'AI Personality Extraction from Faces: Labor Market Implications' by Marius Guenzel, Shimon Kogan, Marina Niessner and Kelly Shue comments
Human capital — encompassing cognitive skills and personality traits — is critical for labor market success, yet the personality component remains difficult to measure at scale. Leveraging advances in artificial intelligence and comprehensive LinkedIn data, we extract the Big 5 personality traits from facial images of 96,000 MBA graduates, and demonstrate that this novel "Photo Big 5" predicts school rank, compensation, job seniority, industry choice, job transitions, and career advancement. Using administrative records from top-tier MBA programs, we find that the Photo Big 5 exhibits only modest correlations with cognitive measures like GPA and standardized test scores, yet offers comparable incremental predictive power for labor outcomes. Unlike traditional survey-based personality measures, the Photo Big 5 is readily accessible and potentially less susceptible to manipulation, making it suitable for wide adoption in academic research and hiring processes. However, its use in labor market screening raises ethical concerns regarding statistical discrimination and individual autonomy. ...
The authors go on
In this paper, we depart from using survey-based personality measures, and instead leverage recent advances in artificial intelligence (AI) that enable us to extract personality traits from a single facial image of a person. These advancements, which facilitate the construction of large-scale personality datasets, reflect a broader trend in which AI facial recognition is increasingly adopted across various settings, including matching in dating markets, political affiliation analysis, and targeted marketing.
Using new alternative data—photos from LinkedIn and photo directories of several top U.S. MBA programs—we extract the Big 5 personality traits for 96,000 MBA graduates, for whom we also observe detailed employment outcomes and education histories. We then assess the ability of the novel “Photo Big 5” to predict labor market outcomes such as school rank, compensation, and advancement within organizational hierarchies. We find that, while the vast majority of variation in labor outcomes remains unexplained, the Photo Big 5 provides predictive power comparable to a person’s race, attractiveness, and educational background. Moreover, because the Photo Big 5 exhibits weak correlations with traditional cognitive measures—such as grades and test scores—typically used in labor market screening, it delivers high incremental predictive power. For example, the compensation disparity between individuals in the top quintile versus the bottom quintile of ‘desirable’ Photo Big 5 personality traits is larger than the compensation gap observed between Black and White graduates for men, and about 65% of the Black-White compensation gap for women.
We focus on the Big 5 personality traits because they are the most widely used and extensively studied measures of ‘soft skills’ in finance and economics (e.g., Heckman and Kautz (2012)). The five traits are: Openness (curiosity, aesthetic sensitivity, imagination), Conscientiousness (organization, productiveness, responsibility), Extraversion (sociability, assertiveness, energy level), Agreeableness (compassion, respectfulness, trust), and Neuroti- cism (anxiety, depression, emotional volatility). We study the labor market for MBA grad- uates, as survey and task-based measures of personality are already heavily used as part of hiring and job screening in the MBA labor market. The focus on MBAs also allows us to examine a high-skill population for which we can compare the predictive power of the Photo Big 5 against cognitive measures such as school rank, GPA, and standardized test scores.
The face-based personality extraction draws upon a robust body of scientific research in genetics, psychology, and behavioral science that has empirically established three primary, non-exclusive channels linking facial features and personality. First, an individual’s genetic profile significantly influences both their facial features and personality. Certain variations in DNA correlate with specific facial features, such as nose shape, jawline, and overall facial symmetry, defined broadly as craniofacial characteristics (Claes et al., 2014). Related evidence indicates that 30%-60% of the variance in Big 5 personality traits across individuals is attributable to genetic factors (Vukasovi´c and Bratko, 2015). Further, a growing body of literature has used large-scale genome-wide association studies (GWAS) to investigate the genetic underpinnings of personality traits (e.g., De Moor et al. (2012), Lo et al. (2017), Nagel et al. (2018)), finding that individual genetic variants collectively contribute to the heritability of personality traits and identifying specific genes linked to cognitive performance and personality traits.
Second, a person’s pre- and post-natal environment, especially hormone exposure, has been shown to affect both facial characteristics and personality. Verdonck et al. (1999) and Whitehouse et al. (2015) study the link between post- and pre-natal testosterone exposure and facial structure. Cohen-Bendahan et al. (2005) focus on prenatal hormone exposure and personality traits such as aggression, empathy, and social interest. Szyf et al. (2007) investigate the postnatal effects of the environment on gene expression (i.e., epigenetics) and behavior.
Finally, perceptions of one’s facial features, whether by oneself or others, can influence and be influenced by personality traits (e.g., the “Quasimodo Complex” as described in Masters and Greaves (1967)). For example, Umberson and Hughes (1987) show that others’ assessments of attractiveness correlate with achievement and psychological well-being. Other studies show that others’ perceptions of personality traits influence behavior such as friendliness and sociability (Snyder et al., 1977). Moreover, Zebrowitz and Montepare (2008) show that “babyfaced” individuals are stereotyped as more naive, warm, and submissive, often leading them to adopt more agreeable behaviors. In this project, we focus on evaluating the predictive potential of the facial-image-based Big 5 assessment, leaving the inquiry into the precise mechanisms underpinning the link between facial features and personality traits to other researchers.
'Stressing the ‘body electric’: History and psychology of the techno-ecologies of work stress' by Jessica Pykett and Mark Paterson in (2022) 35(5) History of the Human Sciences comments
This article explores histories of the science of stress and its measurement from the mid 19th century, and brings these into dialogue with critical sociological analysis of emerging responses to work stress in policy and practice. In particular, it shows how the contemporary development of biomedical and consumer devices for stress self-monitoring is based on selectively rediscovering the biological determinants and biomarkers of stress, human functioning in terms of evolutionary ecology, and the physical health impacts of stress. It considers how the placement of the individual body and its environment within particular spatio-temporal configurations renders it subject to experimental investigation through standardized apparatus, electricity, and statistical normalization. Examining key themes and processes such as homeostasis, metricization, datafication, and emotional governance, we conclude that the figure of the ‘body electric’ plays a central limiting role in current technology-supported approaches to managing work stress, and that an historical account can usefully open these to collective scrutiny. ...
Stress is increasingly a global problem that affects both mental and physical health outcomes. Globally, work stress is thought to be a key contributor to the rise in mental health problems such as anxiety and depression. The World Health Organization recognizes stress as a growing problem in developing countries, and the next International Classification of Diseases (ICD-11), to be published in January 2022, defines the phenomenon of chronic workplace stress as ‘burnout’ (World Health Organization, 2019). Nationally, governments are increasingly concerned about the links between work stress, absenteeism, and presenteeism, including the impact on economic productivity. In the United States alone, stress costs enterprises $300 billion, or 2.6% of GDP in 2006 (Brun, 2007). Strategic responses across government and the private sector variously involve risk assessment and organizational procedures (Health and Safety Executive, 2019), and guidance on the overuse of digital technologies. Many workplace well-being initiatives offer advice on work–life balance, and a conviction to reconnect people (with their bodies, minds, nature, or each other) through such individual activities as mindfulness, yoga sessions, and opportunities for exercise. As we will argue, such initiatives portray a set of assumptions in the public sphere around selectively rediscovering the biological determinants and biomarkers of stress; human functioning in terms of evolutionary ecology; and the physical health impacts of stress. Rather than offer historical accounts and reports of stress or the phenomenon of burnout as such (see Hoffarth, 2017; Jackson, 2013), then, our emphasis is on how individualized, physiological, and evolutionary accounts of stress have recently become mediated through technological formations that represent embodied stress as an aggregation of externalized data points. We will explore how these manoeuvres can serve to hollow out collective claims to being well in the contemporary workplace.
A key feature of emerging responses to work stress is the development of biomedical and consumer devices for self-monitoring. These involve combinations of novel software and wearable sensor hardware to measure electrodermal activity, heart rate variability, or to assess electrochemical or volatile organic biomarkers for emotional stress found in sweat (e.g. Zamkah et al., 2020). An article in Nature heralds the development of ‘epidermal electronics’ as the inevitable next step: ‘wireless sensors mounted directly on the skin, where they can pick up a host of vital signs, including temperature, pulse and breathing rate’ (Gibney, 2015: 27). Therapies including eMental health apps and transcranial magnetic stimulation have variously been posited as the future of medical treatments to address national ‘epidemics’ of stress (Kim et al., 2016; Phillips, Gordeev, and Schreyögg, 2019). Less well known is the development of ‘electroceuticals’, treatments based on electromagnetic fields (Famm et al., 2013). The American Institute of Stress (n.d.) believes that electroceuticals signify the future of medicine. As we discuss below, this apparently futuristic approach is entirely in keeping with the historical fascination with the medical applications of electricity (e.g. Morus, 1999; Parisi, 2018; Peña, 2005) and the use of electricity to induce and also to gauge stress within physiological experimentation.
This article considers the development of an historical consensus on the science of stress and its measurement from the mid 19th century onwards. It is a consensus that placed the individual body and its environment within a particular spatio-temporal configuration, making it subject to experimental investigation through standardized apparatus, electricity, and statistical normalization. We argue that these historically emergent spatio-temporalities of stress within and across bodies have shaped contemporary approaches to stress. By using Foucault’s (2003) descriptions of the social apparatus of anatomo-politics and biopolitics, we show how the history of stress research is also one of emotional management, normative ideals, and subjectivity formation. We demonstrate how an imaginary of the ‘body electric’ persists as a central figure in the knowledge, epistemologies, and methodologies of stress research and workplace well-being initiatives. The historical trajectory encompasses techniques and technologies that emerged from early experiments with electrical stimulation of nerves in the late 18th century, and the psychophysics of emotions and sensation in the late 19th century. This is followed by the identification of physiological mechanisms of homeostasis and adaptation to external stimuli in the mid 20th century, right up to the contemporary use of wearable digital biosensors. These diverse strands converge by means of distinct yet related processes of metricization and then datafication. Hence a key tension in the stress management practices of the contemporary workplace is a result of novel spatio-temporalities of data standardization, ownership, and analysis. The aggregation of biophysical data is no longer contained solely within the physical hardware of the wearer's device, but shared globally through proprietary cloud-based software platforms, with potentially negative implications for the possibility of collective organization at work. We explore how such wearable technologies have suddenly become adopted into the therapeutic regimens of workplace stress. By doing so, we foreground the conceptual development of what has become a subfield of biology, the physiology and neurophysiology of stress. We explore how this subfield can be complemented by more interdisciplinary insights to suggest how researchers, organizations, and governments might deal with this contemporary occupational health crisis.
Since working conditions (relations of power, labour, and capital) are central to understanding work stress, it is somewhat surprising that some research developments in the treatment of work stress are often concerned with individual biological aetiology (Polacchini et al., 2018; Ryff, Singer, and Dienberg Love, 2004; Sumner et al., 2020). The core aim of this article is to explore how this rift between individual aetiology and work conditions has come to pass, and to consider how the emphasis on self-management might be impacting workers’ collective capacity to monitor and address their workplace stress. We outline the historical specificity of the way in which wearable digital technologies – which sense, measure, compute, and visualize physiological phenomena – transform the body into data that is then algorithmically correlated with particular psychological states. This mechanism encourages users to ‘capture’ and categorize certain psychological traits or habits of thought and ‘improve’ their responses to specific triggers or environments through biofeedback, that is, biological data presented to them. This is often via obscure techniques of analysis and proprietary software channels – in order to elicit emotional self-management. We argue these technologies reconfigure mind–body–environment relations leading to the emergence of a specific spatio-temporality of stress. Temporally, stress is conceived as an immediate psychophysiological ‘flight or fight’ response. Spatially, this response relates to proximate stimuli within an individual's perceptual environment. We find that understanding of contemporary articulations of ‘techno-ecologies’ of stress can be illuminated through the historical excavation of the ‘body electric’, and can be understood through a framework that combines historical and sociological analysis. It is by no means inevitable that the workplace technologies outlined here should prioritize emotional self-management at the apparent expense of workers’ capacities for collective organizing. Yet, as we argue, the ways that data from such technologies is processed, visualized, shared, and stored provides grounds for cautioning against the particular form of emotional governance enabled by these processes.
The article consists of three substantial sections, each of which maps onto models or frameworks for defining and interpreting stress, namely the organismic, the psychological, and the ecological. We begin with ‘Defining stress’, a review of the origins of scientific definitions of stress within early 20th-century physiology and psychology, which takes the organism and its interior and exterior milieu as the main unit of analysis, prior to the employment of stress in relation to the workplace and technology. The following section, ‘Metricization and bodily sensation’, then outlines the processes of metricization achieved through scientific instrumentation to constitute the figure of the body electric, the productive worker's body as monitored and measured. This places the measurement and modelling of stress and its effects more firmly into the realm of psychology, and especially the emerging field of industrial psychology. In the third section, ‘Datafication and the biopolitics of machinic emotions’, we consider the legacy of this process by showing how specific workplace ecologies of digital wearable technologies render the ‘body electric’ visible and sharable through data, framing it as the biopolitical target-object of emotional governance. Emotional governance can refer to social practices of regulation, management, government and policymaking, service design, or state-citizen relationships, which see emotions as central to their conception and operation (Jupp, Pykett, and Smith, 2017). We argue that platforms for obtaining and sharing data captured by wearable biosensors constitute an ecology of data and devices, and a corresponding ecological approach to stress and stressors. We thereby show how deepening the understanding of the convergent role of scientific instrumentation and digital technologies in the historical and contemporary conceptualization of work stress advances this agenda by considering the processes by which the body electric shapes the discourses and regulatory practices of emotional governance. Finally, in the conclusion we draw out the significance of this account for evaluating digital workplace well-being interventions.
