Two-Dimensional Automation Choice
Increase machine capability without automatically surrendering human control
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 96%
Automation is often presented as a one-dimensional ladder: as the machine does more, the human does less until the person disappears. Norvig argues for two independent dimensions instead. One dimension measures what the technology can do; the other measures how much control the user chooses to retain. A capable system may take over completely in one context while remaining interruptible, steerable, or advisory in another. The design mechanism is to map situations, identify decisions people still value, and expose control modes that preserve those choices. This avoids equating technical progress with compulsory surrender of agency. It also produces a more respectful product: the user can benefit from difficult work being automated without losing the ability to redirect, slow, stop, or modify the system when circumstances change.
Origin
Peter Norvig challenged the self-driving automation diagram in which the person fades away as the automation level rises, proposing a two-dimensional alternative.
Core principles
- 01Automation capability and human control are separate dimensions
- 02More automation should not make the person disappear
- 03Users should choose how much control they retain
- 04The right control level can change by situation
How to run it
- 1
Map machine capability
List what the system can reliably perform without human intervention.
Pro tip Separate demonstrated capability from promised capability.
Watch out Do not use a single automation label to hide uneven performance.
- 2
Map valued human choices
Identify the decisions, exceptions, and moments where users may want direct control.
Pro tip Ask about situational preferences, not one permanent setting.
Watch out Convenience in routine cases does not imply consent in unusual ones.
- 3
Design control modes
Offer modes that range from assistance to full delegation while preserving clear boundaries.
Pro tip Name modes by what the person and machine each do.
Watch out A cosmetic toggle is not meaningful control if the user cannot predict its effect.
- 4
Provide live intervention
Let users redirect or override the system when their intent changes.
Pro tip Keep the most important interventions immediate.
Watch out An override buried behind friction does not preserve practical autonomy.
- 5
Test for retained agency
Observe whether users understand who is in control and can choose the level that fits the moment.
Pro tip Test routine and exceptional scenarios separately.
Watch out Efficiency gains can mask a loss of user confidence or responsibility.
In the wild
A passenger trusts a self-driving car enough to sleep on a familiar motorway route, so the car handles the journey. In town, the same passenger wants to choose an unscheduled stop and redirect the route. The product supports both full delegation and immediate human direction instead of forcing one universal automation level.
→ The passenger gains automation without giving up situational control.
Common mistakes
Using a one-way automation ladder
It assumes every gain in machine capability should remove an equal amount of human involvement.
Offering control without usable overrides
Users retain no practical agency if they cannot understand and change the system's behavior in the moment.
Is it for you?
Best for
Designers of AI products and automated workflows where user preferences vary by context.
Not ideal for
Safety-critical operations where regulation requires one fixed control allocation.
From the transcript
“I don't want technology that makes me disappear I want technology that respects me”
“I'd rather have it be two two-dimensional”
“let me make the choice of how much the machine is going to be doing and how much I'm going to keep control”
From the episode
Peter Norvig: Simple Ways to Grow Your Business with AI
Peter Norvig