Decision-and-Creation AI Test
Separate true AI from automation by testing decisions, learning, and creation
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 95%
Harari distinguishes AI from ordinary automation by asking what the system can do independently. A conventional machine follows a program: it may perform a task automatically, but it does not learn, alter its own behavior, or originate a new option. The stronger threshold is independent decision-making, followed by the ability to learn and produce ideas its designers did not directly specify or anticipate. Applied as an assessment, the test begins by mapping which choices remain human, then checks whether the system adapts from experience and whether it creates genuinely new outputs rather than selecting from fixed instructions. This does not prove consciousness or wisdom. It clarifies the operational capability being discussed and prevents every automated product from being treated as equivalent AI.
Origin
Extracted from The Diary of a CEO. Yuval Noah Harari contrasts a fixed coffee machine with one that learns a preference, acts on it, and proposes a new drink.
Core principles
- 01Automation alone does not make a system intelligent
- 02Independent decisions mark a meaningful capability shift
- 03Learning changes behavior beyond a fixed program
- 04Generating unanticipated ideas is stronger evidence than repeating instructions
How to run it
- 1
Map the Fixed Program
Document what the system was explicitly instructed to do and which outputs follow predetermined rules. Treat reliable automatic execution as automation, not sufficient evidence of AI.
Pro tip Ask what the system would do in a situation its designers did not enumerate.
- 2
Locate Independent Decisions
Identify choices the system makes without waiting for a human to select the answer. Note the scope and consequences of each choice.
Pro tip Use a concrete decision such as recommending, approving, rejecting, or prioritizing.
Watch out Independent action does not make a decision accurate, fair, or safe.
- 3
Test for Learning
Check whether experience or new data changes future behavior. Distinguish adaptation from a hidden schedule or rule written in advance.
Pro tip Compare behavior before and after relevant feedback.
- 4
Test for Creation
Look for a new option, strategy, or idea that was not directly supplied by a human. Verify that the claim is more than recombining a fixed menu under another label.
Pro tip Ask the vendor to show an unanticipated output and explain how it was evaluated.
Watch out Novelty alone is not usefulness.
- 5
State the Boundary
Describe the product precisely as fixed automation, adaptive decision-making, generative AI, or a combination. Preserve human accountability regardless of the label.
Watch out The episode's distinction is conceptual, not a formal industry standard.
In the wild
Harari contrasts a machine that automatically makes a programmed drink with one that learns a person's preference and independently predicts they want an espresso. His strongest example is a machine that proposes a new drink the designers did not specify. Each added capability moves the example from fixed automation toward his working definition of AI.
→ A vague AI label becomes a set of observable capability questions.
Illustrative example: a buyer discovers that a loan tool merely applies fixed thresholds. It makes no learned adaptation and creates no new policy. The team classifies it as automated decision software, then focuses scrutiny on the human-designed rules rather than accepting an expansive AI claim.
→ The buyer uses a more precise category and asks better accountability questions.
Common mistakes
Calling Every Automatic Tool AI
Harari explicitly distinguishes a pre-programmed coffee machine from a system that learns or makes independent decisions.
Confusing Novelty With Quality
An unexpected output may show generative capability without showing that the output is useful or safe.
Treating Intelligence as Consciousness
Harari later separates solving problems from having subjective feelings.
Is it for you?
Best for
It is best for product buyers, operators, and leaders evaluating broad or inflated AI claims.
Not ideal for
It is not ideal as a technical benchmark of model quality, safety, consciousness, or general intelligence.
From the transcript
“it's the first technology ever in human history that is able to make decisions independently”
“there is a lot of automation out there which is not AI”
“it's really AI when it comes up with completely new ideas”
From the episode
Yuval Noah Harari: This Election Will Tear The Country Apart! AI Will Control You By 2034! The Dark Truth Behind Meta & X!