It takes a lot of (other people’s) effort to look effortless: Fluent prose, vibecoding, and the future of deskilling

Mark Alfano (Macquarie University), Melina Tsapos (Lund University)

One of the paradoxes of the last half-decade has been the capacity of low-skill individuals to exploit the affordances of LLMs and other AI tools to give the appearance of knowledge, skill, and even intellectual virtue. We normally associate such appearances with an underlying reality because they are hard to fake and hard to conceal. In other words, we treat them as costly signals. The novice might make what seems like a brilliant move in a chess game, but two moves later it’s revealed that they have no idea what they’re doing. The expert might pretend to be a novice, but sooner or later the iron hand shows through the velvet glove. Recent advances in AI have made it harder to detect both false positives (cases where someone seems to have expertise but does not) and false negatives (cases where someone has expertise but it seems that they do not). This has reduced the costliness of the signal in both directions. For example, a recent study by Machery and colleagues showed that people found AI-generated poems more compelling than their training data (e.g., Chaucer, Shakespeare, Plath) until they were told that the slop they had rated highly was AI-generated. And a recent large study out of China found that students who were compelled to use AI as part of their coursework initially experienced gains and then plummeted by approximately 20% from baseline. In this paper, we discuss AI-facilitated deskilling at the ontogenetic and phylogenetic scales. We also explore the ways in which gradient descent and similar techniques are liable to produce prose, images, audio, video, and other outputs that unsuspecting and even suspecting audiences find more compelling than the flawed but authentic work of humans. We call back to Wittgenstein’s use of composite photography to argue that, while the mean or median representation may be most appealing in the short term, overuse of this heuristic is homogenizing in many spheres – from prose and poetry, to coding, to skill development more broadly. We conclude on a bleak note, pointing to the fact that synthetic outputs are now being used as training data for LLMs and other AI tools. This does not portend the singularity so much a singularly predictable loss of culture.