The Mozart, Picasso and Einstein test: AI is already creative (and superhuman in some domains), but it isn't changing the rules of the game yet

THESIS: measured against Margaret Boden's classic framework, AI already exhibits real combinatorial and exploratory creativity —and in verifiable domains like mathematics or structural biology, at a level no human reaches— but transformational creativity, the kind of Mozart, Picasso and Einstein, the kind that doesn't solve the problem but redefines which problem is worth solving, remains human territory. And there is a second layer the debate usually ignores: artistic greatness is not just a property of the object produced, but a social act of intention, risk and meaning that, as of today, a machine cannot author. The honest answer to the test is neither 'yes' nor 'never': it is 'not yet, but the frontier moves faster than the contempt admits and slower than the hype sells.'

Let's start with the conceptual map, because without it the discussion collapses into anecdotes. Boden distinguishes three forms of creativity: combinatorial (mixing known ideas in new ways), exploratory (finding unprecedented solutions within a given style or rule-space) and transformational (changing the space itself: inventing new rules). The recent academic literature on language models converges on a diagnosis we share: LLMs master the first, reach the second when there is an evaluator to separate wheat from chaff, and have not demonstrated the third. This is no minor nuance: it is the difference between composing a fine quartet in the style of Haydn and being Beethoven deciding that the quartet can no longer sound that way.

Now the facts that force us to take the affirmative case seriously. In December 2023, FunSearch (DeepMind) produced new constructions for the cap set problem —the first time an LLM-based system made a genuine discovery on an open mathematical problem, with the largest improvement in twenty years to the asymptotic lower bound. In 2025, AlphaEvolve found an algorithm to multiply 4x4 complex matrices with 48 scalar multiplications, beating the record Strassen set in 1969: fifty-six years of human effort surpassed by an evolutionary agent with Gemini inside. And as documented by Terence Tao —arguably the most respected living mathematician— AI assistance has helped move on the order of a hundred Erdős problems into the 'solved' column within a few months, including cases Tao himself describes as resolved by the system with near-total autonomy and formal verification in Lean. This is not marketing: these are verified, published, reproducible results. In the exploratory space of mathematics, where validity can be checked mechanically, AI already creates things no human had created.

AlphaFold is the other pillar. The 2024 Nobel Prize in Chemistry to Hassabis and Jumper rewarded not a promise but a fact: predicting the structure of virtually all 200 million known proteins, an open database used by more than two million scientists in 190 countries, with real applications in antibiotic resistance, plastic-degrading enzymes and accelerating drug discovery for neglected diseases like Chagas. Here AI does not imitate a genius: it expands what the entire species can do. And this connects directly to our underlying line —the horizon of eradicating disease and extending life— without needing to exaggerate anything.

Our reading of all this has three layers. First: there is a clear pattern —AI shines exactly where an objective evaluation function exists. FunSearch, AlphaEvolve and AlphaFold are not 'models that dream'; they are generation engines coupled to relentless verifiers that discard the garbage. Creativity emerges from the generate-evaluate-select loop, not from an inner spark. That is enormously valuable —it is, in fact, a decent description of how much of science works— but it also explains why art, where there is no objective verifier, is a radically different case.

Second layer: in art, the data are both spectacular and misleading. GPT-4 scores in the top 1% of the Torrance Test of Creative Thinking, beating human controls in originality and elaboration (University of Arkansas, Scientific Reports, 2024). A University of Pittsburgh study (Scientific Reports, 2024) showed that non-expert readers cannot distinguish AI-generated poetry from that of Shakespeare, Dickinson or Plath —and tend to rate it higher. But beware the easy conclusion: the authors themselves note that readers prefer AI poems because they are simpler and more accessible, mistaking the complexity of human poems for incoherence. In other words, the AI 'triumph' partly measures the comfort of average taste, not an artistic peak. And Torrance tests measure divergent potential in lab conditions, not the creation of a work that reorganizes a field. That AI wins on statistical fluency and originality does not mean it wins at what makes Picasso great.

Third layer, and for us the decisive one: artistic greatness lies not only in the artifact but in the act. Here the first-rank voices matter. Boris Eldagsen won the 2023 Sony World Photography Award with an AI-generated image and rejected the prize —not to humiliate the machine but to force the debate: 'AI is not photography,' he said, and the two should not compete in the same category. Nick Cave, faced with a ChatGPT-generated song 'in the style of Nick Cave,' called it 'a grotesque mockery of what it is to be human,' with an argument worth taking seriously even if one doesn't fully share it: songs 'arise out of suffering,' and the algorithm 'has been nowhere, it has endured nothing.' We shouldn't idealize this stance —the romantic myth of the tormented artist is also a myth— but it points to something real: we value art in part because it is a message from one consciousness to another, a testimony of having lived. AI produces the object; it has lived nothing. And when a photography or fiction prize is given to an AI work presented as human, what breaks is not the quality of the pixel but the contract of intention that underpins the value.

There is also a collective risk a serious optimist cannot stay silent about. The Doshi and Hauser study (Science Advances, 2024) showed something uncomfortable: generative AI improves each individual writer's creativity but homogenizes the whole —AI-assisted stories resemble each other more. It is the paradox of our time: each person becomes a little more 'creative' and the culture, in aggregate, a little more monotonous. If Mozart, Picasso and Einstein matter it is precisely because they diverged from the average; a technology that pushes toward the mean could, misused, make it harder for the next outsider to appear. The good news —and follow-up research confirms it— is that this depends on how the tool is deployed, not on an inevitable limitation: with diverse prompts and personas, variety is preserved. The problem is one of design and market, not destiny.

IMPLICATIONS. In the short term, the honest thing is to acknowledge the disruption: mid-tier creative professions —commercial illustration, ad copywriting, stock music, catalog photography— are already suffering substitution, and the conflicts over authorship, rights and training data are real and unresolved. We don't sugarcoat it. But it's worth separating panic from trend: as in other sectors we've analyzed, the routine and reproducible is the most exposed; judgment, a distinctive voice and the intention to say something resist better. In the medium term, the most likely scenario is not the autonomous AI-artist dethroning the geniuses, but the augmented creator: the mathematician with a copilot who explores in hours what once took years, the musician using the model as an idea bank, the scientist delegating the search and reserving the question. Transformational creativity —deciding what is worth asking— remains the human bottleneck, and it is precisely the part we are least eager to cede to anyone.

In the long term, and here is our qualified optimism: AI's capacity to make verifiable discoveries —proteins, algorithms, theorems— is one of the most powerful levers ever invented for an abundance of knowledge. If that frees human time from the routine to devote to what truly excites us —including artistic creation without economic pressure— the balance is deeply positive. But let's not confuse the planes. AI has passed the Einstein exam in its 'solve hard problems within given rules' version, and is on track to pass it ever better. The Picasso exam —break the rules and convince the world the break matters— is still pending, and it isn't clear it is an exam a machine, having lived nothing, could even want to take. Our conclusion, without hype or contempt: AI is already creative in a real and sometimes superhuman sense; it is not yet a genius, because genius is not only producing the new and valuable, but deciding what is going to matter. That decision, for now, we still make. And that is no small thing.