Given the reviews I got back on a recent manusript, it seems that people are having a hard time grasping the main point of a recent study I did. I thought it might be a useful exercise to write a concise version of that result here and see if I can get the point across.

Humans often get enjoyment from visual experiences, and can readily express preferences for some images over others. Whether in an art museum or driving down a highway, some things we see grab us more than others. The scientific study of preference, while extending back over centuries, has recently seen a renaissance using newer methods and based on a firmer understanding of the brain processes underlying visual processes and emotional responses.

In many of these recent studies, investigators have sought to understand whether particular aspects of visual material can reliably predict what people like to look at. This tendency to average results across observers goes against the old adage that "beauty is in the eye of the beholder," stressing group norms over the expression of individual taste. However, many of these studies have indeed found that factors common to the experience of all observers can be powerful predictors of preference. Other studies, however, have stressed this need to focus on individual's idiosyncratic preferences as the only way to elucidate the relationship between how a single observer evaluates an image and their preference for that image.

Is there a more general principle that can guide this discourse? When can we reasonably expect that people will agree in their preferences, and when must we sacrifice the statistical power that averaging provides and focus on individual preferences instead? In this paper, we report that indeed a general principle does exist which predicts whether people will agree in what they like and dislike, and when people will instead show divergent tastes.

We obtained preference judgments from a group of observers on both a set of real-world scenes (photographs of parks, buildings, street scenes, beaches, people, etc.) and a set of abstract images containing no identifiable (e.g. nameable) objects or structure, such as fractals, kaleidoscopic images, and photographs of natural patterns. Using these judgments, we calculated two measures. The first measure, within-observer reliability, indicated that observers were able to make reliable judgments for both sets of images - they were consistent with themselves. The second measure, between-observer agreement, gave us a measure of how much the different subjects agreed with each other. In line with previous findings, observers showed a high degree of agreement with the other observers for real-world scenes - if one observer liked a picture of a beach scene and disliked a picture of a trash dump, the other observers tended to agree. However, observers showed significantly less agreement for the abstract images - a fractal that was highly preferred by one observer was often strongly disliked by another observer. There were even pairs of subjects that showed correlated preferences for real-world scenes, but were then anti-correlated for abstract images.

These results provide strong evidence for the importance of semantics (meaning) in determining a subject's preference - if two subjects interpret an image in the same way, as is the case for photographs of everyday scenes and objects (two people are likely to describe a photograph in the same way), then they are likely to also share their like or dislike for that image. However, if two observers have vastly different interpretations for an image, as is often the case for an abstract image such as a fractal, since they do not contain easily identifiable, nameable features, then the preference of one observer will tell us nothing about the preference of the second observer.

This result, then, suggests a strong unifying principle that can guide future research into human preferences. If a researcher wishes to identify characteristics of images which lead, on average, to high preference across many people, then their stimulus set should first be tested for the degree to which people agree on the images' meaning. On the other hand, researchers looking to capitalize on the idiosyncrasies of individual observers would do well to use stimuli which contain a minimum of semantic associations.