s disease, heart disease and genetic conditions. We have reason to want better understanding and better treatment. AI may help parts of that work. It would be wrong to promise that more computing power will produce a cure on demand. It would also be wrong to treat the chance of better research as an indulgence that can be set aside without serious thought.

The same applies to energy and water. Better forecasting, design, operation and maintenance may help people obtain reliable services with fewer resources. Whether a given tool does so is an empirical question. The aim is worth sustained investment before every answer is known.

Scientific uncertainty is often the reason to conduct research. If uncertainty alone becomes the reason to prevent it, the condition for resuming may never be reached.

There are exceptions. Some inquiries or releases may themselves create dangers that cannot be responsibly contained. Their value does not erase those dangers. But the exception should be explained by the mechanism of harm, not by a general suspicion of discovery.

A person who wants more progress should be willing to make these distinctions. They protect the case from its weakest exaggerations.

The energy bill belongs in the account

It is possible to make a sincere case for AI-assisted energy research while overlooking the energy used to run AI. That would be an incomplete case.

The International Energy Agency estimated that data centers used about 415 terawatt-hours of electricity in 2024, roughly 1.5 percent of global consumption. Its 2025 base-case projection put data-center demand near 945 terawatt-hours by 2030, with AI a major driver. Those figures cover data centers, not AI alone, and the 2030 number is a projection.⁵

The same report examines ways AI could improve energy systems and warns that potential savings depend on adoption, infrastructure and other conditions. A possible saving is not already available to cancel a present demand.

The right response is to build an honest energy account and act on it. Where is the electricity coming from? Does a new facility add pressure at the time and place the grid is most constrained? What investment will meet the demand? Who pays for the connection and the reinforcement? Can the work be scheduled more flexibly? What benefits does the public receive?

A global percentage may be small while a local burden is large. The community beside a facility does not live inside a world average. It lives with a particular grid, a particular water system and a particular public budget.

That is a reason to plan the expansion well. It is not a reason to give up on the technologies that might help make energy and water services better. We should invest in efficient computing, cleaner supply, capable grids and applications with demonstrated value. A policy that merely pushes demand elsewhere may make its own accounting look cleaner without reducing the burden.

Again, the choice has more than two positions. A community can welcome investment and demand that the developer pay appropriate costs. A researcher can favor stronger AI and insist that wasteful work be reduced. A government can speed up necessary infrastructure while preserving scrutiny of its effects.

The courage to build should include the courage to show the bill.

A trial is a decision too

Between unrestricted deployment and a broad halt lies a large field of choices. It is where much of the serious work belongs.

A tool can be tested with a smaller group. Its authority can be limited. Sensitive uses can require stronger evidence. A system can suggest an action without being allowed to carry it out. Some capabilities can remain available to qualified researchers while public release waits. A service can be expanded gradually as its performance becomes clearer.

These options are not interchangeable, and none is a universal solution. A limited trial can still cause serious harm if the thing being tried is dangerous. Keeping a human nominally involved may add little protection if the person lacks time, information or authority. A restriction can be easy to state and hard to enforce.

The value of a staged approach is that it can make the next decision better informed while limiting exposure where that is genuinely possible.

Imagine a public agency considering an assistant for staff who answer routine questions. Before purchase, it identifies the questions for which a wrong answer would be especially harmful. Those are excluded from the first trial or routed to trained staff. The trial records accuracy, time, user experience and the work needed to correct errors. It also records the people for whom the service performs poorly.

A successful trial would support a particular expansion under particular conditions. It would not justify giving the system authority over every decision in the agency. A failed trial would support repair or rejection of that use. It would not prove that all AI assistance should stop.

This sounds modest. It can be a rapid way to accumulate trustworthy knowledge if enough institutions are equipped to do it well and share what they learn.

The choice to conduct such a trial also has an opportunity cost. Staff time spent evaluating one tool cannot be spent on another improvement. That is why the question should be tied to a need. Testing every fashionable product can become its own form of waste. A serious program begins with the problem it intends to solve.

It also defines a stopping condition before enthusiasm makes stopping difficult. What error would end the trial? What result would justify expansion? What evidence would show that another approach serves people better? The people responsible for answering those questions should have the resources and authority to act on the answers.

Speed without that authority can produce a system that everyone is evaluating and no one can stop.

Refuse the lazy choice

The strongest argument for a pause is that some harms could be so large, and the chance to recover so small, that learning by ordinary deployment would be unacceptable. That argument deserves a serious answer. It should change how we handle capabilities that could create such exposure.

It does not establish that every path to greater capability must be blocked until every possible future is understood.

To justify that wider conclusion, one must show why narrower controls are insufficient, what the pause can realistically achieve and how its own costs and concentrations of power will be managed. A promise of safety cannot exempt the institution making it from scrutiny.

Likewise, a promise of abundance cannot exempt a developer from proving that a release meets the standards appropriate to its use. The public owes neither side a blank authorization.

My preferred program is demanding because it refuses the cheap form of certainty. Keep developing useful AI. Expand scientific research. Fund independent testing at a scale that matches the ambition of the builders. Give public institutions real technical capacity. Restrict dangerous actions where the evidence warrants it. Make the conditions for changing those restrictions visible. Invest in the means by which benefits reach people.

This program will not eliminate disagreement. It will give disagreement something better to work on than a single imagined switch.

A broad pause is reckless when it substitutes the stillness of a sector for the safety of a society. People remain exposed to old dangers, institutions remain weak, and the work of building a better alternative may remain undone. We should judge the policy by what it changes, not by the comfort of having said stop.

The future can be made worse by careless action. It can also be made smaller by a failure of nerve disguised as responsibility.

We have work to do on both fronts. Let us fund it, test it and get on with it.

Sources and notes