Aesthetic Dysphoria: Asking For Directions in the Uncanny Valley
In the rush to be the brand name of artificial intelligence, risks underlie current models’ abilities to become ubiquitous in our society. One of the pervasive qualities of AI-generated images and film is the so-called ‘uncanny valley’ effect, loosely translated from the original Japanese bukimi no tani. This phrase originated with roboticist Masahiro Mori in 1970 to describe the sense of repulsion or aversion to objects and machines that are engineered to resemble or perform in a human-like fashion. While it is most often associated with visual media, it’s been well-documented in other fields as well, including humanoid robotics, and may even explain the emotional volatility directed toward self-driving cars.
One potential cause of this effect is a mismatch between stimuli and response that occurs in the parietal regions of the brain, the location of the mirror neurons, leading to a physical reaction within the sympathetic nervous system, raising cortisol and adrenaline levels in a fight, flight, or freeze response. This explanation suggests that there may be parallels with other involuntary responses to external physical stimuli like cognitive dissonance, and misophonia, where repetitive sounds or movement can trigger spontaneous anger and rage. Other threads connect it to conditions like anaphylaxis and Irukandji syndrome, where an acute allergic reaction or poisoning can cause feelings of impending doom before any physical effects present themselves. Strobe-induced epileptic seizures and arc flash, where welders experience a microsecond of high-intensity light before their visor darkens, leading to episodes of intense negative emotions are other examples that encompass similar processes where, like a black box, the output is not easily determined from input, but has an immediate effect in the real world.
A not insignificant amount of research and development has been invested into studying the effects of the uncanny valley in film and animation-3d modeling in particular, but there has been less focus on how these reactions relate to other forms of media using algorithms to approximate interpersonal observation and interaction. With AI image generators becoming integrated into social media, art, and design, there is reason to be concerned about the potential of stress-inducing qualities inherent in LLM-driven writing, speech, music, and chat models- possibly even with automated vehicles. Does regularly interacting with a generative pre-trained model instill a measurable sense of disquietude?
It might not be the most dramatic or dangerous example of generalized artificial intelligence backfiring, but the off-putting quality of the uncanny valley could be a substantial obstacle. If not analyzed and effectively countered, even the most advanced generative pre-trained models could soon go the way of Zoltar.
This project explores some of these questions by examining several datasets associated with the ChillsDB database (available here) a collection of self-reported responses to chills-inducing stimuli in media, including YouTube videos and Reddit threads. It is a fascinating collection recording the phenomena of chills induced by various media sources, with intensity, frequency, and other metrics documented rigorously across a sample size of over 600 participants. An in-depth discussion of the rigorous process used to determine the nature and differentiate between ‘goosebumps’ and ‘psychogenic shivers’ is available here.
In the initial phase of this project, I chose to focus on film, music, and speech to comb through each CSV to see if there were commonalities between different mediums, any ‘syn-aesthesia’, at work, such as the feeling of ‘frisson’ in music correlating to being moved by an inspiring speech, or a thread that resonated across age, gender, or social strata.
Upon digging deeper into the initial exploratory analysis, I found that patterns within each of the datasets were more compelling than the comparisons between them. I had not anticipated the measurable differences associated with pleasure difference and the positive versus negative effect of the chills-inducing stimuli. The concept of polarity played a crucial role in the story of how and why we subliminally respond in both sublime and spine-tingling ways to certain elements that register beyond merely observing the actions.
By performing some routine sentiment analysis, I validated this assumption to establish a working baseline between “good” chills and “bad” chills, it dawned on me that these preliminary functions could be relevant to current questions surrounding general artificial intelligence. Once developed further, they could conceivably be applied to studying how AI-augmented products like Stable Diffusion or Gemini affect their users. Determining the aesthetic impact of interacting with emerging technologies could assist in their adoption by society at large.
While the source material is not readily available, the pure expression of the responses allows for extrapolation and engineering of features that perform better than any individual feature run against predictive models. In the final combined dataset, a battery of predictive models for each of the sentiment features, along with quick visualizations reinforce this abstracted approach while illustrating an unexpected issue. Because the responses vary so widely (as in most human sample groups), the outliers make predictions more difficult, but including them gives a more accurate picture. Allowing for more data results in a more tuneable model, which leads to deeper, more fine-grained analysis. This analysis, in turn, serves to identify less apparent trends and temper confirmation bias. Keeping these variables gives a bigger picture that might have been lost if the data was pared down to only the information most readily identifiable as relevant.
Although this research might not be thrilling enough to cause many chills in and of itself, it does offer a peek through the keyhole at what constitutes an expanding universe of opportunities to examine the interface between human and machine minds. By extracting the ‘Chill Rating’ and other reproducible predictive modeling features, a decent standard can be compiled to compare against media augmented or created by AI. With enough testing, these models will produce outcomes that more closely resemble the actual data, allowing the effects of the uncanny valley, in theory, to be minimized or neutralized, and giving products that access these transformative technologies a chance to be more accessible and widely adopted.
Assuming there is a workaround, the next step is to examine other technologies that might fall under the same umbrella. If correlations exist between AI applications, they can be used to inform best practices as well as support ethical considerations. Some of the areas where AI can be readily applied happen to overlap with acquired skills and knowledge highly prized by humans, creating tension between developers of disruptive technologies and areas of enterprise and development that could be supplanted or supported.
Many engineers and programmers might say that AI is great for art and writing, but will never be a substitute for business acumen or where human safety is at risk. A sculptor or writer might take issue with machine learning tools' ability to create meaningful works of art but have no problem with self-driving transportation, and automated stock trading. There may be inherent differences in the way different brains are wired to experience adverse effects, or it could be attributed to a saturation of AI tools within the programming and developer community that lessens the response in general. Being able to map out the topography of our emotional responses to machines could help us have more empathy for the tools we use, and maybe even help develop future AI models’ emotional intelligence.
Every technology faces the challenge of being relevant, and of being replaced by something more relevant. A fundamental issue that could undermine the ability of our current models to be incorporated as functional tools is the barrier of overcoming aversion induced by the uncanny valley effect. This research presents the idea that by being able to determine the physical effects of the media we consume, we can engineer models that are more compatible with human psychology and proprioception by being more palatable to measurable metrics of aesthetics. Even if the best predictive models can’t foresee the future, hopefully, they can point us in the right direction.
Repository for notebooks including datasets can be found at:
ChillsDB url:
https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/ADLSZE
