Part of Temporal Constructs Appendix, which contains details behind the research: dates, quotations, descriptions of instruments, and sources.
General definition
A study by Anil Doshi and Oliver Hauser in Science Advances (2024) found that generative assistance raises the novelty of an individual’s work. The same study found that generative assistance reduces the diversity of what a group of people produce.
Relation to the practice
It is one of the two studies the account’s weakest position rests on. It supplies the distinction that position turns on: the harm it documents belongs to open-ended delegation, in which a person directs the collaboration.
Abstract
Creativity is core to being human. Generative artificial intelligence (AI), including powerful large language models (LLMs), holds promise for humans to be more creative by offering new ideas, or less creative by anchoring on generative AI ideas. We study the causal impact of generative AI ideas on the production of short stories in an online experiment where some writers obtained story ideas from an LLM. We find that access to generative AI ideas causes stories to be evaluated as more creative, better written, and more enjoyable, especially among less creative writers. However, generative AI enabled stories are more similar to each other than stories by humans alone. These results point to an increase in individual creativity at the risk of losing collective novelty. This dynamic resembles a social dilemma: with generative AI, writers are individually better off, but collectively a narrower scope of novel content is produced. Our results have implications for researchers, policy-makers, and practitioners interested in bolstering creativity.
Findings of the study
The experiment is preregistered and runs in two phases. In the first phase, 293 writers each wrote an eight-sentence story for a teenage and young adult audience. They had been assigned at random to one of three conditions: no assistance, the option of one story idea from GPT-4, or the option of up to five. Of the writers offered an idea, 88.4 per cent asked for at least one. In the second phase, evaluators read the stories without knowing which condition produced them. They rated novelty, usefulness and a set of emotional characteristics.
Stories written with an idea scored higher on novelty, on being well written and on being enjoyable. The gain grew with the number of ideas: 5.4 per cent more novel with one idea and 8.1 per cent with five. The benefit fell unevenly. Writers who scored lower on a divergent association task beforehand gained most. This is the finding the paper draws on when it says the tool narrows the gap between more and less creative people.
The cost appears in a different measure. Comparing the stories to one another through OpenAI embeddings, the authors found the assisted stories closer to each other than the unassisted ones. That closeness held both within a condition and against the ideas the model had supplied. Individual work improved, and the range of what the group produced contracted. The paper describes this as a social dilemma.
Reading notes
The reading notes welcome the research, and hold that tests of this kind will have to account for a person’s previous experience with generative AI as the technology is adopted. They state that every technology changes the ways people think as those people become more experienced with it. They hold that neutrality cannot be assessed beforehand unless the individual experience and its processes are measured. They note that this becomes harder to measure as integration increases. They state that the uniqueness of each individual may lie in that person’s approach to AI use.
The notes state that learning to steer the process towards a particular outcome, with varying levels of thinking with and thinking for, is required to gain this experience. They describe this as an experimental mindset, with no set method. They define the experimental mindset as testing the technology and pushing into new territory, trying different things on each occasion of collaboration with AI. They compare playful interaction and experience with this to learning to play an instrument, which may affect how people engage with large language models.
The study was from 2024, and by 2026 chatbots have become more common in daily use, giving users more experience, both negative and positive. The notes state that this experience leads to a particular kind of understanding. They state that whether this is the understanding needed for appropriate steering is unknown, and hard to obtain through measurement. They applaud the existence of such a study and further research, and hold that testing becomes harder as the technology becomes an extension of the person.
They state that the extension is always an extension of the person, so a pure AI outcome cannot be isolated from human experience. They add that the human is inevitably biased by the AI in return. They hold that how far this leads, and whether it is good or bad, is hard to measure. They hold that a lack of negative experience, with AI outputs that people have reviewed critically, produces the increased homogeneous output. The more people engage critically with AI-generated outputs, the more they can find a particular kind of human extension in which AI bias remains present but differs between people.
They state that the difference comes only from the nature of the technology, which changes rapidly, faster than human nature. They hold that human nature being part of this makes a sample of a few hundred people insufficient. They find it problematic to engage with this topic from a scientific standpoint, because self-expression may evolve in more complex ways than a linear process that all people follow equally. They state that commercial success can be measured, and artistic integrity cannot be measured in the same way.
Verification
Read in full on 7 November 2026 from the PDF in the library (Zotero item doshiGenerativeAIEnhances2024, 9 pages). The abstract is quoted with its dashes replaced by commas. The archive has no em dashes. The wording is otherwise unchanged.
The figures the summary rests on are on the pages it cites. Page 2 has the three conditions, the 293 writers and the 88.4 per cent who called on the model. The 5.4 and 8.1 per cent gains in novelty are on p. 3, where the authors also report the larger gains for less creative writers. The similarity result, measured as cosine similarity of OpenAI embeddings, is on p. 4.
What the source does not support: the study measures one sitting with GPT-4 in 2024, so it says nothing about writers who have used these systems in the two years that followed. Its diversity measure compares stories written to one brief in one session. The reading notes above turn on that gap, and the paper does not address it.
Further reading
Clark & Chalmers 1998 “The Extended Mind” : because his central
objection is that the technology becomes an extension of ourselves, making “pure AI outcome ” impossible to isolate. That’s exactly the extended mind thesis, and it provides a useful vocabulary (coupled system, active externalism) plus a counter-argument (the “cognitive bloat”/ “coupling-constitution fallacy” from Adams & Aizawa) that could sharpen the claim.