There is a reassuring argument about AI: it learns from the songs we write, the pictures we paint, and the things we say. Everything it produces must therefore be imitation and recombination. The original creator will always be ahead; the imitator will always follow.
Creators might breathe a little easier. At least creativity seems safe.
But I would ask you to pause.
If learning from existing work rules out anything genuinely new, human creativity faces the same objection. People also learn from languages, works, and experiences that precede them. The origin of the learning material does not settle what can happen after learning.
Chapter 1: The Library of Babel↗
Imagine a library containing every possible book. Not every book already written, but every book that could be written within a fixed alphabet and length.
In Borges’s The Library of Babel, each book has 410 pages, each page forty lines, and each line about eighty characters. Its alphabet contains twenty-two letters, a space, a comma, and a full stop. Fixing those quantities gives:
Finite, but almost impossible to imagine.
Somewhere in this library would be a text claiming to describe your entire life, including the years you have yet to live. Other texts would claim to prove the Riemann hypothesis, explain consciousness, or cure cancer. If a correct answer can be written within the allotted length, it too would be there. So would its apparent refutation, countless wrong answers, and meaningless strings.
Containing every text does not tell a reader which one is true. It does not guarantee that every question has an answer.
When a human writer writes a book, then, is she creating or discovering?
The combinations can be defined in advance. Writing still involves finding, selecting, and realising a particular combination—and making it worth reading. Babel unsettles the idea that creation must come from nothing. It does not establish that humans and AI create in the same way.
Using existing material does not, by itself, decide whether a work is creation or imitation.
Chapter 2: Where does Pegasus come from?↗
Close your eyes and imagine a horse. Now give it wings.
You have just “created” Pegasus. But where did the horse come from? Horses you have seen, pictures, other people’s descriptions. And the wings? Birds, butterflies, perhaps aircraft. A new image does not require every part of it to be new.
In Section II of An Enquiry Concerning Human Understanding, published in 1748, Hume describes mental creation in terms of combining and transforming material supplied by sensation and experience. This is a philosophical account of the origin of ideas, rather than a scientific theorem that exhausts human creativity.
Unicorns borrow horses and horns; many imagined ideal worlds borrow from earthly life. Such examples suggest that recombination and creation need not be opposites. They do not establish that all gods resemble humans, or that every new concept is a simple assembly of familiar parts.
The question remains: why count some human recombinations as creation while excluding AI’s in advance?
Perhaps the difference is meaning, emotion, or intention, whereas AI performs statistical operations. Those differences deserve discussion. First, though, consider a narrower question: can AI produce outputs different from its training examples? Calling its operation statistical does not supply a negative answer.
Chapter 3: Combinatorial explosion and an unfinished map↗
3.1 Finite can still be vast↗
How many ways can two hundred distinct elements be arranged in a row?
A 375-digit number. Both the elements and the sequence length are finite, yet exhaustive enumeration is already impractical. Infinity is unnecessary: combinatorial explosion makes clear why a finite training set need not restrict a model to reproducing a finite list of examples.
Generative models can learn patterns in language, melody, or images and use them in new outputs. Learning regularities and memorising examples can also coexist. Carlini and colleagues extracted training images from several diffusion models. Sampling, therefore, is no guarantee against copying.
Not every generation is retrieval; not every generation is original. Whether a particular work closely resembles training material requires a separate check, especially when the training data are unavailable.
3.2 A continuous space is not a road everywhere↗
Consider another analogy. Having seen only red and yellow, how might someone understand a colour between them? The example gestures towards variation beyond discrete assembly. But pigments, light, and digital colour coordinates mix differently; “red plus yellow makes orange” is not a universal law of interpolation.
For a generative model with continuous latent variables, intermediate representations can be constructed and decoded. A simple linear interpolation is:
Here, are two latent representations and is a chosen decoding map. This gives a way to construct candidate outputs. It does not guarantee that every point along the path yields something coherent, meaningful, or new. Distances in different models need not express the same semantic similarity, and not every LLM is a single map of complete works with such a decoder.
AI might be doing more than rearranging blocks. It might be exploring a map on which training data mark some locations while others remain to be found. Some paths lead to new works; others lead to noise. A vast map is not evidence of artistic quality.
Chapter 4: Are those possibilities worth anything?↗
A large space of possibilities does not establish that its contents are good creations. That calls for concrete achievements—and attention to what each achievement shows.
4.1 Proteins: prediction is not confirmation of every structure↗
Half of the 2024 Nobel Prize in Chemistry went to David Baker for computational protein design; the other half went jointly to Demis Hassabis and John Jumper for protein structure prediction. The researchers received the prize, rather than AlphaFold itself.
AlphaFold extended prediction to roughly two hundred million protein sequences. That does not amount to two hundred million individually confirmed structures. Comparisons against experimental data found many highly accurate predictions, alongside discrepancies even in some high-confidence ones.
Its value does not require universal correctness. Useful structural hypotheses that can be tested further already change where research can begin. Knowledge grows through the combination of prediction and experiment.
4.2 Knots: machine-generated clues, human proofs↗
A 2021 collaboration between DeepMind and mathematicians used machine learning to explore relationships between geometric quantities and algebraic invariants of knots. Models and attribution analyses supplied clues; mathematicians investigated natural slope and signature and established a new theorem.
AI did not independently write the proof, and connections between knot geometry and algebra were not wholly unimagined before this work. A particular new relationship was clarified through machine guidance and human reasoning. Existing material could still lead to a previously unknown mathematical result.
4.3 A painting: there is still a process behind the prize↗
In 2022, Jason Allen’s Midjourney-assisted Théâtre D’opéra Spatial won an art competition at the Colorado State Fair. The administrative record also describes repeated prompting, image selection, and subsequent editing.
This shows that a creative process involving generative AI can receive artistic recognition. It does not establish that an autonomous system defeated human art under controlled conditions. A competition result cannot substitute for a systematic comparison of judging conditions and creative contributions.
4.4 Music: what does failing to identify the source mean?↗
In 2025, a Deezer-commissioned Ipsos survey covered nine thousand adults across eight countries. Participants heard two AI songs and one human song; the release reported that 97% failed the source-identification task.
This does not mean that every participant lacked any ability to classify any song. Nor does it establish a general musical Turing test. It is a result for particular tracks and a particular task, rather than a direct measure of lasting aesthetic value, originality, or creative intention.
It remains worth taking seriously: a work’s origin is not always immediately audible. But identifying who made it and understanding why it is worth hearing are different questions.
These examples need not become a single scoreboard of AI victories. Scientific prediction, mathematical discovery, artistic judgment, and source identification ask for different evidence.
Learning from existing works is no longer a sufficient reason to declare that AI can never contribute to creation. Which abilities belong to the system, which depend on human selection and verification, and what this changes about human distinctiveness remain questions worth pursuing.