Essay / Tommaso Di Bartolo / Published 2026-11-20

The Habit of Handing Over: AI and Human Agency

Handoverism begins with a question about who should shape our future. It also asks us to examine the choices we make with AI today.

AuthorFull articleEnglish
Adam reaches toward an AI figure with a chrome face and illuminated circuitry, set against Michelangelo’s warm Renaissance fresco.
AI-assisted reinterpretation of Michelangelo’s The Creation of Adam.

By Tommaso Di Bartolo

I would like to spend less time on tasks such as writing drafts, collating information and considering standard alternatives. If an AI can perform well on these tasks, spending longer on them would not be impressive. However, I still want to understand the decision-making that we do need to perform in these situations.

Suppose the recommendation produced by your AI includes good supporting evidence, a logical organization and reasoning for why other decisions would’ve been poor. If you accept this recommendation, it is unclear whether this was a good decision. It may be that an equally well written recommendation for the opposite decision would have been accepted. The debate has a name: handoverism. At its largest scale, it considers entrusting the direction of civilization to AI. While delegation of reports and political decision-making are different, in both cases, it is important to consider what meaningful control people in these situations have.

In her November 11, 2026 essay, Ruxandra Teslo describes a conversation with someone working at a major AI lab. While he was resigned to the idea of machines in charge in the future, he was unable to realistically see any other option. He regarded rule by a benevolent machine as preferable to extinction. Based on his intelligence prediction, he had made an inference about who would be in charge. This seems like a large leap. As systems become more capable, we should be given more options. It is therefore important to pay equal attention to the decision making around using these systems.

Learning is part of the work

Much professional judgment occurs during the tasks we will later delegate. During these tasks, people may discover a mistaken assumption, understand why a well-intentioned decision didn’t work, or recognize what a user meant by their criticism. Once an individual has experience performing these tasks, they may be able to skip large parts of the process. However, if a novice is given this shortcut, it will be difficult for them to gain this experience.

Before considering the quality of students’ submissions in an AI learning system, we should consider how the system allows people to learn. In a 2025 experiment by Hamsa Bastani and colleagues, nearly a thousand high-school mathematics students in Turkey completed practice questions with different types of support. Students using a basic GPT-4 interface improved their assisted practice performance, but performed worse in the final assessment (without support) compared to students that completed the practice questions without AI support. Teacher-informed safeguards in a different tutor largely mitigated this harm.

It is easy to miss this during an exciting product demo. If the final practice sheets are presented, it appears the basic support is helping. Students’ subsequent work without the tutor gives us another way to judge that apparent success. Fortunately, the design of this tutor means we can create good support based on learning, as long as we care to measure it by learning. Simply reducing the time taken is not sufficient.

Similarly, in a workplace, care should be taken when a large volume of output has been generated through automation. We may be removing tasks with no learning component. However, we also may be removing tasks that are learning experiences. If this is the case, we run the risk of our new reviewers struggling with judgment. It is important to understand what parts of a job are being automated.

Age alone is unlikely to be very informative in this situation. An executive may be inexperienced in this domain, and a junior expert may have very good reasons to doubt a model. My hypothesis is that competence may become harder to assess if its principal signal is the quality of AI-assisted work. We may place too much emphasis on the quality of output in order to infer understanding without verifying a link between them. The studies discussed here do not establish a generation becoming less competent. Instead we have strong reasons to improve how we allow people to demonstrate what they have learned.

Review requires resources

While “Keep a human in the loop” sounds simple, if there is no specification of what humans should do, then it may be an inefficient use of resources. Having someone’s name on an approval flow is very cheap. However, providing them with the information, allowing them to investigate, and giving them authority to delay a release costs more.

If a reviewer believes the marketing strategy recommendation is based on incorrect estimates of customer demand, they may need to investigate this. This investigation may involve speaking to customers and lead to a different conclusion. If the launch date cannot be changed, and the budget has been spent, how much influence has the review provided? The reviewer may still be held accountable, despite no longer having the opportunity to change the outcome.

An April 2025 CHI study by Hao-Ping Lee and colleagues surveyed 319 knowledge workers. The more confidence participants expressed in generative AI, the less critical-thinking effort they said they put in. They described checking generated material and bringing it into their work. However, since this is a self-report study, it does not show a decline in ability. I would give that checking work the resources it requires. Just because verification has been moved to the end of a process, it does not mean it is free.

More of the work preceding review can now be delegated. Anthropic examined samples of Claude.ai conversations from late 2024/early 2025 and August 2025. Its September 2025 Economic Index put directive exchanges at 27% in the earlier samples and 39% in August. Here, users delegated a complete task without much discussion. Shorter conversations may reflect improvements in the model. The figures concern how people used Claude, without measuring their cognitive ability. However, it does mean that we have a greater design challenge in considering under what conditions approvals occur. Good recommendations can be given to users without them seeing much of what would help them identify a bad recommendation.

Opportunities exist for designers and managers. Allow assumptions to be verified. Provide easy options to check the recommendation. Give time for differing opinions. Include information on what will happen if a reviewer believes that the plan is incorrect — and who can give approval for changes to be made.

Other people can change the objective

Another person can bring considerations we may not have included. For example, feedback from a customer may mean we’ve got the problem definition wrong. Similarly, a colleague may explain that the increased efficiency is just shifting work to their team. Although we can predict these using AI, we still want to have the discussion.

Research by Julian De Freitas and colleagues, published online in June 2025, found that AI companions could reduce momentary loneliness, including repeated relief over a week. This was partly explained by the feeling of being listened to. This is valuable. If a person feels listened to, they may be in a better place to engage with others. Products that encourage this engagement would be desirable.

I would want to know whether, over time, companionship helps someone engage with other people or takes the place of that engagement. The study leaves this open. We should be cautious in concluding that the social benefit of a product is solely determined by providing quick relief. In human relationships, we must compromise to meet needs that we have not defined, and accept responses that we may not like. We want this to be part of our learning to work and live well with others.

Within a professional context, we want to allow recipients to provide feedback that can still impact the work. If feedback is only collected once decisions have been made, even with a good system in place to listen to people, the recipient will have very little influence.

The case for better machine judgment

Despite this, we need to consider that AI may often make better decisions than humans. I expect to welcome many of these improvements. It would be strange for a techno-optimist to wish to maintain tasks for humans simply for the sake of it.

Scott Alexander makes a serious case for this in “Against Against Handover,” published October 20, 2026. He considers how aligned AI can benefit humans, and how local decision-making can also be maintained. We should pay more attention to this than to simply stating that humans should always be better.

However, we should be cautious about giving too much weight to good decision-making by AI. Even though an expert may understand better what the likely consequences of my decisions are, I still expect to have some control over which consequences I am willing to take. If a very capable system can help people understand a wide range of options (including those it does not recommend for them), this would be a considerable benefit. However, we must still consider within institutions how people may change the objectives of the system, or take back delegated decision-making.

Humans have difficulties maintaining these responsibilities in institutions. If only the head of an institution is human, but no-one else has meaningful ways to challenge decisions, it will have little impact. I would ask the same demanding question of proposed AI governance. How much change can people in the situation actually make if they disagree?

The question, the decision and the relationship

In the Agency Reel on Instagram, I demonstrate how to keep the question, the decision and the relationship in your hands while delegating the work to AI.

In a hypothetical marketing plan, asking “Write me a marketing plan” is very vague. Asking “What would our competitor never expect us to do?” leads the investigation in a different direction. However, in both cases, we must decide whether the planned action would be beneficial to the customer. While surprising actions may be useful, this must be for a worthwhile goal.

Before reviewing options, write down which criteria you are applying. Perhaps reaching a particular group of new customers matters more than generating an immediate response. If during analysis you decide to prioritize differently, make clear why. Now you have considered decisions that may have otherwise been hidden within the request.

Give the recipient the opportunity to provide feedback. What would they now have to do? Why would they trust you enough to take that next step? Give them the opportunity to change the proposal, even if it means going back to previously finished work. However, if the team has only budgeted for creating the plan, they have made a decision on how much input they will allow.

The habits we choose

We shape technology. Through repeated use, it can start to shape us.

Using social media should make this familiar. Alongside its benefits, many people are consciously trying to change habits they have formed. In one randomized study, people who removed themselves from Facebook for four weeks showed an improvement in well-being, and continued to use it less after the experiment. By taking a break, people changed what they wished to return to.

With AI, we can consider the habits we want to develop. If we were to use AI every day for five years, what would we want to be better at? Asking more precise questions? Understanding new topics? Making decisions we can justify?

Working backward, does our current usage of AI give us practice in these skills?

Handoverism is one possible outcome. If we begin to develop a habit of asking AI for an answer, without thinking for ourselves, and taking recommendations without considering how to evaluate them, this could become our default relationship with intelligence.

Alternatively, we can be very intentional. Putting our phones away creates an opportunity to focus. Pausing before entering a prompt allows us to think for ourselves. “Lock your prompt,” as I put it in the video: think first, then open the tool. Even if our starting answer is tentative, we can explore it, challenge it and improve it using AI.

This is the habit we should aim to develop in a Participation Era: using greater intelligence to become more capable participants.

Handoverism happens one prompt at a time. So can independent thinking.

AI isn't the enemy. The old script is.

Language: English. English original written by Tommaso Di Bartolo.

Source: Tommaso Di Bartolo · Published 2026-11-20

Original publication ↗
Ask me on Lana