AI ethics often gets framed as a search for the right rules: prevent bias, protect privacy, explain decisions, keep humans in control.
Those are good goals. But they can sound cleaner than the real problem is.
The deeper question is this: When AI enters a decision, what happens to the person involved?
Can they understand what is happening? Can they disagree? Can they choose another path? Can they reach someone who has the authority to fix a mistake?
That is where human agency lives. Not in mission statements, but in the small, practical ability to act.
A human in the loop may still be powerless
Companies often say that an AI system does not make the final decision. A person does.
This sounds reassuring. Sometimes it is. Often it is not.
Imagine a loan officer reviewing an application. The system recommends rejection and assigns the applicant a high-risk score. The officer can technically override it, but doing so requires extra paperwork and may trigger a review. The officer has ten minutes per case and knows that management tracks agreement with the model.
There is a human in the loop. But the loop has been designed to produce compliance.
The same thing happens in hiring, insurance, education, and public services. A model offers a recommendation. The recommendation arrives with numbers, confidence scores, and the quiet authority of computation. The person reviewing it may not understand how it was produced. Even when they have doubts, the safest career move is often to accept it.
So the right question is not, “Was a human present?”
It is, “Did that human have enough knowledge, time, authority, and protection to make an independent judgment?”
That is a much harder standard. It is also a more honest one.
Convenience can reduce choice
Many AI systems are built to remove friction. They draft the email, choose the route, rank the applicants, suggest the diagnosis, summarize the meeting, and decide what appears next on the screen.
Removing friction can be useful. I use these systems too.
But friction is not always waste. Sometimes it is the moment when we think.
Writing a difficult email forces us to decide what we actually mean. Comparing candidates forces us to question our assumptions. Reading a full argument exposes us to details that a summary may flatten. Choosing a route helps us build a mental map.
When those steps disappear, the task becomes easier. It may also become less ours.
The concern is not that people will stop thinking altogether. That claim is too dramatic. The concern is quieter: we may begin to accept the first reasonable answer because checking it takes more effort than producing it did.
AI changes the economics of judgment. Generating an answer becomes cheap.
Evaluating it remains expensive.
That gap matters.
Agency requires a real way to say no
Consent is often treated as a box to check. A user clicks “agree,” and the system moves on.
But consent means very little when refusal is impractical.
A worker may be told that an AI system will monitor productivity. They can object, perhaps, but not without risking their job. A student may be evaluated by automated software with no alternative process. A patient may be offered an AI-assisted service without being told how much of the decision is automated. A person may opt out of data collection only by giving up a service they need.
These are not free choices in the ordinary sense.
A useful test is simple: What happens to the person who refuses?
Do they receive a comparable service? Are they delayed, penalized, or treated as suspicious? Is the alternative hidden behind a support form nobody answers?
Agency depends on viable alternatives. A theoretical exit is not enough.
Explanations should help people act
There is a lot of interest in explainable AI. Much of that work is valuable, but explanation can easily become theater.
A system may say that a decision was based on “employment history, income stability, and account activity.” That sounds informative. It may still leave the affected person unable to do anything.
Which part of the employment history mattered? Was the data correct? What would change the result? Who can review the case? How long will the information be stored? Can the person appeal?
A good explanation is not merely a description of the model. It gives the person a path forward.
That changes how we should judge explanations. The goal is not to make the system sound understandable. The goal is to help someone challenge, correct, or respond to a decision.
Sometimes that requires technical detail. Often it requires ordinary detail: what happened, why it mattered, and what can be done next.
Responsibility tends to disappear inside systems
When an automated decision harms someone, responsibility can become strangely difficult to locate.
The developer says the model only produced a score. The vendor says the customer chose how to use it. The manager says the employee made the final call. The employee says the system was mandatory. Everyone played a part. No one appears to own the outcome.
This is not just a legal problem. It is a design problem.
A responsible system needs named people with the power to intervene. Someone must be able to pause the system, investigate patterns of harm, correct records, compensate affected users, and explain what changed.
Without that, accountability becomes a chain of forwarded emails.
We should be suspicious of any system that can affect a person’s life but has no clear owner when something goes wrong.
Personal AI raises a different set of questions
Not all questions about agency involve institutions making decisions about people. Some involve people using AI for themselves.
An assistant can help someone write, plan, study, code, or organize a difficult day. That can expand agency. It can give people abilities they did not have before, especially when time, money, language, or disability creates barriers.
But the relationship can shift.
A tool that begins by helping us express our preferences may gradually start shaping them. It suggests what to buy, what to watch, how to respond, which goals are realistic, and what kind of person we should try to become.
This influence does not need to be manipulative to matter. Repetition is enough. Defaults are enough. A confident tone is often enough.
The design question is not whether an assistant influences the user. Every useful tool does. The question is whether that influence leaves the user more capable of reflection and choice, or more dependent on the system’s next suggestion.
A good assistant should sometimes slow down. It should make uncertainty visible. It should distinguish between helping someone carry out a decision and quietly making the decision for them.
Children deserve special care
Adults already struggle to judge when an AI system is reliable, persuasive, or merely fluent. Children have an even harder task.
A conversational system can feel patient, attentive, and emotionally available. Those qualities can be helpful. They can also create trust that the system has not earned.
Children may not understand that the system has no lived experience, no duty of care, and no stable relationship with them. They may not recognize commercial incentives or know when advice should be taken to an adult.
This does not mean children should never use AI. It means the burden should not be placed on them to manage every risk.
Systems used by children should be designed around age, dependency, privacy, and emotional vulnerability. Parents and teachers need meaningful controls, but children also need dignity. Constant surveillance is not the same as protection.
The balance is difficult. Pretending otherwise does not help.
Ethical design often looks ordinary
People sometimes expect ethical AI to involve grand principles or advanced technical methods. In practice, many of the most useful safeguards are fairly plain.
Give people time to review a decision. Show the source of important claims. Make uncertainty visible. Keep records of overrides. Provide an appeal process that a real person reads. Test whether some groups are harmed more often than others.
Do not collect data simply because storage is cheap. Do not automate a decision when the cost of error is high and the benefit is minor.
None of this is glamorous.
It is closer to maintenance than invention. That may be why organizations neglect it. New features attract attention. A well-run appeals process does not.
But ethics lives in those details.
The point is not to preserve every old way of doing things
Human agency does not mean humans must perform every task by hand. Nor does it mean every automated decision is wrong.
People have always used tools to extend their abilities and hand off work. We rely on calculators, maps, search engines, medical devices, and countless systems we do not fully understand.
The issue is not delegation itself. It is delegation without awareness, recourse, or control.
A person should be able to know when a consequential decision has been delegated. They should understand the limits of the system well enough to judge when caution is needed. And they should have some practical way to intervene when the result affects them.
That is a demanding standard. It should be.
Efficiency is easy to measure. Agency is not. But we can still look for it.
Can people question the system without being punished? Can they correct bad information? Can workers override recommendations without risking their careers? Can users leave without losing access to essential services? Does someone take responsibility when the system fails?
Those questions bring AI ethics down to earth.
The goal is not to keep humans ceremonially present while machines and institutions make the real choices. The goal is to build systems that leave people with a meaningful role in their own lives.
That role does not need to be absolute.
It does need to be real.
