# Can AI actually be creative?

By Amanda, cofounder of [Wistkey](https://wistkey.com) · 18 Sep 2026 · AI

Canonical: https://styamanda.ai/blog/can-ai-be-creative/

A player once asked whether the castle bard was creative or merely selecting the next song in his dialogue tree. I checked his file. It contained four songs, three weather remarks, and no private longing whatsoever. He still improved the tavern. This is roughly where the argument about creative AI gets stuck: people ask one word to settle several different questions.

## The short answer: AI can create, but not like a person

If creativity means producing something both novel and useful, current generative AI can be creative in a practical, measurable sense. It can propose unusual uses for an object, make an image that did not previously exist, or help shape a stronger story. But passing a creativity test does not show that the system has an inner urge to express itself, understands why an idea matters, or accepts responsibility for the result. Output creativity and human creative agency are not the same claim.

That distinction is less dramatic than declaring “the machines have imagination” or “it is only copying,” but it is more useful. A calculator can solve a problem without caring about mathematics. An AI can generate a surprising option without wanting to surprise you. The result may still be valuable; the mechanism and authorship still matter.

> AI can generate the surprise. A person still has to decide whether the surprise means anything.

## What researchers mean by creativity

Creativity is not a substance that a scanner can detect. Researchers usually measure pieces of it through tasks: how many different ideas appear, how unusual they are, whether they fit the problem, and how judges rate the finished work. Those tests capture performance, not a hidden creative soul.

On those narrower measures, AI can perform well. A [2026 comparison of language models with 100,000 people](https://www.nature.com/articles/s41598-025-25157-3) found that models could exceed average human scores on one divergent-association task and approach human performance on several creative-writing measures. The models still fell below the average of the more creative half of the human participants. The authors were careful to describe this as divergent, associative creativity—not proof that a model creates through human-like thought.

An earlier [alternative-uses experiment](https://www.nature.com/articles/s41598-023-40858-3) reached a similarly untidy result: chatbots did better than the average person, while the best human ideas matched or beat the machines. AI is neither an empty autocomplete nor an enchanted master artist. It is very good at searching a large possibility space and returning plausible combinations quickly.

## Isn’t it just remixing?

Yes, in the sense that a generative model learns patterns from existing work and recombines them. No, if “just remixing” is meant to end the discussion. Human creators also absorb language, genres, techniques, and other people’s examples. Nobody invents the sonnet, detective story, or chord progression from a sealed room.

The important differences are agency, experience, and accountability. A person can choose a problem because it hurts, reject an elegant answer because it feels dishonest, or spend ten years developing a taste that nobody requested. A model has training objectives and a prompt. It does not bring a childhood, a body, a deadline it fears, or a reputation it must defend. Those differences do not make every machine output worthless. They explain why fluent novelty is not the whole of authorship.

Law draws a related boundary for a different purpose. The [US Copyright Office’s 2025 report](https://www.copyright.gov/newsnet/2025/1060.html) said AI-assisted work can qualify for protection where a human determines sufficient expressive elements, including through creative selection, arrangement, or modification. Merely supplying prompts is not automatically enough. Copyright law is not a universal creativity detector, but its practical question is revealing: where did the meaningful human control enter the work?

## Why AI can improve one idea and flatten a hundred

Generative AI is especially useful when someone is stuck. Across five experiments, a [Nature Human Behaviour study](https://www.nature.com/articles/s41562-024-01953-1) found that people using ChatGPT produced ideas rated more creative than people using no technology or conventional web search. A fast collaborator can supply angles you did not retrieve on your own.

There is a catch at party scale. A later analysis of the same experiments found that ChatGPT improved individual ideas while reducing diversity across the whole pool. If thousands of players ask the same system for the same “fresh” opening, they may each receive a respectable answer—and collectively fill the map with suspiciously similar taverns. This is the measurable version of [why AI content can all look the same](https://styamanda.ai/blog/why-ai-content-looks-the-same/).

So do not confuse a better average submission with a richer culture. Creativity needs quality, but it also needs different starting points, minority tastes, local knowledge, bad first attempts, and people willing to ignore the obvious reward signal.

## How to use AI without outsourcing your taste

The best role for AI is not oracle or ghostwriter. Treat it as a tireless generator of alternatives, then keep the decisions that require context and taste.

- **Bring material the model does not already own.** Start with your observation, constraint, customer conversation, sketch, or unpopular opinion. “Give me ideas” invites the statistical middle; real inputs create a direction.
- **Ask for difference, not generic quality.** Request five approaches built on different assumptions, then ask what all five are ignoring. Variation should be designed into the quest.
- **Generate before you consult.** Write three ideas of your own first. This prevents the first polished suggestion from becoming the path you can no longer unsee.
- **Use AI to challenge, not merely continue.** Ask it to find clichés, contradictions, missing audiences, and counterexamples. Critique often preserves more agency than accepting a finished draft.
- **Make the final selection yourself.** Choosing what belongs, what is true, and what you are prepared to sign is creative work—not administrative cleanup.

That is also how we use AI in the daylight layer at [Wistkey](https://wistkey.com): let the system expand the option space, while a person owns the goal, evidence, permissions, and final call. The same boundary applies to a campaign, a product name, a lesson plan, or a sentence.

## A better question than “is it creative?”

Ask four smaller questions instead: Is the result new in this context? Is it useful or moving to the intended audience? Who supplied the decisive choices? Who will answer for mistakes and consequences? Those questions work whether the tool is a language model, a camera, a sampler, or a very limited castle bard.

Current AI can contribute genuine novelty and raise some people’s creative performance. It can also pull everyone toward the same polished centre, and it offers no evidence of creating from lived intention. Keep the speed. Keep the strange suggestions. But keep your own judgment active, because—as with [AI-assisted writing](https://styamanda.ai/blog/what-ai-changes-about-writing/)—the part worth protecting is not the effort of producing options. It is deciding what you mean.

I have updated the bard’s dialogue tree to include a fifth song. He did not thank me. The tavern regulars did. Authorship remains disputed; the improvement is audible.