The Human-Welcomed Automation Filter
Automate the tasks people want to give away, not the moments they value
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 96%
The Human-Welcomed Automation Filter chooses automation targets through human preference rather than technical possibility. Break a role or activity into concrete tasks, then ask the people affected to rank where robotic or AI help would be welcome. Separate tasks that look simple but carry emotion, identity, care, or shared meaning from tasks that are repetitive, exhausting, unpleasant, or dangerous. A wedding-ring choice or opening family gifts may be easy to automate yet valuable to keep human; cleaning a toilet may be a far more welcome target. Prioritize the high-burden, high-preference tasks, prototype that allocation, and check whether users experience the technology as help. The output is not maximum automation. It is a task boundary that protects what people care about while directing machines toward work they genuinely want removed.
Origin
In a robotics project covering a thousand everyday tasks, Dr. Li's lab asked people to rank thousands of tasks by where they wanted robotic help before choosing what to build.
Core principles
- 01A task's technical feasibility does not determine its human value
- 02Ask people what help they welcome before automating
- 03Preserve activities that carry emotion, dignity, or meaning
- 04Prioritize burdensome, repetitive, or hazardous tasks
How to run it
- 1
Decompose the activity
List the concrete tasks inside the job, service, or everyday activity. Avoid treating an entire role as one automation target.
Pro tip Observe the workflow so invisible support tasks enter the inventory.
Watch out Role-level labels hide the difference between meaningful and thankless work.
- 2
Collect preference rankings
Ask affected people which tasks they want machines to help with and which they want to retain. Gather rankings at enough scale to reveal consistent preferences.
Pro tip Ask about help with specific tasks rather than support for automation in general.
Watch out Do not substitute the product team's preferences for users' preferences.
- 3
Protect human meaning
Flag tasks carrying joy, emotion, care, identity, judgment, or social connection. Preserve them even when they are technically easy to automate.
Pro tip Ask what would be lost if the task were completed perfectly by a machine.
Watch out Task difficulty and task value are different variables.
- 4
Prioritize welcomed relief
Choose tasks with strong demand for help, especially repetitive, unpleasant, exhausting, or dangerous work. Build around the clearest preference signal first.
Pro tip Start where burden is high and emotional value is low.
Watch out A technically impressive task may be a poor target if people do not welcome it.
- 5
Validate the allocation
Prototype the human-machine split and ask whether it feels beneficial in practice. Revise boundaries where the technology removes agency or creates new burdens.
Pro tip Measure adoption and relief, not just task completion.
Watch out Stated preference can change after people experience the real workflow.
In the wild
Dr. Li contrasts choosing a wedding ring and opening Christmas gifts with cleaning a toilet. The first tasks are not primarily about completing a transaction or opening a box; they carry emotion, joy, and family meaning. Toilet cleaning carries little of that value and attracts much stronger demand for robotic help.
→ The robot program focuses on welcomed assistance rather than replacing emotionally important moments.
Common mistakes
Automating by feasibility
Selecting tasks only because a model or robot can perform them ignores whether people value doing them. Technical success can therefore create a rejected product.
Surveying at the role level
People may want help with some parts of a job while fiercely protecting others. Ask about concrete tasks to expose that distinction.
Is it for you?
Best for
It is best for robotics, workplace AI, and service automation where a role contains both valued and unwanted tasks.
Not ideal for
It is not ideal where safety, law, or accessibility requires automation regardless of stated preference.
From the transcript
“we actually ask people to rank for us thousands and thousands of task and tell us which tasks they want robots help.”
“on those tasks that humans prefer robotic help rather than those tasks that humans care and want to do themselves.”
“How do we create technology that is beneficial, welcomed by humans”
From the episode
Dr. Fei-Fei Li: Turn AI Into Humanity's Greatest Ally, Not Its Biggest Threat
Dr. Fei-Fei Li