What Must We Be Free to Choose?

A choice we can exercise

Cooperation can make a capability available that no participant could sustain alone. The previous chapter considered that possibility for governing AI. It left a question about the dependence cooperation creates: can people and institutions still change course when their ability to function rests on an arrangement they share?

Meaningful choice requires more than permission to choose. Someone must be able to exercise the choice, with the knowledge and means it needs. A public institution may have a right to replace a system while lacking a workable way to keep its service available during the change. Its freedom remains important, but part of the capacity needed to use it is missing.

Consider a further extension of the hypothetical public service discussed in earlier chapters. Suppose its supplier's integrated AI assistance has become important to how the service operates. Staff use it to organize work and help people obtain access.

The institution retains a formal right to change providers. When it considers doing so, it discovers that a replacement would require expertise, time, and continuity arrangements it cannot immediately supply.

Suppose earlier integration choices have also shaped staff expertise and the systems used by partner organizations. Each further connection increases the work needed to leave, while everyday use deepens proficiency with the current arrangement. The original choice now influences which alternatives are practicable. Understanding that accumulation matters before replacement becomes urgent, without making useful dependence itself a wrongdoing.

The current supplier need not have done anything wrong. The integration may provide valuable help. The difficulty is that changing providers before continuity arrangements are ready could interrupt the service. People who depend on the service could bear consequences they did not choose. This remains a hypothetical situation, illustrating a possible relationship between useful reliance and constrained choice.

Simply exercising the right to leave could make matters worse. Simply remaining could deepen a dependence the institution cannot adequately govern. The responsible question is how to preserve useful service while making a change of course practicable. That may require investment, a more limited reliance, or cooperation with others. None is established as a successful remedy by the illustration.

This chapter asks what practical agency should survive our reliance on AI. We will consider whose choices matter, which dependencies can support them, and whether essential oversight can continue when its own tools must change. The final question concerns what today's arrangements leave others able to choose.

We value meaningful choice over operational convenience.

Convenience deserves its place in that comparison. Integration can reduce repeated work and make complex assistance easier to use. For someone seeking help, fewer transfers between systems can mean less time spent explaining the same need. For a small institution, shared capability may make a service possible at all. Preserving meaningful choice should protect these benefits where justified, while keeping their terms open to consequential revision.

Whose freedom matters

An institution's choice of supplier and a person's choice about treatment are different matters. Officials might replace an AI system while leaving the same policy, inaccessible process, or mistaken decision in place. The change would improve one institutional option without necessarily restoring the affected person's agency. We should ask what each freedom makes possible and for whom.

Someone seeking an essential public service may have no reasonable substitute for it. Telling them to use another provider would leave the institution's responsibility unresolved. Their agency may depend on intelligible reasons, usable review, and a way to obtain appropriate help when the usual route fails. Institutions should carry the work needed to make those protections effective, rather than make protection conditional on a person's ability to leave.

Meaningful choice therefore includes influence over the terms of reliance. Can an institution examine an important change, seek assurance beyond its supplier's account, and require a justified limit? Can a person question a consequential choice without mastering the technology behind it? Replacement matters partly because a practicable alternative can give a refusal force. But changing the terms or correcting treatment may be the relevant response, even when the system remains useful.

At the scale of societies, this is the emphasis the book gives to AI sovereignty: the practical ability of people and societies to choose, govern, challenge, and replace the AI systems on which they depend. The term describes human and institutional self-determination. It gives no sovereignty to an AI system and establishes no legal doctrine. Its value here is to direct attention to capacities that claims of independence can leave unexplained.

A government could gain independence from a foreign supplier while imposing an arrangement its population cannot meaningfully question. Domestic ownership would change the source of supply; it would not by itself justify the power exercised through the system. People still need protection against arbitrary state action. Nor should the ability of a powerful organization to choose its preferred technology be confused with freedom for everyone affected by that choice.

Sovereignty understood this way also has obligations. One institution's freedom to change course cannot erase duties to people who depend on it or consequences imposed elsewhere. Common rules can give participants dependable terms on which to cooperate. Binding commitments may be necessary for a shared capability to work. Meaningful choice does not require a universal veto, exemption from valid obligations, or a costless service tailored to every preference.

The useful expectation is that consequential terms remain answerable and appropriate alternatives remain practicable. A person can disagree with a decision without acquiring the authority to decide it. An institution can enter a durable commitment without surrendering every means of challenging its operation. What matters is whether the arrangement preserves the capacities needed for the responsibilities and interests at stake.

Dependence on workable terms

Dependence is not in itself a failure of sovereignty. An institution relying on shared infrastructure may gain expertise and practical options it could not otherwise afford. A society cooperating with others may become better able to examine evidence or sustain an essential function. Refusing that reliance could leave it with less capacity to govern, including less capacity to make an independent judgment.

The question is which dependence is acceptable and what must accompany it. In the public-service illustration, integrated assistance may be worth retaining. The institution nevertheless needs enough understanding to judge what it is entrusting to the supplier and what a consequential change would involve. If all relevant expertise sits outside the institution, it may struggle to recognize when reliance needs reconsideration or to assess a proposed alternative.

Retained expertise does not mean reproducing every technical skill internally. It means having access to competent judgment that can serve the institution's duty, including when that judgment is inconvenient to the supplier.

Evidence access matters for the same reason. A choice among providers has limited governing value if each account can be examined only on the provider's terms. Appropriate access must preserve privacy and security while allowing material assurances to be questioned.

Selective domestic capability can support practical agency where a consequential function cannot otherwise be sustained or independently examined. Its justification should identify the capacity being preserved. Owning every component would not automatically supply reliable operation, relevant expertise, or effective scrutiny. The location of a system matters where it changes these conditions; ownership alone cannot stand in for examining them.

Shared infrastructure offers a different possibility. Several institutions could support expertise or an alternative method that none could maintain alone. That cooperation would create its own dependence, but could also make replacement or independent inquiry more workable. The important comparison is between feasible arrangements and the choices they preserve, including their costs. Technological isolation is neither a prerequisite for agency nor a guarantee of it.

Alternatives need examination too. A second provider might depend on the same essential infrastructure as the first. If the constraint lies in that shared infrastructure, changing providers would leave it unchanged. A provider might also offer a similar service while lacking the evidence access or capacity the institution needs; that alternative would leave those needs unmet. These are possibilities to investigate, not reasons to presume every alternative inadequate. Counting suppliers would answer less than asking what can actually be sustained or changed.

An alternative need not duplicate the existing service in every respect. A simpler arrangement might preserve the essential function while offering less convenience. In the hypothetical service, a temporary route could preserve access while requiring more staff effort or longer waits. That would be a real cost. It might still be preferable to a change that makes access unavailable, but its adequacy would need examination rather than assurance by name.

The strongest objection is the expense. Maintaining expertise and a practicable alternative uses resources that could improve the current service. A small institution may be unable to sustain two comparable arrangements.

People with urgent needs may reasonably question why help is delayed to protect a future option. The argument for choice must answer them, rather than treat every investment in continuity as self-justifying.

Proportion matters. Preserving the ability to challenge a policy may require different resources from replacing the system that administers it. A replaceable analytical aid may warrant less preparation than a dependence whose failure would interrupt essential access.

The institution should explain which function matters, what consequence it seeks to avoid, and who bears the preparation cost. Preserving every feature at all times would be an extravagant standard; abandoning essential protection to save effort would be inadequate.

Where a workable alternative is not yet available, the limit should affect reliance. The institution may need to retain a narrower use while developing the capacity to change it. Some useful integration may have to wait. In other circumstances, shared assistance may make justified reliance possible sooner. The decision must respect applicable duties and the stakes, with neither a promise of future independence nor an urgent benefit supplying missing grounds on its own.

Oversight when tools change

The same reasoning applies to the means by which institutions govern. Chapter 6 proposed computational assistance for connecting evidence, examining interactions, and carrying out authorized work. Chapter 8 extended that capacity across institutions. If essential scrutiny comes to depend on a shared tool, changing or losing that tool becomes a governing event. The activity being overseen and the means of overseeing it need to be considered together.

Early procurement and standards choices can shape this independence. A reviewing method adopted for convenience may become the basis of training, shared records, and acceptable evidence. Later alternatives must then replace those connections as well as the tool. Preserving room to question and change the method is part of governing its adoption.

Suppose a shared analytical method can no longer be relied on. Its provider might withdraw support, or examination might reveal a material weakness in its findings. The question is not simply whether another tool can be purchased. Can responsible institutions continue to examine the evidence and keep consequential activity within justified limits during the change? An earlier permission may have depended on capacity the new arrangement no longer supplies.

Preserving records helps, but records do not interpret themselves. An institution may retain the history of an authorization while lacking a workable method for examining how later changes affect its conditions. Reasons and supporting evidence should remain accessible to appropriate scrutiny, and people need the competence, resources, and methods to use them. A readable archive would otherwise preserve an account that nobody can adequately assess in time.

Nor is a human fallback adequate merely because someone is assigned to it. If the interactions exceed what reviewers can responsibly examine unaided, removing computational help can leave the duty beyond their capacity. Accountability to people remains essential. Fulfilling it may require another demonstrated method or a narrower activity, rather than a signature attached to work the reviewer cannot meaningfully perform.

Oversight must survive its tools

A replacement need not reproduce every feature of the old tool. It must support adequate scrutiny of the activity that remains permitted. Preserving that duty may require less activity during the transition. The capacity to keep governing matters alongside the capacity to restore the technology.

That directs attention to the essential work. Which judgments can still be supported, which constraints remain effective, and which permissions depend on the missing contribution?

A different analytical method or shared expertise could help. A simpler operation could reduce the demands on scrutiny. Where adequate oversight cannot be maintained, the affected activity must be narrowed to what can actually be overseen, or paused if no adequate scope remains. The response should follow the missing capacity and consequences, rather than impose one rule for every interruption.

Replacement itself needs justified reliance. A tool suitable for locating records may not be suitable for evaluating their significance. Evidence that an alternative performs well on separate decisions does not establish that it can examine their interactions. Independent inquiry should be able to question the proposed substitute, including whether it changes what gets counted as a material concern. Continuity must preserve the duty, not merely the appearance of uninterrupted review.

The Governance Labs proposed in Chapter 8 could investigate whether shared governing capacity remains usable when a particular method is unavailable. Such an exercise would need to examine the whole arrangement, including its powers and resources.

A successful exercise would supply evidence within its conditions, leaving performance in consequential operation to be established. This is one possible purpose for joint inquiry, not an additional mandate or an achieved result.

Practical independence can also require the ability to pursue an inquiry outside the shared method. Replacing a supplier would accomplish little if the same unexamined assumptions continued to determine every question. The aim is room for competent scrutiny that can reach a different conclusion. It need not require every institution to work alone or disclose protected information to everyone.

These capacities deserve deliberate development alongside the governing tools. Institutions should not discover only at the moment of failure that their duty has become impossible to discharge without one service. Investing in continuity can make useful assistance less precarious. Its success would be the ability to sustain appropriate scrutiny and respond within legitimate limits, with the costs and remaining constraints made clear.

What remains possible

The remaining question is what recovery and future choice mean for people. A system can return to operation while access remains interrupted, a mistaken decision remains in force, or someone still bears a loss. Restoring the technology and repairing its consequences are different responsibilities. We should judge recovery against the function and protection people need, rather than the moment a tool becomes available again.

Return to the hypothetical public service. A successful provider change, if one became possible, would not by itself resolve the earlier questions about access or treatment. The institution would still need to examine consequential decisions and respond within its authority. People should be able to obtain help through the transition without first understanding why the institution chose its supplier or how its systems fit together.

Some consequences cannot be undone. Replacing a system may not restore a missed opportunity or undo the loss of privacy caused by disclosing private information. The capacity to recover therefore has limits, and those limits can justify protection before exposure. Preserving a future choice sometimes means refusing a dependence whose consequences would leave no responsible route back. Other dependencies may be justified despite incomplete reversibility. The reasons must address the actual interests and available alternatives.

We value lasting benefit over immediate gains.

Immediate help matters. A useful service can relieve hardship now, and postponing it can impose a serious burden. Lasting benefit asks whether that help can remain dependable and whether its terms leave future people able to act. The preference cannot justify sacrificing urgent needs to an unspecified future. It calls for reasons about durability, distribution, and the options today's decisions preserve or close.

Those reasons must also identify who carries the cost. An institution might protect its own continuity by transferring extra work to applicants or leaving smaller partners without practical alternatives. Its resilience would then coexist with diminished agency elsewhere.

Shared investment can be worth pursuing when it makes essential capacity available to those who could not sustain it alone. The burden remains part of the judgment even when the proposed arrangement is useful.

We should expect institutions to explain what they depend on, which consequential choices remain practicable, and how essential functions and protections could survive a change of course. This expectation supports cooperation whose terms people can govern. It gives deliberate development of governing capacity a lasting purpose: useful assistance can expand what we accomplish while leaving us able to reconsider how we accomplish it.

That capacity rests on beliefs about what will continue to work. A supplier will remain available; an alternative will be adequate; evidence will remain examinable; someone will notice a consequential change. Before relying on those beliefs, we need to ask what supports them and what could make us reconsider. The freedom to change course depends partly on recognizing when the grounds for staying have changed.