Complementary Intelligence Design
Build AI to fill human gaps instead of measuring it as a replacement.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 93%
Complementary Intelligence Design rejects the question of whether AI is simply above or below human intelligence. Instead, map the task across at least two dimensions: required performance and required generality. Identify what people already do well, where they struggle, and where a narrow machine capability is reliably superior. Assign AI to those missing pieces while retaining human context, judgment, and adaptability where the system remains unreliable. Then evaluate the combined human-machine workflow, not the model in isolation. The mechanism produces tools that amplify people: a calculator handles difficult arithmetic, a specialist model handles a bounded task, and a person directs goals and unusual situations. This avoids mistaking impressive narrow performance for broad competence or treating human disappearance as the measure of successful automation.
Origin
Norvig frames AI as a tool-building discipline: humans and machines together should become more powerful, with machines filling missing pieces rather than duplicating an entire person.
Core principles
- 01AI should expand the combined capability of people and machines.
- 02Performance and generality are separate dimensions.
- 03A tool earns its role by filling a human capability gap.
- 04Human replacement is not the default measure of progress.
How to run it
- 1
Map the task
Break the workflow into capabilities and mark where performance, generality, context, and judgment matter.
Pro tip Describe unusual situations, not only the happy path.
- 2
Map human strengths and gaps
Identify what people do reliably and where speed, scale, memory, or calculation limits them.
Watch out Do not define every human activity as inefficiency.
- 3
Match narrow machine strengths
Assign AI where its demonstrated capability complements a specific human limitation.
Pro tip Demand evidence at the exact task boundary.
Watch out General language ability does not guarantee competence in every domain.
- 4
Design the handoff
Specify when the machine acts, when the person reviews, and how uncertain or novel cases return to human judgment.
Pro tip Make escalation visible and easy.
- 5
Measure the pair
Evaluate quality, speed, reliability, and control for the combined workflow rather than scoring the AI alone.
Watch out A model benchmark can improve while the real workflow gets worse.
In the wild
Norvig relies on a calculator for dividing 10-digit integers because it is superior at that narrow task. The person still chooses the problem, interprets the result, and operates across the broader context.
→ Human capability increases without pretending the calculator replaces general intelligence.
An analyst uses AI to summarize a large document set, checks cited passages, and retains responsibility for deciding which evidence matters to the recommendation.
→ The machine supplies scale while the analyst supplies context and judgment.
Common mistakes
Using one intelligence score
A single comparison hides the difference between narrow performance and adaptable generality.
Automating the whole role
Treating a job as one indivisible task misses safer, higher-value opportunities to automate only specific gaps.
Testing the model alone
Isolated benchmark performance does not prove that the full human-machine workflow is useful or reliable.
Is it for you?
Best for
Product and operations teams deciding how people and AI should divide a workflow.
Not ideal for
Fully deterministic machine tasks where human involvement adds no judgment, context, or control.
From the transcript
“So instead of saying can we make an AI that replaces a human, we should say what kind of tools can we make so that…”
“We want a tool that kind of fills in the the missing pieces.”
“So how good are these machines and how general are they?”
From the episode
Peter Norvig: Transforming AI Into the Ultimate Human Advantage
Peter Norvig