When beauty filters reality

Cognitive biases and heuristics profoundly impact our perception, memory and decision making. While they have been and are commercially leveraged to manipulate human behavior in a variety of scenarios—including casinos, addictive apps, advertisement, marketing strategies and social media campaigns—there is an unexplored opportunity to create a more constructive interplay between cognitive biases and AI systems, shifting from manipulation to collaboration. A first step in this direction consists of studying the existence of cognitive biases in human-AI interaction, including scenarios where humans are augmented by means of AI algorithms, such as beauty filters. Beauty filters are a popular family of face filters that leverage computer vision and augmented reality methods to automatically beautify selfies uploaded by social media users. With millions of users worldwide, these filters profoundly impact user self-presentation, raising questions about authenticity, self-esteem, mental health, diversity and racism.

In this work, we study the impact of beauty filters on the attractiveness halo effect, a cognitive bias whereby individuals tend to associate positive attributes, such as intelligence and honesty, to those who are perceived as more attractive. In our research, we selected a gender-race-age balanced sample of 462 images from two popular face datasets—the Chicago Faces Database and the FACES database. Both of these image sets consist of individuals with a neutral expression facing the camera and wearing the same clothes in front of a uniform background i.e., minimizing confounds. The image sets are diverse and balanced across age, ethnicity and gender. We refer to these images as the PRI set. We then applied a popular beauty filter used extensively on social media to each of these images to create a new set of beautified images (POST set) depicting the same individuals as in the PRI set but in an “attractive” (beautified) condition.

Samples of female and male face images used in the study, before (left) and after (right) the application of the beauty filter.

We recruited 2,748 participants from Prolific to rate the images on a 7-point Likert scale on perceived attractiveness, and 6 other attributes, including intelligence, trustworthiness, sociability and happiness. Each participant evaluated a gender-balanced sample of 10 distinct individuals with an equal number of samples from the PRI and POST sets, but were not told about the beauty filters being applied to some of the images they saw. Each face image was rated by at least 25 participants. This large dataset enables us to study the attractiveness halo effect on the same individual in two conditions (original and beautified)—minimizing the possibility of factors other than attractiveness impacting the effect; at scale and with diversity in the age, gender and ethnicity of the stimuli, which significantly expands the studies carried out to date. Furthermore, we shed light on the impact that pervasive beauty filters have on this cognitive bias.

Our analysis was split in two levels. First, using the medians of all the ratings each image received which allowed for pairwise comparison between images in the PRI and POST set, but masked the impact of raters age and gender. The second level used all the scores, but after they were passed through an Ordered Stereotype Model to account for the ordinal nature of the data. While details about the methodology used can be found in our manuscript, here we summarize some of the key findings of the study:

• Beauty filters increase perceptions of attractiveness and other behavioral traits: Our results show that almost everyone was perceived as more attractive after the filters were applied and no one was perceived as less attractive. Additionally, there were statistically significant increases in perceptions of intelligence, trustworthiness, happiness and sociability simply by applying a beauty filter.

• Gender, age and initial attractiveness matter; ethnicity does not: We found that females were perceived as significantly more attractive than males and younger people were perceived as significantly more attractive than older people. Additionally, the attractiveness scores for females increased significantly more than for males after applying filters and middle-aged individuals benefited more from filters than their younger counterparts. Individuals who were rated with lower attractiveness levels in the PRI set benefited significantly more by applying filters. Interestingly, ethnicity had no impact on perceptions of attractiveness or the impact of the filters.

• Beauty filters partially mitigate the attractiveness halo effect: Attributes like intelligence and trustworthiness show a saturation in their relationship with attractiveness. Thus, applying beauty filters reduces the impact of attractiveness and significantly weakens the halo effect for these attributes. Such an effect however is not seen for sociability and happiness. Our study is one of the first to report such a saturation effect and look at filters as a tool to mitigate the halo effect.

• Beauty filters exacerbate existing gender stereotypes: While females were perceived as more attractive than males in both the PRI and POST set, males were perceived as more intelligent indicating a stronger gender stereotype in perceptions of intelligence. Concerningly, not only were men perceived as more intelligent, but the gap in intelligence scores provided to men and women was higher in the POST set indicating that beauty filters not only propagate gender stereotypes but they exacerbate them as well.

While our findings not only provide conclusive results but also new insights into the halo effect and how it presents in the modern digital age, they also raise several ethical concerns regarding the use of beauty filters. Our study reveals that females are often perceived as more attractive but less intelligent than males, reflecting entrenched gender stereotypes and the belief that physical appearance inversely correlates with intelligence. It also finds that male raters are more influenced by beauty filters than females, which raises concerns about how these tools may reinforce societal biases, perpetuate gender discrimination, and uphold traditional gender roles. While beauty filters have the potential to reduce the intensity of the attractiveness halo effect, their use brings ethical questions about authenticity, honesty, and the risk of influencing decision-making without users’ knowledge.

These findings highlight the complex relationship between technology, social perceptions, and ethics, underscoring the importance of transparency and ethical guidelines for the use of beauty filters.

Read the full paper here: What is beautiful is still good: the attractiveness halo effect in the era of beauty filters 

Royal Society Open Science is an open access journal that welcomes the submission of all high-quality science. More information about the submission process can be found on our webpage.

Images: From figure 6 of the paper. Samples of male and female face images used in the study before (left) and after (right) the application of the beauty filter. As illustrated in the examples, the beauty filter modifies the skin tone, the eyes and eyelashes, the nose, the chin, the cheekbones and the lips in order to make the person appear more attractive.