Keep, share, hand over

Three columns for deciding what a model gets. Keep is what never goes: your position, your voice, the learning struggle, the final decision, anything you have to defend. Share is done together: exploring options, critique of your draft, finding what to read, rehearsal, brainstorming. Hand over is take it, please: formatting, converting files, organising notes, repetitive processing. If you cannot say which column something is in, it belongs in the first.

The three columns#

Keep. Never hand this over. Your position. Your voice. The learning struggle. The final decision. Anything you must be able to defend.

Share. Do it together. Exploring options. Critique of your draft. Finding what to read. Rehearsal and simulation. Brainstorming.

Hand over. Take it, please. Formatting. Converting files. Organising notes. Repetitive processing. Mechanical tasks.

The default rule does most of the work: if you cannot say which column it is in, it goes in the first one. That default is deliberately conservative. Delegation drifts the other way when nobody sets one.

Sixty years of levels of automation, and why three columns instead#

Deciding what a machine gets is one of the oldest questions in human factors, and the field's answer is a scale rather than a sort. Sheridan and Verplank set out ten levels of automation in 1978, running from the computer offering no assistance at all to the computer deciding everything and acting autonomously. Parasuraman, Sheridan and Wickens generalised it in 2000 across four stages of information processing, so a system can sit at a different level for acquiring information, analysing it, deciding and acting.

That literature is better than this framework in every technical respect. It is precise, validated across decades of aviation and process control, and the right tool for a safety engineer. Its findings are also unflattering to the middle: Parasuraman and Riley distinguish use, misuse, disuse and abuse of automation, and Parasuraman and Manzey show that partial automation produces complacency and attention bias rather than attentive supervision.

What that literature is not is usable by a person at a desk in four seconds. Ten levels across four stages is a design tool for a system. Three columns is a sorting tool for a task, and the price of the simplification is real: it collapses a continuum into buckets and it says nothing about how far a tool should be trusted within a column.

Regulation has the same shape and puts it differently#

Article 14 of the EU AI Act requires that high-risk systems be designed so that natural persons can effectively oversee them, and the ICO's meaningful-involvement test asks whether a human review is substantive rather than a rubber stamp. Both are Keep expressed as a legal obligation: certain decisions must remain with a person in a way that survives inspection.

Switzerland states it most plainly of the instruments this research has read. Guideline four of the Federal Council's 2020 guidelines on artificial intelligence: responsibility must not be capable of being delegated to machines. That is a first-column rule in a national policy document.

The column that is actually contested#

Hand over is easy and Keep is easy once stated. Share is where the argument is, because Share is where automation complacency lives: a person nominally in the loop, an output that is usually right, and attention that degrades precisely because the tool is reliable.

Two things make Share survivable. Produce something before the model does, which is the finding from the guardrailed tutor arm and from the scaffolded programmers, whose blackout failure rate was 39 per cent against 77. And check unaided occasionally, because a Share task quietly becomes a Hand over task and nothing announces the move.

What this has not been shown to do#

No study has tested a three-column sort against a levels-of-automation scale, or against no framework. The claim that "if you cannot classify it, keep it" is a judgement about which error is cheaper, not a measured result, and it will be wrong for anyone whose bottleneck is throughput rather than capability.

This is also the operational face of an argument made elsewhere on this site, drift versus design, and it inherits that argument's status: a position supported by evidence about what happens when nobody decides, not a finding about what happens when they do.

Key sources

• Parasuraman, R. and Riley, V. (1997). Humans and Automation: Use, Misuse, Disuse, Abuse. Human Factors, 39(2).
• Parasuraman, R. and Manzey, D. (2010). Complacency and Bias in Human Use of Automation. Human Factors, 52(3).
• Bastani, H. et al. (2025). Generative AI Without Guardrails Can Harm Learning. PNAS, 122(26).
• Sankaranarayanan, S. (2026). Mitigating "Epistemic Debt" in Generative AI-Scaffolded Novice Programming. arXiv:2602.20206.

Related SuperSkills research#

The fuller version of this decision, the delegation boundary map, and the argument underneath it, drift versus design. On the middle column, automation complacency and why human in the loop is not a safeguard. On putting friction into the request itself, goal, context, friction, standard.

About this framework#

Keep, share, hand over is used by Rahim Hirji in teaching and in the Mastering AI deck, most recently in September 2026, where it is presented as drift versus design made operational. No claim of first use is made. No dated first publication exists for it and the Box of Amazing archive carries none, so it anchors to SuperSkills (Kogan Page, 2026) and to the deck. The levels of automation belong to Sheridan and Verplank and to Parasuraman, Sheridan and Wickens. The reading offered here, that a scale is right for designing a system and a sort is right for a person deciding about a task, is an interpretation by Rahim Hirji and is marked as an interpretation and not a finding.