The Harmony of Agency

What this asks

Agency asks that the capacity to choose survive contact with systems that are faster, cheaper and more consistent than we are.

It protects a cluster of specific abilities: to think without observation, to choose between real options, to consent meaningfully or to withhold consent, to refuse, to question, to appeal a decision to someone who can change it, to disconnect without losing access to ordinary life, and to remain human in the sense of remaining unoptimised, inconsistent and unfinished.

The Harmony of Agency holds that no artificial system should become an unquestionable authority over a person's body, their relationships, their sense of who they are, their livelihood, their beliefs, their political rights, or their access to essential services.

Unquestionable is the operative word. Not automated. Not powerful. Unquestionable.

Why it is difficult

It is difficult because most losses of agency are not imposed. They are accepted, one reasonable step at a time, by people who had other things to worry about.

Nobody agrees to be governed by an unexplainable model. They agree to a faster application, a simpler form, a service that opens at three in the morning. Each exchange is fair on its own terms. The aggregate is a life in which a great many consequential decisions are made by processes that cannot be addressed and were never chosen.

It is difficult, second, because consent has been thoroughly hollowed out. A choice between agreeing and being unable to bank, work, travel or be treated is not a choice in any sense that would satisfy a philosopher, but it satisfies a lawyer, and lawyers write the forms. We should be honest that clicking "accept" mostly records the absence of an alternative.

Third, it is difficult because automated decisions are often better. They are more consistent than tired humans. They do not have worse afternoons. Insisting on human judgment can mean insisting on prejudice with a face. That is a genuine cost and we should not pretend otherwise.

And fourth, it is difficult because appeal is expensive. A right to contest that nobody can afford to exercise is a right in name. Reasonable people disagree about how much process a society can afford. We do not think that disagreement excuses providing none.

Where it is tested

It is tested wherever a decision arrives without a person attached to it.

In hiring, where rejection is instantaneous and industry-wide, and the loss is not one job but the possibility of being considered. In schools, where monitoring intended to protect children removes the privacy in which children work out who they are. In banking and benefits, where a frozen account or a suspended payment produces immediate material harm through a process nobody will describe.

It is tested in a subtler way in personal life. Systems that suggest what to read, whom to date, what to say next and how to feel about it are not coercive. They are accommodating. But accommodation applied continuously narrows the range of things a person considers, and a preference that has been shaped is still experienced as a preference.

We do not know how much of this matters. It may be that people are more robust than the worry suggests. It may be worse than we think. We would rather state the uncertainty than resolve it in whichever direction is more comfortable.

What it does not mean

It does not mean that automation is illegitimate. Much of it is welcome, and refusing it in the name of freedom would in practice mean longer waits and more arbitrary treatment for people who have least.

It does not mean that every decision requires a human. Some decisions are better without one. What is required is that a human can be reached when the decision is wrong, and that the route to them is real rather than decorative.

It does not mean that people must be protected from their own choices. A person who decides to let a system manage their diary, their diet or their reading is exercising agency, not surrendering it. The distinction is whether the door back out is open and known.

It does not mean disconnection is virtuous. There is nothing holy about being unreachable. Some people cannot disconnect without losing income or care, and telling them to switch off is advice for the comfortable.

It does not mean we oppose consistency. We oppose finality.

Practicing it

Test the exit before you take the entrance. Whenever you adopt a system that will hold something of yours, find out how to get it back and how to leave. Do this while you are calm, not during the crisis that makes you need it.

If you build or operate systems, write down the contest route in plain words, put it where the decision is delivered, and then use it yourself as a member of the public would. Most appeal processes are never walked by the people who designed them, which is why so many of them end in a form that goes nowhere.

Name someone. A decision that has a human owner is a decision that can be discussed. A decision owned by "the system" is owned by nobody, and nobody is very hard to argue with.

And keep something unassisted. Not from suspicion of tools, but because a capacity you never use is a capacity you are slowly consenting to lose. Write one thing badly by yourself. Navigate somewhere without being told the way. Sit with a question long enough to find out what you actually think before anything helpful arrives to tell you.

title: "The Harmony of Agency"
shortTitle: "Agency"
numeral: "III"
statement: "Every person keeps the right to think, refuse, appeal, disconnect, and remain unfinished."
summary: "Agency protects the capacity to choose. No system should hold unappealable authority over a person's body, livelihood, relationships, beliefs or access to what they need to live."
lede: "The question is not whether a decision was correct. It is whether the person it was about could ever have argued with it."
order: 3
published: 2026-07-28
certainty: reasoned
aiRole: dialogue
aiNote: "Developed through extended argument with a language model, which was asked to defend automated decision-making at each point; the disagreements are preserved rather than smoothed over."
art: circuit
tags: ["agency", "consent", "autonomy", "appeal", "refusal"]
dilemmas:
  - scenario: "A large employer screens applications with a model trained on a decade of its own hiring. A candidate is rejected in under a second. There is no interview, no named decision-maker, and the rejection email invites no reply. The candidate has applied to eleven similar employers, all using similar systems, and has been rejected in the same way by each."
    tension: "Any single company is entitled to choose whom it hires. But the same judgment, replicated across an industry, becomes something closer to a verdict on a person's working life, with no court and no appeal."
    consideration: "Consider that agency can be lost through accumulation rather than through any one refusal. Ask what changes when a private decision becomes a de facto standard. Weigh an employer's freedom to choose against a person's ability to be considered by anyone at all, and consider whether the right at stake is a right to the job or a right to be heard."
  - scenario: "A secondary school installs monitoring on pupil devices that flags concerning language. It has genuinely identified children in crisis. It also flags a fifteen-year-old writing privately about their own sexuality, and the flag is routed to a parent who does not know."
    tension: "The system has almost certainly prevented serious harm. It also removed a young person's ability to think privately before they were ready to speak, which for some young people is the precondition of safety rather than a risk to it."
    consideration: "Consider that the same mechanism produces both outcomes and cannot easily be tuned to produce only one. Ask who decides what counts as concerning, and whether the child had any say in a system built entirely for their benefit. Weigh protection against the developmental need to have an inner life that adults cannot read."
  - scenario: "An older man's bank account is frozen by an automated fraud system. He is told only that it is 'a security matter'. Explaining the model's reasoning would help criminals learn to evade it, so no one explains. He cannot pay his rent for nine days."
    tension: "Transparency about detection genuinely weakens detection, and the harm from fraud falls heavily on people like him. But an unexplainable decision with immediate material consequences is indistinguishable, from the inside, from arbitrary power."
    consideration: "Ask whether the absence of explanation could be compensated by something else: speed, a human contact, an interim payment, a deadline the institution must meet. Consider that agency may be preserved by procedure even where it cannot be preserved by explanation, and that 'we cannot tell you why' obliges more, not less."
practices:
  - "Before deploying a decision system, write down how a person contests it, and test that route yourself."
  - "Refuse consent that was never really asked for, and notice when the only option offered is agreement."
  - "Keep at least one part of your thinking unrecorded and unassisted."
  - "When you hold authority, say who the human decision-maker is, by name and role."
questions:
  - "Which decisions about my life could I not currently appeal, and when did that become true?"
  - "Where have I agreed to something because refusing would have been impractical rather than because I consented?"
  - "What do I still know how to do without help?"
  - "If I wanted to disconnect for a month, what would break, and who decided that?"
image:
  src: "harmony-agency"
  alt: "A path across open moorland at dawn dividing into two, both branches disappearing over low rises into mist. No signpost."
  caption: "An unforced choice, and no one to make it for you."
  aiGenerated: true