AI and Creativity — Can Machines Be Genuinely Creative?
GPT-4 made 293 writers look more creative one story at a time, then made their stories more alike as a set. Doshi and Hauser’s 2024 experiment captures the central problem: a machine can raise the floor of creative production while lowering the variance from which new forms emerge.
Two tests hiding inside one word
Creativity can describe an artifact or its maker. Those claims require different evidence:
| Test | Question | What present models demonstrate |
|---|---|---|
| Artifact | Is the result novel, useful or surprising? | Often, within bounded tasks |
| Process | Did the system search, revise and select? | Partly observable |
| Agency | Did it intend the result or care whether it succeeded? | Unknown |
Haase and Hanel’s 2023 divergent-thinking experiments found chatbots performing near human levels on tasks such as proposing unusual uses for ordinary objects. That supports artifact-level creativity. It does not settle concept chinese room or concept hard problem consciousness because neither novelty nor fluency establishes experience.
Anthropic’s 2025 attribution-graph work traced causally active internal features during operations including planning and translation. The result makes “mere phrase retrieval” a poor mechanical description, but a computational graph is not evidence that anything inside the model feels surprise.
The diversity trap
Doshi and Hauser assigned writers no AI assistance, one GPT-4 idea or five GPT-4 ideas. Assisted stories received higher novelty and usefulness ratings, with the largest gains going to writers who scored lower without assistance. Semantic comparisons also found that the assisted stories resembled one another more closely.
The loop changes the unit of analysis. One writer gains options; 10 million writers drawing from related distributions may inherit correlated defaults. The danger is not repetition word for word. It is convergence on the same structures, metaphors and acceptable surprises.
Three kinds of machine creativity
Margaret Boden’s taxonomy separates three operations that “creative” tends to blur:
| Type | Operation | AI evidence as of 2026 |
|---|---|---|
| Combinational | Join familiar elements | Common |
| Exploratory | Search within existing rules | Strong in games, images and music |
| Transformational | Change the rules defining valid work | Contested |
A model can combine a Bach chorale, a tabla cycle and a synthesizer patch. The harder case is a new musical grammar whose first appearance sounds wrong, as bebop did to some 1940s listeners, but later changes what musicians can hear. Assumption: preference-trained systems may discard such errors before a culture has time to recognize them as inventions.
What’s contested
John Searle’s 1980 Chinese Room separates correct symbol manipulation from understanding. Functionalists dispute that separation: if intention, revision and judgment arise from organized information processing in brains, they ask why comparable functions would be disqualified in another substrate.
Transformational creativity presents a second dispute. A machine may fail to alter its conceptual boundaries, or observers may reserve “genuine” creativity for artifacts attached to biography, struggle and social risk. The disagreement concerns both machine capacity and the human rules for assigning authorship.
The cultural evidence remains thin. A 2024 short-story experiment cannot tell us whether decades of assisted writing will compress culture into familiar forms or let millions of technically untrained people express ideas that otherwise remained private.
Why this has to do with other realms
The problem resembles biological monoculture, with one crucial difference: stories mutate whenever readers interpret them. Even so, dependence on a few model distributions creates correlated cultural failure. concept svalbard seed vault stores crop variance against catastrophe; no equivalent archive deliberately preserves rejected drafts, eccentric forms and low-probability ideas before ranking systems bury them.
Claude Shannon’s 1948 concept information theory gives the bridge a number: an event with probability (p) carries (I=-\log_2 p) bits of information. More predictable output carries less surprise. A culture optimized for immediate plausibility may become easier to consume precisely as it becomes easier to predict.
Key Sources
- Margaret A. Boden, The Creative Mind: Myths and Mechanisms (1990; second edition, 2004). Introduces combinational, exploratory and transformational creativity.
- Anil R. Doshi and Oliver P. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content” (2024), Science Advances 10, eadn5290. Reports the 293-writer experiment.
- Jennifer Haase and Paul H. P. Hanel, “Artificial Muses: Generative Artificial Intelligence Chatbots Have Risen to Human-Level Creativity” (2023), Journal of Creativity 33, 100066. Tests divergent-thinking performance.
- John R. Searle, “Minds, Brains, and Programs” (1980), Behavioral and Brain Sciences 3, 417–457. States the Chinese Room argument.
- Anthropic, “Circuit Tracing: Revealing Computational Graphs in Language Models” (2025). Examines causally active features inside a language model.
- Claude E. Shannon, “A Mathematical Theory of Communication” (1948), Bell System Technical Journal 27. Defines information through probability.
Further Reading
- Arthur I. Miller, The Artist in the Machine (2019). Follows AARON and other machine artists before large language models.
- Douglas Hofstadter, Gödel, Escher, Bach (1979). Tests whether formal symbol systems can acquire meaning.
- concept outsider art: asks what creativity looks like beyond professional institutions and shared taste.
- concept frisson: examines whether knowing an artifact’s author changes the body’s response.
- concept default mode network: follows the brain networks associated with spontaneous idea generation.
See Also
- concept generative art
- concept outsider art
- concept chinese room
- concept frisson
- concept hard problem consciousness
- concept embodied cognition
- concept information theory
- concept svalbard seed vault
Abhishek's take
I use models to widen a draft, then distrust the first clean answer. Their most consequential bias may be convenience: the plausible sentence arrives polished enough to stop the search. If every assistant recommends the same sensible next page, which unsuggested page does this wiki most need?
Tags: #ai #creativity #divergent-thinking #generative-art #consciousness #originality