Catherine Wearing is Associate Professor of Philosophy and Director of the Cognitive and Linguistic Sciences Program at Wellesley College. She works chiefly in philosophy of mind and philosophy of language, with particular interests in figurative language and the imagination.
A post by Catherine Wearing
Recent gains in the capabilities of generative AI systems pose a dilemma for our thinking about creativity. On the one hand, some of the products of these systems seem pretty creative. Some of the images generated by Midjourney and stories generated by ChatGPT are good enough that, had they been produced by humans, they would count as creative. Similarly, systems like AlphaFold (which predicts the 3D structures of proteins) and AlphaGo (which plays the boardgame Go to a level of proficiency sufficient to beat human grandmasters) produce results which have some claim to count as creative. On the other hand, most people are reluctant to grant that these AI systems themselves are creative.
One tempting reaction to this clash of intuitions is to divorce the criteria for a creative product from those for a creative producer. If there’s a way to acknowledge the creativity of a product while denying creativity to its producer, then we can respect both of the intuitions which generative AI systems provoke. This is the strategy I want to explore here.
Here is one way we might separate creative products from creative producers: first, adopt something like the standard characterization of a creative product as something which is novel, surprising, and valuable. This characterization looks as though it will cover those products of generative AI which we judge to be creative. Second, specify a necessary condition on being a creative producer which is independent of the features of the standard characterization of creative products and which rules out any (current) generative AI system.
Candidates for such a necessary condition have included: being conscious (Hoel 2022), being an agent (Moruzzi 2022), having autonomy (Boden 2010, 2014), and having meta-cognitive awareness (Kind 2022). The absence of any one of these properties would allow us to disqualify generative AI systems from counting as creative producers, for as long as these systems lack that feature. However, it’s arguable that none of these conditions is truly essential to human exercises of creativity. If, for example, some instances of human creativity spring from the subconscious, then consciousness might not be a necessary condition on creative production after all. (See Langland-Hassan (forthcoming) for a fuller discussion of this line of objection.)
Whether or not these are necessary conditions for being a creative producer, it’s noteworthy that none of them has anything specifically to do with creativity itself. To approach creativity more closely, we need to consider what the producer is doing. Properties like agency and consciousness don’t begin to capture the idea that it matters how an object is produced, even if its creator is a conscious, autonomous, and meta-cognitively aware agent. In other words, details of the creative process matter too. For example, we generally judge that at least two kinds of process are not creative: the utterly random and the overly ‘automated’ (in the sense of following a recipe or working through a list).
Let’s shift, then, to considering whether our opening goal can be achieved by seeking independent characterizations of creative products and creative processes. I think there are two hurdles facing this objective. The first problem has to do with the sorts of ‘disqualifying’ processes just mentioned, and specifically, with the impact that learning of the involvement of such a process can have on our judgements about the creativity of what is thereby produced. There are at least three cases in which details about how a thing was produced tend to weaken or overturn our judgement that the thing itself is creative:
objects created by natural, but non-living, means: If we discover that what we took to be a creative sculpture is in fact a product of wind, rain, and time, we rescind our judgement that it is creative.
objects created entirely randomly: Consider a painting produced by someone flailing in the dark (cf. Gaut 2003). It might be a ‘stunningly good abstract painting’, but it’s not creative. [Gaut exploits this judgement to support introducing a feature of the process of creation (which he calls ‘flair’) as a constraint on the creativity of the product, while Langkau (2022) allows that the painting may be creative while its producer is not.]
objects produced by means of a systematic search through an exhaustive list of possibilities or by following an existing recipe: If the process is too much like following an already specified procedure or systematically checking entries on a list, then our inclination to say that the result is creative wanes.
If the creativity of a product is independent of the process by which it was created, then these revisions in judgement don’t make sense. Insofar as they do make sense, they suggest that the criterion for a creative product is tied to the process by which it was created. It's open to us, of course, to stand our ground and overrule the judgement that some processes undermine the creativity of their products. But there’s another problem in the offing, namely, that it’s not clear that the standard characterization of a creative product is in fact independent of details regarding the process by which that product was created.
Consider the condition that a creative product must surprise us. Langkau (2022) rightly worries that Boden makes a mistake when she ties different kinds of creativity to differences in the phenomenology of surprise, for our experiences of surprise may vary in unpredictable ways. But for present purposes, what matters is the way Boden distinguishes the various kinds of surprise from one another, namely, by looking at what causes them. And this, Boden suggests, is the recognition of various kinds of processes: “making unfamiliar combinations of familiar ideas”, “exploring a conceptual space”, and “tweaking or even radically transforming a conceptual space” (Boden 2010; we see a version of this same three-way distinction in much earlier work as well, e.g. Boden 1994). In practice, we may know nothing about the process by which something was created, but when we make the judgement that it is creative, we’re making assumptions about what kind of process seems to have been involved. This explains our sense of deflation when we learn that the outcome was in fact a matter of chance or mechanical search.
The worry, then, is that the standard characterization doesn’t identify a creative product independently of details about the process by which it was created. What are the prospects for specifying the creative process without attending to its products? The situation in this direction looks more promising. We’ve already seen how certain processes (which can be specified independently of their products) sit outside what we judge to be creative. Consider next some of the features which psychologists have identified as relevant to creativity: flexible thinking (pulling in lots of diverse ideas, not perseverating on a single perspective); being able to generate many ideas or possible solutions, rather than just a few (sometimes called ‘fluency’); a well-developed capacity for pattern recognition (especially of novel patterns); as well as more general traits such as perseverance, attention to detail, and patience. These traits have been identified by working backwards from cases which have resulted in creative products, so in that sense, they are shaped by their relation to their results. But the features themselves aren’t defined in relation to their products; indeed, there’s no assumption that they will lead to creative products in every instance, only that they tend to. So they have an independence that the standard characterization of a creative product seems to lack.
What does this mean for the opening dilemma? If we can’t characterize creative products independently of creative processes, then we can’t reconcile our conflicting intuitions. Insofar as it looks more plausible that creative processes can be defined independently of their products than the other way around, we should deny the first horn of the dilemma and hold that the products of generative AI aren’t creative. They’re all kinds of good things – original, impressive, perhaps even something no human could have achieved – but they’re not creative.
References
Boden, Margaret A. 1994. “What is Creativity?” in Dimensions of Creativity, M. Boden, ed. MIT Press.
Boden, Margaret A. 2010. Creativity and Art: Three Roads to Surprise. Oxford University Press.
Boden, Margaret A. 2014. “Creativity and Artificial Intelligence: A Contradiction in Terms?” in The Philosophy of Creativity: New Essays, E. Paul and S.B. Kaufman, eds. Oxford University Press.
Gaut, Berys. 2003. “Creativity and Imagination” in The Creation of Art: New Essays in Philosophical Aesthetics, B. Gaut and P. Livingston, eds. Cambridge University Press.
Hoel, Erik. 2022. “AI-art Isn’t Art” The Intrinsic Perspective, May 18, 2022.
Kind, Amy. 2022. Imagination and Creative Thinking. Cambridge University Press.
Langkau, Julia. 2022. “Two notions of creativity” in Wittgenstein and beyond, C. Pfisterer, N. Rathgeb, and E. Schmidt, eds. Routledge.
Langland-Hassan, Peter. Forthcoming. “Creativity, Imagination, and Artificial Intelligence” in The Oxford Handbook of Philosophy of Imagination and Creativity. A. Kind and J. Langkau, eds. Oxford University Press.
Moruzzi, Caterina. 2022. “Creative Agents: Rethinking Agency and Creativity in Human and Artificial Systems” Journal of Aesthetics and Phenomenology, Vol. 9.2.