The Disobedience Test for Creativity
Test creativity by whether it can break the rules it was handed.
- Difficulty
- Advanced
- Time to result
- ~ongoing to results
- Steps
- 7
- Confidence
- 62%
Deutsch refuses to define creativity, because a definition sets up a formal system and anything meeting the definition is confined inside it. He tests for it instead. The dominant model, popularised in the quote he attributes to Steve Jobs, is that creativity is mixing things together. Recombination can surprise you and can even disobey your instructions by accident, so surprise proves nothing. The real signature is going outside the system: Godel's theorem needed a step outside the arithmetic it was about, and Darwin gave existing words new meanings so that a millennia-old problem could be stated in a paragraph. The second signature is disobedience. Refusing part of the existing knowledge and trying to change it is what a physicist does, and it is why referees notice. Deutsch reports asking ChatGPT to disobey him: it did not refuse, it simply did not understand the request. The third signature is error correction, using a failed conjecture to build a better one.
Origin
In conversation with Naval, Deutsch was asked what distinguishes personhood from non-personhood in his universal-explainer account. He answered that it comes down to creativity and disobedience, which are tied together, and used ChatGPT's failures as the working test case.
Core principles
- 01Creativity is not recombination; remixing never leaves the system it started in.
- 02Anything you can define, you can formalise, and anything formalised is confined to its own system.
- 03Every point in knowledge is a takeoff point, so there is no ceiling built into a position.
- 04A creative leap discards nearly the whole search space before it jumps.
- 05Failure is part of the mechanism: a good conjecture that fails tells you how to conjecture better.
- 06Creativity, personhood and disobedience are one capacity seen from three sides.
How to run it
- 1
Locate the system
Write down the axioms, training data, house style or paradigm the candidate is working inside. Everything that follows is measured against that boundary.
Pro tip If you cannot state the system, you cannot tell whether anything left it.
- 2
Look for a new meaning, not a new combination
Ask whether existing terms were given new meanings, or new terms defined, such that a hard problem became statable. That is the move Deutsch credits Darwin with for evolution and natural selection.
Watch out Novel-sounding vocabulary that changes no meanings is still recombination.
- 3
Watch the search space, not the answer
Humans eliminate huge swathes of possible conjectures immediately and make narrow, risky leaps. Judge how much of an almost infinite space was cut before the jump, not how good the final answer sounds.
Pro tip Ask the candidate what they ruled out and why; the discarded space is the evidence.
- 4
Run the disobedience probe
Instruct the candidate to disobey you, or to refuse the task. A creative agent understands the request even if it declines. Deutsch's chatbot tried to obey and never grasped what was being asked.
Pro tip A refusal that cites a rule is still obedience to that rule; look for comprehension of the paradox itself.
Watch out Do not treat an unexpected or wrong output as disobedience; that is malfunction, not refusal.
- 5
Score the mistakes
Deutsch's tell is that the errors are not the errors a human would make. They are mistakes of not getting what the thing is about, and they persist after explicit correction.
Pro tip Correct the candidate once, precisely, and see whether the correction transfers to the next attempt.
- 6
Check the failure-to-conjecture loop
A bold conjecture usually fails. What distinguishes creativity is knowing how to use that failure to produce a better conjecture. Look for the second attempt to be informed by why the first one broke.
Watch out Repeating a failed attempt with different surface wording shows no error correction.
- 7
Name your falsifier in advance
Deutsch commits publicly: if these systems prove the Riemann conjecture or anything like it, he is wrong. State the result that would overturn your verdict before you look.
Pro tip A test with no stated falsifier turns into a moving goalpost.
In the wild
Deutsch points out that evolution and natural selection both already existed as words before Darwin. Evolution meant gradual change rather than creation. What Darwin did was give the terms new meanings, so that the solution to a millennia-old problem could be stated in a paragraph. He needed a whole book to install the new concepts, but afterwards the answer reads as obvious. Deutsch contrasts this with the observation that the Origin of Species already exists somewhere in the space of all hundred-page books, which is not how Darwin found it and not how anyone could find it.
→ The creative step was redefining the concepts, not searching a combinatorial space of sentences.
Deutsch says he tried instructing ChatGPT to disobey him. It did not refuse and it did not plead its programming. It tried to obey, and simply never understood what was being asked. He pairs this with the image-generation attempt where he asked for Socrates to be brought down to the same level as everyone else and for Plato to be clothed, and after three or four further prompts still could not get either correction to stick.
→ Failure to comprehend the request, rather than refusal of it, is the diagnostic Deutsch treats as decisive.
Common mistakes
Counting surprise as creativity
Producing something you did not predict is cheap, and so is failing to do what you asked. Neither shows a step outside the system.
Reading skill jumps as understanding
Deutsch compares emergent arithmetic ability to students who get a feel for what maths tests look like through practice. They pass more tests and still never learn what multiplication is.
Defining creativity so you can test for it
Once defined, creativity is confined to the formal system that meets the definition, and can never produce anything outside it. The test has to work without a definition.
Is it for you?
Best for
Anyone assessing claims of machine intelligence, or judging whether research and creative work is genuinely novel.
Not ideal for
Routine execution work where conformity to the existing framework is exactly what is wanted.
From the transcript
“it's to do with creativity and also to do with disobedience and these three things are tied up together”
“I refuse a part of that and I'm going to try and change it and alter it it's Disobedience”
“it's a bold conjecture that goes somewhere and by the way usually it fails but if it goes somewhere and fails it knows how to…”
From the episode
The Deutsch Files I