AI can produce genuinely novel combinations, but the honest answer is that it recombines patterns from its training data rather than creating from lived experience — which means it can be impressively creative in some tasks and fundamentally limited in others.
The distinction matters because it tells you when to trust an AI's output and when to reach for a human. A tool that generates a fresh marketing slogan is doing something different from a tool that writes a memoir about grief, even if both are called "creative AI." Understanding which is which saves you from disappointment and from over-trusting output that sounds original but isn't.
The mechanism, stated plainly
Most generative AI systems work by predicting what comes next in a sequence — given a prompt, they estimate the most likely continuation based on patterns absorbed during training. That's the writer's framing of a widely discussed mechanism, not a claim from a single citable study, and it's worth flagging as such.
The practical consequence is that the model is doing interpolation: it navigates a space of combinations it has seen, weighted by how often those combinations appeared. When you ask for a poem about autumn, it draws on thousands of autumn poems and produces something that sits plausibly within that space.
It is not retrieving a stored poem verbatim, which is why outputs feel new. But it is also not inventing a category that didn't exist. According to the AI-Mind AI Tool Database, which tracks 360 AI tools with capability snapshots recorded at verification time, the tools that score highest on "creative" tasks are typically those with the broadest training exposure — more material to recombine. That's a pattern, not proof of creativity.
The contested middle
Some of this is genuinely unsettled. There's a real philosophical argument, associated with thinkers like Margaret Boden, that combinational creativity — making unfamiliar combinations of familiar ideas — is a legitimate form of creativity, not a lesser imitation. Under that view, AI is creative in the same way a collage artist is.
Others, including many working artists, argue that creativity requires intentionality and stakes: you have to care about the outcome, and you have to risk something. I won't pretend there's a consensus, because there isn't. What I can tell you is that the disagreement is about definition, not about what the tools do. Both sides agree the tools recombine. They disagree about whether recombination counts.
A concrete example
Say you run a small bakery and you want a name for a new sourdough loaf. You ask an AI for twenty options. It returns things like "Midnight Crumb," "Hearth & Honey," "The Slow Rise."
These are novel strings — none of them existed before — but each is assembled from familiar bakery vocabulary. For this task, that's fine. You're judging the output by whether the arrangement sounds appealing, and you can pick the one that fits your brand.
Now say you want the story on your packaging about why you started baking. The AI can produce grammatical, warm-sounding prose. But it has no access to the actual reason — your grandmother's kitchen, the year you spent unemployed, the loaf that finally worked.
It will invent something plausible and generic. For that task, recombination fails, because the value is in the lived specifics, not the arrangement.
Where recombination fails, and the decision rule
Here's the criterion I'd actually use. Ask yourself: is the output judged by the novelty of its arrangement, or by its connection to lived stakes? If it's arrangement — slogans, brainstorming, code patterns, chord progressions, alternative phrasings — AI is often genuinely useful, and the recombination is a feature.
If it's lived stakes — testimony, memoir, original research, a eulogy for someone you knew — recombination produces something that sounds right and is hollow. The failure mode isn't that the AI lies; it's that it fills the gap with plausible-sounding filler, which is the same mechanism behind AI hallucinations.
There's a related trap: AI creativity tends to cluster around the average of its training data, so outputs can feel samey across many attempts. If you ask for fifty slogan ideas, the first ten may be strong and the next forty increasingly interchangeable. The fix is to constrain the prompt hard — give it a specific audience, a tone, a forbidden word list — because constraints push it away from the statistical centre.
That's a real technique, not a trick. The limitation is that it still won't give you the grandmother's kitchen. For that, you're the only source.