Perhaps there is nothing more closely tied to one’s identity than their own face. Much of the human experience surrounding faces relates to how faces are recognized or processed. Facial recognition has been found to be processed as a whole, rather than by parts [1]. Additionally, facial recognition is sensitive to the context in which it is being processed [1]. However, beauty itself is found to be relatively resistant to external factors and context [1]. It is found, though, that judgments of morality based on faces is influenced by external factors, which is one reason why facial recognition can have large contextual influences [2]. Facial attractiveness can be measured and tracked in many ways; one approach is eye tracking. It has been found that greater eye movement is correlated with viewing a more attractive face [3]. Developmentally, eye tracking is found to increase throughout infancy and directly correlate with increases in facial recognition and bias [4]. Thus, psychological research on faces is important, particularly in an era when realistic faces are being produced by generative Artificial Intelligence (AI).

Recently, AI image generation by text-to-image models, has impressed the public with its sophistication and realism. Whereas traditional AI algorithms have relied on structured data for model building and information processing, machine learning techniques have fundamentally altered AI research in recent years [5]. Machine learning is crucial as it allows systems to acquire new information and expand their knowledge, judgments, and conclusions through experience and data without being explicitly programmed to do so beforehand [6]. More advanced generative AI algorithms have evolved to process data in natural language, enabling the generation of images from text. These algorithms have evolved to where convoluted neural networks and recurrent neural networks have gained the ability to analyze and produce images, audio, and video [5].

Natural Language Processing (NLP) has been important in the development of chatbots as the use of NLP techniques allows these programs to understand and interpret human language input [6]. OpenAI © is an artificial intelligence company that has released several generative AI applications including DALL·E and ChatGPT. DALL·E, along with many other chatbots, uses NLP input to create imagery in multiple styles, including photorealistic imagery, paintings, and emoji [6]. Any user can input text, such as a descriptive caption, into DALL·E, prompting it to generate images based on the input received. Given the power of these tools, research on the ability of generative AI, particularly DALL·E, to process text descriptions of faces has theoretical and practical implications.

Facial Appearance as Core Expression Scale (FACES) was developed to assess how individuals perceive themselves and how well it represents their ideal self. This instrument can help to identify and reduce disparities in expectations of facial surgical patients who may not have the same perceptions as their surgeons. This instrument has many uses including the ability to quantify facial satisfaction the patient experiences, ability to compare surgical techniques, and the ability to identify patients who may be unlikely to perceive benefits from surgery.

Patients answer seven questions about how much they agree or disagree with how much those statements apply to their face today by moving a slider from 0 to 100. These seven statements include: my face is pretty or handsome, my face is like I want it to be, my face is like I want others to see me, my face is appealing, my face is attractive, my face is likable, my face is like my ideal self. This FACES score is an average of these seven responses, with higher numbers corresponding to a more positive view of one’s own face. This set of seven items was highly reliable in previous research with Cronbach’s α = 0.94 [7]. Research by Wolfe et al. indicates that in a non-surgical population most individuals rated themselves between 40 and 80, and less than 3% would rate themselves lower than 20 [7]. In another study, participants rated their own face as well as several other facial images using the FACES instrument. Eight of the pictures shown were of 4 patients who had undergone maxillofacial surgery with before and after surgery pictures. The results from this study suggested that FACES items are sensitive to the results of surgical interventions and validated this instrument in detecting predicted differences with high sensitivity of about 0.77 standard deviations [7]. A third study investigated individual differences with participants rating their own faces and completing the Body Image Avoidance Questionnaire, Rosenberg Self-Esteem, and State Self Esteem Scale, and each predicted FACES outcomes with R2 = 0.43 [8]. This indicates a strong relationship between self-esteem and people’s perceptions of their own faces as measured with the FACES instrument.

The purpose of this study was to investigate the potential for DALL·E 2 to create useful stimuli for medical and psychological research on faces using FACES. If the text interface of DALL·E is sensitive to the characteristics of human faces captured with the FACES instrument, then textual descriptions of faces based on those characteristics should produce visual stimuli (pictures of faces) that participants rate differently depending on whether the generated faces were generated with positively-worded instructions (e.g. “the face is like I want others to see me”) on negatively-worded instructions (e.g. “the face is not like I want others to see me”). The hypothesis was that faces generated by DALL·E using positively worded instructions will score significantly higher on FACES than those that are worded negatively by the addition of “not” before the same positively worded statements. This hypothesis was tested across a range of instructions to DALL·E for apparent age, ethnicity, and gender.