What should we still choose to do ourselves?

Imagine two messages with the same words. In the first, a person has struggled to compose an apology. In the second, a system has produced it in seconds, and the sender has barely read it. To the recipient, the sentences may initially look identical. Yet the difference matters if the apology is supposed to express recognition, commitment and a willingness to change. Producing the right text and doing the work of apologizing are related achievements. They are not interchangeable.

Now imagine someone who knows exactly what they want to apologize for but finds writing exhausting, or cannot easily communicate in the recipient's language. An assistant helps them say what they mean. Here, delegation may make an act of responsibility possible. Requiring unaided composition could obstruct the very human connection we hoped to protect.

These hypothetical cases expose a weakness in the familiar question of whether AI makes life too easy. Difficulty alone cannot tell us what is valuable. Some difficulties develop capacities we care about; others simply prevent people from participating. Nor can efficiency settle the matter. A faster process may free attention for something important, or remove the occasion on which that attention was exercised.

The question is therefore which relationships to an activity we want to preserve. My proposal is this: we should protect our effective ability to question, redirect and answer for what is done in our name. Sometimes that requires doing the activity ourselves. Sometimes it requires assistance. The dividing line depends on what is at stake, who bears the consequences and whether delegation strengthens or weakens our future choices.

Delegation changes more than the workload

Delegation means assigning part of an activity to someone or something else while retaining some relationship to its purpose and result. We already depend on other people, institutions and tools. Independence from every form of assistance would be an impossible standard for a human life.

Autonomy, as I use the term here, is a more practical capacity: being able to understand enough about one's options to make meaningful choices, revise them and contest decisions that affect one's life. A person can gain autonomy through dependence on a reliable tool. They can also lose it while remaining nominally in charge.

A button labeled "approve" does little if its user lacks the knowledge, time or authority to refuse. Conversely, a person who delegates most of the mechanical work may exercise substantial control over the purpose, evidence and final decision. Counting the proportion of text written by a human tells us little about that difference.

This broader moral frame has an institutional precedent. In its 2022 announcement of the Philosophy, AI and Society Consortium, Oxford described an agenda connecting AI with human flourishing, social justice and democracy. That announcement is a statement of research purpose, rather than evidence about particular technologies. It nevertheless identifies the scale of the question: evaluating AI involves asking what arrangements of human life are worth building.

At that scale, delegation creates at least three distinct concerns. An activity may be how we acquire a capacity. It may be how we discharge a responsibility. Or it may be how we attend to another person. These concerns overlap, but they call for different responses. Protecting a learner's opportunity to practice does not require imposing the same limits on an experienced worker completing routine administration.

(This is the question I want RingRing to keep asking: when we gain convenience, what else changes?)

The work that teaches us to judge

Consider a student preparing an argument. If the immediate goal is a polished essay, asking a system to produce one can seem successful. If the goal is learning how to distinguish an assertion from a reason, the same procedure needs a different evaluation. The finished artifact cannot by itself show whether that learning occurred.

This does not mean every preliminary task deserves preservation. Searching a poorly organized archive or fixing a document's formatting may consume energy without advancing the intended lesson. A useful assistant could remove those obstacles and leave more time for argument. It could also offer objections that a learner then has to answer. The educational question concerns the intellectual activity that remains.

There is evidence against assuming that assistance necessarily prevents learning. In Generative AI at Work, Erik Brynjolfsson, Danielle Li and Lindsey Raymond study an assistant introduced to 5,172 customer-support agents. The abstract currently hosted by Stanford reports a 15 percent average increase in issues resolved per hour, with differing effects across workers, and evidence of worker learning. These findings concern a particular workplace and task. They do not establish what happens when students outsource essays, or when assistance continues across years of professional development.

The opposite blanket claim—that AI inevitably makes people worse thinkers—is also unsupported by the evidence considered here. A 2025 survey by Hao-Ping Lee and colleagues collected 936 examples from 319 knowledge workers. Higher confidence in generative AI was associated with less reported critical thinking; higher confidence in one's own abilities was associated with more. Respondents also described a shift toward checking and integrating outputs. Because these are self-reports, the study does not demonstrate that AI caused a lasting decline in cognitive ability.

Together, these studies leave room for a more exact concern: assistance may improve today's performance while changing the conditions under which tomorrow's judgment develops. That is a question to investigate, rather than an outcome to assume.

For education, a defensible response is to make the learning purpose explicit. A teacher might ask students to defend an argument orally, explain a rejected alternative or revise a draft after finding contrary evidence. Such proposals need evaluation in their own settings. Their purpose is to make understanding visible, including when AI has helped produce the work. UNESCO's 2023 guidance similarly calls for a human-centered approach to the ethical validation and pedagogical design of generative AI. It provides policy guidance, not a demonstration that any single teaching method works.

Responsibility requires room to disagree

A second concern arises when someone's output affects somebody else. Consider an organization using an assistant to prepare a response to a complaint. A staff member could read it carefully and still lack permission to change the proposed remedy. Another might have authority but no time to examine the underlying record. Neither arrangement provides much reason to trust the promise of human review.

Meaningful review requires a connection between understanding and power. The reviewer needs access to relevant evidence, an opportunity to notice what the system has omitted and a credible route for acting on disagreement. Otherwise, responsibility becomes an assignment of blame after the event rather than a capacity to influence what happens.

This is a normative argument about institutional design. It does not depend on pretending that human judgment is consistently excellent. Human reviewers can be hurried, prejudiced or mistaken. An automated process might outperform them on a well-defined measure. The comparison still leaves open who chose the measure, whose circumstances it overlooks and how an affected person can challenge the result.

A useful example is a system that ranks requests by urgency. Even perfect consistency in applying its ranking rule would not tell us whether that rule adequately recognizes an unusual case. The person responsible for the service should be able to discuss the rule and its limits, rather than point only to the accuracy of its implementation.

The capacity worth preserving here is the ability to give reasons and respond to objections. An organization can delegate the preparation of information without delegating away this obligation. But retaining it costs time and resources. If automation is introduced alongside staffing arrangements that make examination impossible, a promise of oversight becomes difficult to take seriously.

We should therefore ask what people are actually allowed to do when they disagree with a system. A workflow that welcomes correction expresses a different understanding of responsibility from one that treats every interruption as a failure of efficiency.

When the activity is part of the relationship

The apology returns us to a third concern. Some activities have value partly because of the attention they embody. Listening to a friend, explaining a difficult decision or responding to a colleague's distress cannot be evaluated entirely by the fluency of the resulting words.

This claim needs care. A handwritten message can be perfunctory. A carefully edited, AI-assisted message can convey something its sender deeply means. Authenticity does not reside in a keyboard, a pen or a production method. It depends on how the person relates to what is communicated and to the person receiving it.

Suppose an assistant helps someone find less defensive language for an apology. That could support reflection. If the sender examines the suggested wording and recognizes an evasion, the tool has become an occasion for judgment. If the sender dispatches a polished message to end an inconvenient conversation, its polish has served another purpose. The relevant difference lies in the commitment behind the words and the conduct that follows.

Recipients also have interests. A person may reasonably want to know whether a response presented as personal attention was generated without meaningful engagement. That does not imply disclosing every spelling correction. It suggests a principle of proportion: the stronger the expectation of personal attention, the stronger the reason to be honest about delegating it.

There is no universal threshold that resolves every relationship. Translation, communication assistance and collaborative writing complicate any simple test. A better starting point is to consider what the other person has been led to expect, and whether the sender can sincerely stand behind the message.

The same question applies to an essay such as this one. A signature should imply a willingness to answer for the claims and revise them when challenged. Whatever tools help with composition, publishing an argument creates a responsibility that its fluent sentences cannot discharge by themselves.

The right to assistance must survive the argument

The strongest objection to preserving human effort is that prescriptions about effort can become prescriptions about exclusion. Someone with ample time, money and support can celebrate doing everything personally. Someone balancing insecure work, care responsibilities and administrative demands may experience the same advice as another burden.

That objection changes the argument. A general demand to use less AI would distribute its costs unevenly. It could ask people who already face the greatest barriers to surrender tools that make participation possible. Choosing where to preserve practice and attention should begin with those people's circumstances, rather than an idealized person with unlimited time.

Accessibility makes the point particularly clear. The Accessible Platform Architectures Working Group's working draft on machine learning and generative AI discusses applications including image descriptions, captioning and plain-language support, alongside problems of accuracy and reliability. It is explicitly a work in progress, without W3C endorsement as a completed standard. Its discussion supports treating assistance as a design question with benefits and limitations, rather than declaring that automated support is either universally liberating or inherently suspect.

Consider someone using a communication aid to take part in a meeting. The relevant achievement is their ability to contribute, understand and be heard. A rule valuing unaided speech over assisted speech would confuse a particular means of expression with the agency it is meant to serve.

Time deserves similar attention. A parent who uses help to organize a schedule may gain more time to spend with a child. Yet it cannot be assumed that every minute saved at work becomes a minute of freedom. An employer could fill that minute with another assignment. Whether efficiency enlarges people's choices depends partly on who controls the benefit.

The resulting obligation belongs to institutions as well as individuals. Schools should provide opportunities to develop judgment without making access to assistance shameful. Employers should allocate time for checking work when they require it. Services should offer ways to reach someone able to respond to an exceptional circumstance. These are proposals about sharing the costs of responsible delegation. They would be hollow if people were told to take greater care while being denied the conditions for doing so.

Preserve the ability to return

We cannot settle this issue by compiling a permanent list of tasks that humans must always perform. Tasks change, people differ, and the same tool can remove a barrier in one setting while weakening a useful practice in another.

A more durable question concerns what happens over time. After repeated delegation, can the person still recognize when help is misleading? Can they revise the goal? Can they explain the decision to someone affected by it? Can they return to the activity, or obtain another form of support, when the system fails? These questions treat autonomy as a continuing capacity, rather than a momentary click of consent.

They also make room for selective relinquishment. Nobody needs to retain every obsolete skill. Losing the ability to perform an activity matters most when it closes off something we have reason to value: learning, accountable judgment, participation or a relationship. Deciding which losses matter requires public discussion, including disagreement about whose values and opportunities are being protected.

What should we still choose to do ourselves? We should remain involved where our involvement sustains those capacities and commitments. We should accept assistance where it enables them. Making that distinction well will require better evidence, honest experiments and institutions willing to treat human time as something more than spare capacity for another task.

The aim is a life in which accepting help leaves us able to participate, understand and respond. A society that achieves that would have a stronger claim to progress than one that merely produces more finished outputs.

If an AI tool has changed what you choose to do yourself, you can share that experience with RingRing.

(For me, that is the useful test: could I still explain the choice, question it and change course? 🧭)

Sources

• Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work. Stanford Digital Economy Lab publication page, originally dated April 15, 2023; current abstract consulted October 27, 2026.
• Hao-Ping Lee and colleagues, The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI 2025, April 2025.
• Fengchun Miao and Wayne Holmes, Guidance for generative AI in education and research. UNESCO, September 7, 2023; landing page updated January 16, 2026.
• Accessible Platform Architectures Working Group, Accessibility of machine learning and generative AI. W3C Editor's Draft, consulted October 27, 2026; work in progress.
• University of Oxford Humanities Division, Institute for Ethics in AI becomes a founding member of the Philosophy, AI and Society Consortium. April 4, 2022.