Human creativity in the age of text generators

Will synthetic text generators usher in a new age of creative thinking? The remarkable fluency of large language models may make them interesting tools for rapidly exploring semantic and stylistic spaces, yet the deceptive ease with which they generate output also provides countless new ways of appropriating ideas and erasing authorship. To better understand challenges and opportunities in this domain, we consider the interactional foundations of human creativity. Drawing on documented cases of creative collaboration, we argue that there is a world of a difference between text generation and true grasp — and that the friction and interaction that comes with the latter may be a meaningful element of human collaborative creativity.

TL;DR With synthetic text generators on the rise, there was never a better time to cultivate the artisanal and interactive roots of human creativity

Synthetic text generation, long confined to arcane sessions at computer science conferences, has gone mainstream.¹ In December 2022, a carefully orchestrated launch coupled with nominally ‘free’ access made OpenAI’s ChatGPT a darling of the press and the public. Already we can see a new wave of text generator-adepts. They think of themselves as having an edge, measuring their advantage in the speed at which they are now able to generate smooth prose and paper over the most blatant spots of ignorance. To be fair, for authors who know what they are doing, these tools may allow rapid exploration of semantic and stylistic possibilities. But will text generators herald a tectonic shift in how we work with ideas?

Text generators, by design,² make people feel they have found a thinking buddy: an ever-present, ever-patient companion that supports their ruminations, irons out the wrinkles of their prose and even helps them write computer code. Some seem to hope the text generator will be their Mileva Marić, brilliant collaborator and source of inspiration to her husband Einstein during his most productive years. Others perhaps wish for an intellectual companion like Émilie du Châtelet, whose influence is written all over Diderot’s famed Encyclopédie. Will clever prompts herald the next revolution in physics, and will text generators allow us to pen the next universal encyclopedia in record time?

What we know today about the scientific contributions of the likes of Marić and Du Châtelet is that they were rooted in close collaboration and conversation with their intellectual peers. Marić was part of a small cohort of brilliant physics students that also included Einstein; her mathematical skills greatly impressed him, and his correspondence bears witness to their intense collaboration. Du Châtelet formed a lifelong intellectual companionship with Voltaire, in which they ran a joint experimental lab, competed in essay competitions and even co-authored a book (crediting only him as named author).³ Many of the best ideas are forged in intense collaboration, heated debate and playful conversation.

---

It is tempting to think that text generators might play similar roles: to hope they may help us unlock a whole new level of human creativity.⁴ But text generators, by design, cannot get us there. A key reason is that they only have access to the calcified textual products of human creativity, not the actual process of situated creative collaboration. They are stochastic parrots,⁵ producing statistically likely text by regurgitating words to the tune of latent patterns in billions of words of training data. They may rehash basic textbook arithmetics, but won’t help lay the mathematical foundations for a new theory of relativity, as Marić was able to do.⁶ They may one day produce a passable translation of Newton’s Principia, but won’t produce a new synthesis that revolutionizes our understanding of the nature of energy, as Du Châtelet did.⁷

If anything, today’s text generators are more similar to the generations of bumbling male scholars who were happy to take credit for the work of their brilliant female peers.⁸ Except that text generators don’t even have the redeeming feature (if we can call it that) of being capable of original thought. Yes, they can produce novel arrangements of words that can be invested with meaning by eager interpreters⁹ — but to call that thought is the same as to attribute foresight to horoscopes. Text generators spin authoritative-sounding prose but have no sense of authorship and sunder all links of attribution. (Bropropriation is when a man takes credit for an idea first expressed by someone else.¹⁰ In this sense, text generators are the ultimate bropropriators.)

---

We are currently witnessing the first cohort of people tempted into treating text generators as thinking companions.¹¹ They are aided in this by the irresistible illusion of understanding that computer scientist Joseph Weizenbaum already saw applied to the proto-chatbot Eliza.¹² In time, many may come to depend on text generators to streamline prose, co-write computer code, and cover for gaps in their knowledge. The excitement of initial speed gains will obscure that this is at root a process of deskilling (with attendant opportunities for large-scale monetization).¹³ After all, there is a world of difference between text generation and true grasp. A mind propped up by a text generator is a mind populated by premade thoughts and prepared only for the most probable patterns. In contrast, grasp presupposes grappling, putting in the work that comes with acquiring a practiced skill. In this sense, text generators don’t aid original thinking but are the antithesis to it.¹⁴

Just as there is value in stammering in a foreign language on the way to greater fluency, so there is value in grappling with complex ideas and struggling to put them into words.¹⁵ Removing that friction doesn’t smoothen the creative process — it makes it slip away. Perhaps the most important lesson from synthetic text generators, then, is that they show us, in photo negative, what it takes to really think together. As Marić, Einstein, Du Châtelet and Voltaire demonstrate, human creativity flourishes in interaction. The miracle of human creativity is fundamentally a feat of human collective intelligence.¹⁶

How can we ensure that human creativity continues to flourish in the age of text generators? No amount of spinning the wheels of text generators will substitute for the power of recognising and harnessing the diversity of original thinkers around us. Amidst a rising tide of synthetic text, let us keep cultivating opportunities for co-present interaction, generous co-creation, and spirited debate.

Mark Dingemanse, Radboud University Nijmegen
Gijske de Boo, Rijnstate, Arnhem

1. Floridi, L. & Chiriatti, M. GPT-3: Its Nature, Scope, Limits, and Consequences. Minds & Machines 30, 681–694 (2020).
2. Chan, A. GPT-3 and InstructGPT: technological dystopianism, utopianism, and “Contextual” perspectives in AI ethics and industry. AI Ethics (2022) doi:10.1007/s43681-022-00148-6.
3. Fara, P. Pandora’s breeches: women, science and power in the Enlightenment. (Pimlico, 2004).
4. Erscoi, L., Kleinherenbrink, A. & Guest, O. Pygmalion Displacement: When Humanising AI Dehumanises Women. Preprint at https://doi.org/10.31235/osf.io/jqxb6 (2023).
5. Bender, E. M., Gebru, T., McMillan-Major, A. & Shmitchell, S. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜. in Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency 610–623 (ACM, 2021). doi:10.1145/3442188.3445922.
6. Troemel-Ploetz, S. Mileva Einstein-Marić: The woman who did Einstein’s mathematics. Women’s Studies International Forum 13, 415–432 (1990).
7. Zinsser, J. P. Translating Newton’s Principia: The Marquise du Châtelet’s revisions and additions for a French audience. Notes and Records of the Royal Society of London 55, 227–245 (2001).
8. Ross, M. B. et al. Women are credited less in science than men. Nature 608, 135–145 (2022).
9. Schlangen, D. Norm Participation Grounds Language. (2022) doi:10.48550/arXiv.2206.02885.
10. Bennett, J. Feminist Fight Club: A Survival Manual For a Sexist Workplace. (Penguin UK, 2016).
11. Rockwell, G., Berendt, B. & Chee, F. On IRIE Vol. 31: On Dialogue and Artificial Intelligence. The International Review of Information Ethics 31, (2022).
12. Weizenbaum, J. ELIZA—a computer program for the study of natural language communication between man and machine. Commun. ACM 9, 36–45 (1966).
13. Vallor, S. Moral Deskilling and Upskilling in a New Machine Age: Reflections on the Ambiguous Future of Character. Philos. Technol. 28, 107–124 (2015).
14. van Rooij, I. (2022). Against automated plagiarism.
15. Aakhus, M. & Laureij, L. V. Activity, materiality, and creative struggle in the communicative constitution of organizing: Two cases of communication design practice. Language and Dialogue 2, 41–59 (2012).
16. Dingemanse, M. et al. Beyond Single-Mindedness: A Figure-Ground Reversal for the Cognitive Sciences. Cognitive Science 47, (2023).

Note. This essay was first written in January 2023 and revised March 2023. Illustration by Suzan Heijink for Trouw (source).