Creativity in AI

In 2012 I did an interview Margaret Bowden, where she discussed her famous three-part taxonomy of creativity: combinational (making new associations), exploratory (working within an established framework), and transformational (breaking and altering the framework altogether).

Margaret provided a highly influential framework for defining creativity in AI, though she did not design it as a rigid metric or scoring tool. Instead of treating creativity as an unmeasurable, mystical concept, she operationalised it by looking at tangible outputs (ideas or artefacts) and evaluating them through three key criteria and three computational types.

The actual ‘measurement’ often relies on human context and domain-specific valuation rather than a simple mathematical formula.¹

• Novelty: The artefact must be new.
• Surprise: The artefact must break or subvert expectations based on standard rules.
• Value: The artefact must be useful, beautiful, interesting, or valuable within a specific domain.

Making the Criteria Measurable: P-Creativity vs. H-Creativity

To make “novelty” and “surprise” measurable, Boden split creativity into two operational levels:

• P-Creativity (Psychological): The idea is fundamentally new to the agent itself. For an AI, this can be measured computationally – did the algorithm generate a state, sequence, or structure that it has never encountered or produced before in its training/programming history?
• H-Creativity (Historical): The idea is new to human history. This is measured by comparing the AI’s output against a global database of existing human knowledge or art.

Can AI abductively create new conceptual schemas?

People argue that AI cannot generate new conceptual schemas, and link to papers like Zahavy’s “LLMs Can’t Jump” paper – which is explicitly a position paper – not an empirical finding.² Its thesis is roughly that current LLMs can perform induction and increasingly deduction, but lack the mechanism required for the kind of abductive leap by which Einstein moved from physical experience/thought experiment to new foundational axioms.

The paper uses Einstein’s development of general relativity as its central case and calls the movement from sense experience to axioms the “Jump”. It argues that this requires embodied simulation or an interactive physical world model, and actually proposes such world models as a route by which future AI could acquire the missing capability.

Zahavy does not demonstrate structural impossibility. He reconstructs one exceptional episode in the history of physics and proposes an interpretation of how Einstein did it. There isn’t an architectural theorem showing that transformers cannot implement abduction, nor an operational benchmark on which all current LLMs necessarily fail.

The inference Zahavy makes is essentially: Einstein’s discovery involved something beyond ordinary induction/deduction -> that something was embodied/manipulative abduction -> current LLMs lack the appropriate embodiment -> therefore current LLMs cannot perform this kind of jump.

This is not established.

More importantly, Zahavy himself has subsequently clarified the strength of his claim. He explicitly said that the paper is not claiming LLMs can never make scientific discoveries, acknowledged that LLM-based systems are already contributing discoveries, and wrote that it is “quite possible that I am wrong” and that scaling current systems might produce new inventions in physics and elsewhere.³

There is already empirical evidence putting pressure on the strong claims that “LLMs cannot abduct”. Google’s AI Co-Scientist was given an unresolved biological question whose answer researchers had discovered experimentally but had not yet published. Its highest-ranked hypothesis independently proposed essentially the experimentally discovered mechanism, and several other generated hypotheses opened new research directions. The work was subsequently published in Cell.⁴ A 2026 Nature paper on Co-Scientist describes it as producing demonstrably novel scientific hypotheses through iterative generation, criticism and refinement.⁵

Current AI has not clearly demonstrated anything comparable to independently inventing and developing a major scientific paradigm such as natural selection, general relativity or quantum mechanics, where it identifies that the existing conceptual ontology is inadequate, proposes a substantially different one, develops it over years, defends it against alternatives and reorganises an entire field around it.

Yet we have substantial evidence that AI can generate novel solutions and explanatory hypotheses, weaker evidence that current systems autonomously invent radically new conceptual ontologies and essentially no basis yet for claiming that such ontology formation is structurally impossible for AI.

Measuring Creativity?

If by creative we mean something like producing novel, non-obvious and valuable solutions, then the claim that AI cannot be creative is already very difficult to sustain. The strongest examples are in domains where novelty and usefulness can actually be checked, rather than debated by art critics over a glass of warm prosecco.

AlphaEvolve

Genuinely new mathematics and algorithms.
This is probably the strongest current example because it is LLM-based. Gemini models generate modifications to programs, automated evaluators test them, and an evolutionary process develops promising ideas. In 2025, AlphaEvolve found a way to multiply 4×4 complex matrices using 48 scalar multiplications, improving on the best-known method in that setting dating back to Strassen’s work in 1969. Across more than 50 mathematical problems, Google reported that it improved the best-known result in about 20% of them. It also established a new lower bound for the 11-dimensional kissing-number problem by finding a configuration of 593 spheres. ⁶

FunSearch

An LLM contributing new mathematical knowledge.
DeepMind combined an LLM generating programs with an evaluator. On the long-studied cap-set problem, it found constructions that mathematicians had not previously discovered, including what the Nature paper describes as the largest improvement in the asymptotic lower bound in roughly 20 years. It also invented new heuristics for online bin packing. This is especially awkward for the “LLMs only remix their training data” argument – the resulting mathematical constructions were literally not in the training literature, because humans hadn’t discovered them yet. ⁷

AlphaTensor

Inventing better matrix-multiplication algorithms.
AlphaTensor searched an enormous algorithmic space and discovered provably correct algorithms that were more efficient than the previously best-known human algorithms for several matrix sizes. For example, where the conventional method for one matrix shape would require 100 multiplications and the previous human best used 80, AlphaTensor discovered one requiring 76. It also improved on a particular Strassen-derived result that had stood for around 50 years. ⁸

AlphaDev

Inventing algorithms humans actually adopted.
AlphaDev discovered previously unknown sorting routines at assembly level. Some were faster than heavily optimised human-designed algorithms, including a new instruction sequence or move that human programmers hadn’t used. The resulting algorithms were incorporated into the LLVM libc++ standard sorting library – the first changes to those particular routines in more than a decade. Human experts inspected the invention and then put it into production software used at enormous scale. ⁹

AlphaGo’s Move 37

Perhaps the cleanest intuitive example.
Against Lee Sedol in 2016, AlphaGo played a move that its own model estimated a professional human would choose only about once in 10,000 occasions. Expert commentators initially thought it was a mistake; it turned out to be pivotal to winning the game. Lee Sedol subsequently said: “Surely, AlphaGo is creative.”¹⁰ Move 37 was unusual – AlphaGo had learned from humans but then extensive self-play led it into a region of strategy that human Go culture had largely failed to explore.

And don’t forget Google’s AI Co-Scientist¹¹ that was mentioned earlier in this post. This evidence is newer and less definitive than the mathematical cases, but it points towards creativity involving hypothesis generation, not merely optimisation.

Footnotes

1. See books ‘Creativity and Art: Three Roads to Surprise‘ by Maggie Bowden and Caterina Moruzzi’s review of Boden’s criteria for creativity – novelty, surprise and value – in ‘Artificial Intelligence and Creativity‘ ↩︎
2. See Zahavy’s paper ‘LLMs Can’t Jump‘ ↩︎
3. See Tom Zahavy’s tweet. ↩︎
4. See Cell paper ‘AI mirrors experimental science to uncover a mechanism of gene transfer crucial to bacterial evolution‘ – Google’s AI Co-Scientist has generated hypotheses about drug repurposing and disease mechanisms that researchers subsequently tested experimentally. In one case it independently reconstructed an unpublished mechanism of gene transfer that another research team had already discovered experimentally but had not made public – other proposed drug-repurposing candidates showed effects in laboratory experiments. ↩︎
5. See Nature paper ‘Accelerating scientific discovery with Co-Scientist‘. ↩︎
6. AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms ↩︎
7. See paper ‘Mathematical discoveries from program search with large language models‘ ↩︎
8. See paper ‘Artificial intelligence finds faster algorithms for multiplying matrices‘ ↩︎
9. See paper ‘Faster sorting algorithms discovered using deep reinforcement learning‘ ↩︎
10. AlphaGo mastered the ancient game of Go, defeated a Go world champion, and inspired a new era of AI systems. ↩︎
11. Co-Scientist: A multi-agent AI partner to accelerate research ↩︎