Policy WindowAI, evidence & public choice
Menu
Human by choice
A Policy Window advocacy programmeWiki · neutral reference ↗

Which work should stay human

Which work should stay human as AI advances, and who decides

Guide v0.5 ·

Suppose AI could do a task well. Would there still be a reason to keep a person involved? The answer depends on what the person contributes, who benefits and what everyone else gives up.

A care service could use AI to organise visits so its staff spend more time with people. It could also use technology to replace some of those visits. Those choices concern different parts of the work, with different consequences for the people receiving care and those doing it.

Human by choice examines when particular tasks should retain people. No single reason settles a case.

The starting point is a specific task. A job can include administration, judgment, physical activity and relationships. Before deciding whether to protect a whole occupation, ask which part of it needs a person and why.

Three reasons to keep a person involved

Human presence. A relationship may matter even if AI performs the task well. People receiving a service may disagree.

Answerability. A person may need to explain a consequential decision, hear a challenge and change the outcome.

Time to adjust. Phased automation may help workers move jobs. Specify an end point and payer; compare income support, shorter hours or retraining.

Doing the task and controlling the decision are different

Keeping a task human differs from controlling AI-assisted work. Meaningful control requires authority, competence, evidence and time to disagree, with involvement recorded.

ICO guidance already addresses human involvement in relevant personal-data decisions, including training and monitoring. Its AI guidance is under review after the 2025 legislation; newer automated-decision guidance remains draft. Extending these expectations to other work needs separate justification and lawful authority.

Review can correct errors or add them. An override count alone cannot establish benefit.

Keeping a person involved has costs too

A rule that preserves a human role can consume scarce staff time, raise prices or lengthen a queue. People waiting for a service matter alongside those already receiving it. A rule can also protect established providers from competition, without improving the service.

Research does not give a general guarantee that adding a person improves an AI decision. A 2024 review of 106 experiments found that people using AI performed better on average than people alone, but worse than whichever performed better alone: the person or the AI. The experiments were published from January 2020 to June 2023. They covered particular tasks and systems; they do not settle performance in every workplace today.

That leaves two questions to examine openly. Does this particular arrangement improve the service? And is there a human relationship or responsibility worth retaining even when it has a cost? Evidence can inform both questions, but a performance score cannot decide how much a community should value a relationship.

A serious proposal should compare keeping a task human with redesigning it. Could AI take on preparation so that staff have more time for people? Could workers share the gains through pay, hours or ownership? Preserving a task is one possible response to change. It should earn its place against the alternatives.

Who gets to make the choice

The power to decide varies with the task and the proposed rule. An employer can change some working arrangements; a union can negotiate an agreement; a professional body can set standards within its authority; Parliament can make law. These routes give different people a say and offer different ways to challenge a decision. A specific proposal has to establish which route can actually deliver it.

State who benefits, decides and pays, how evasion is detected, and effects on access and workers. Say whether protection is permanent or transitional; set a review date.

Those statements do not create public consent. The people whose work, services or money are affected may disagree about the answer. Making that disagreement visible is part of deciding well.

What do people want the human to do?

In an Ada Lovelace Institute and Alan Turing Institute survey, 55% selected more human involvement as something that would increase their comfort with AI. Respondents could select several measures. A separate item asked which statement best reflected their view: 37% favoured humans rather than computers always making decisions and being able to explain them to the people affected. That item introduced AI intended to improve speed and accuracy, alongside possible difficulty explaining its decisions.

These weighted results came from 3,513 UK adults on the probability-based NatCen Opinion Panel, surveyed from 25 October to 24 November 2024. They describe that period. This selects two questions from a study that also examines particular uses of AI.

The first answer leaves the human role unspecified; the second states a categorical principle about decisions and explanations. The figures should not be added or treated as evidence that people changed their minds. The comfort item did not ask people to put safeguards in order of preference. Neither item chooses a staffing rule or establishes the cost people would accept. Broad preferences can inform a discussion without settling a particular choice.

Ask a provider which task the person will actually perform, who can change the outcome and what retaining that role means for access and cost. Use the six questions below to make those choices specific.

Your notes

Use these questions to make a choice about one task visible and open to challenge. They organise a discussion; they do not automatically decide whether a task should stay human.

Your notes stay in this page while it is open. Download a copy to keep them. This worksheet does not send your answers anywhere.

For example, organising a care visit or deciding a benefit claim. Separate the task from the whole job.

Apply the reasons above; identify beneficiaries and disagreement.

If AI assists, test the control conditions above. How would involvement be recorded?

Identify the actual decision-maker and the affected people who might otherwise be left out.

Include effects on waiting, prices, access, pay and working conditions. Compare another way to deliver the same value.

Name the evidence to look for and a review point. A temporary adjustment rule also needs an end point.

Two invented examples

Numbered entries follow the six questions above.

Planning care visits

  1. An invented example: arranging the timetable for care visits, separately from spending time with the person receiving care.

  2. Human presence may matter during visits; scheduling is a separate choice.

  3. One option to examine is AI-assisted scheduling, with staff able to change the plan when it misses a person’s needs. The interaction during the visit is a separate choice.

  4. Assume a care provider is considering the change. Establish what it can decide, and involve people receiving care, carers and staff in judging the options.

  5. Compare staff time, missed or rushed visits, access and working conditions. A faster timetable is not enough if it reduces the value of the visits. A wholly manual timetable has costs too.

  6. A possible trial would compare the service before and after the change, including time with people and their views. Set a review point before adopting the arrangement.

Reviewing a benefit recommendation

  1. An invented example: reviewing an AI recommendation about a benefit claim.

  2. Answerability matters because the outcome affects the claimant's income.

  3. Test the control conditions above; compare other arrangements and record departures and reasons.

  4. Assume a public agency is considering how to organise reviews. Establish its actual authority and the relevant rules. Include claimants and reviewing staff in the discussion.

  5. Compare errors, challenge routes, waiting times and staff workload. A signature that adds delay without improving the outcome is not enough.

  6. Examine whether disagreements improve decisions and whether the arrangement works for different groups. Set a review point and specify what evidence would lead to a different design.

Sources and credit

Bill Gates's Human Reserved discussion informs this guide. Earlier work includes the International Bar Association's report on AI and work (2017). IPPR advocated regulation retaining human responsibility in key areas and incentives for AI to support rather than replace workers (2024). This initiative is unaffiliated with them.

  1. ICO guidance

    Under review; checked 12 November 2026.

  2. ICO draft guidance

    Updated 31 March 2026; checked 12 November 2026.

  3. Vaccaro, Almaatouq and Malone on human and AI performance

    Systematic review and meta-analysis, Nature Human Behaviour, 2024. The comparison with people alone differs from the comparison with the better of person or AI alone.

  4. Bill Gates on Human Reserved

    26 August 2026.

  5. International Bar Association report on AI and work

    April 2017.

  6. IPPR on AI and work

    Official press summary, 27 March 2024.

  7. Findings

    Figures 10 and 12.

  8. Report

    Methods.