MGMT 8825M Week 11 Ethics Argument Example

Not the Face, Not the Voice: An Ethics Argument That an Employer Should Not Let Facial and Vocal Analysis Score Interview Candidates Even if It Outpredicts Human Screeners, Answering the Strongest Case That Regulated Algorithms Make Discrimination Easier to Detect

Student Name

Walden University

MGMT 8825M: Ethics of Artificial Intelligence and Autonomous Systems

Instructor Name

Month Day, Year

Not the Face, Not the Voice: An Ethics Argument That an Employer Should Not Let Facial and Vocal Analysis Score Interview Candidates Even if It Outpredicts Human Screeners, Answering the Strongest Case That Regulated Algorithms Make Discrimination Easier to Detect

Abstract

This paper argues that an employer should not use facial and vocal analysis to score job candidates in video interviews, even where a vendor shows that the scores predict job performance better than human screeners do. The argument rests on an empirical premise, that facial movements do not reliably reveal inner states across people and contexts, and on a moral one, that candidates should be judged on what they can show about their ability to do the job, not on involuntary features of their face and voice. The strongest rival argument, that well-regulated algorithms make discrimination easier to detect than human judgment, is given full weight and answered. The paper concedes that the position gives up some predictive accuracy and some efficiency.

The Position

Halverson Contact Solutions, a composite firm operating call centers in five states, hires about 6,000 customer service agents a year. A vendor has offered it an interview platform that records candidates answering standard questions on video and scores them using, among other inputs, their facial movements and vocal features such as pitch, pace and variation. The vendor reports that, in a study at another call center firm, its scores predicted 90-day retention and customer satisfaction ratings better than the firm's human screeners did. The paper's position is that Halverson should decline the facial and vocal scoring, even if the vendor's claim is accurate.

The Practice as Vendors Present It

Vendors of these tools typically present them as a way to evaluate more candidates, more consistently, than human screeners can. The model is trained on recordings of past candidates whose later performance is known, and it learns which features of the recordings, including words, tone and facial movement, are associated with success. Vendors argue that this removes the inconsistency of human interviewers, who are known to be swayed by first impressions, similarity to themselves and fatigue.

The Empirical Premise

Barrett et al. (2019) reviewed the scientific evidence on whether emotional states can be inferred from facial movements. Their conclusion was that the way people communicate emotion varies substantially across cultures, situations and even across people within a single situation, and that facial configurations are neither reliable nor specific signs of particular emotions. A scowl may accompany anger, concentration or confusion; a smile may accompany happiness, nervousness or politeness. They cautioned that technologies claiming to read emotion from faces rest on assumptions the evidence does not support.

This premise matters for interview scoring in two ways. First, any link the model finds between facial movement and later job performance is unlikely to reflect a stable underlying trait such as enthusiasm or empathy. It is more likely to reflect correlations specific to the training data, which may not hold for a different employer, a different set of questions or a different population of candidates. Second, the features the model uses vary across groups for reasons unrelated to ability: cultural norms about eye contact and expression, disability, accent, neurodivergence and the lighting and camera quality available to a candidate at home.

How Vendors Describe Their Safeguards

Raghavan et al. (2020) examined the public claims of vendors of algorithmic pre-employment assessments. They found that vendors frequently claimed to mitigate bias but disclosed little about how their models were validated or how bias was measured, and that the most common approach centered on checking whether selection rates for protected groups met a statistical threshold. Satisfying that threshold does not show that a model measures something job-related. A model can pass a selection-rate test while scoring candidates on features that have nothing to do with the work, provided the scores happen to balance across the groups tested.

The vendor offering Halverson its platform follows this pattern. Its materials state that the tool "meets adverse impact guidelines" but do not say which features drive the scores or how much weight facial and vocal inputs carry.

The Strongest Rival Argument

Kleinberg et al. (2018) argued that algorithms, properly regulated, can make discrimination easier to detect and correct than human decision-making. A human screener's reasons are often unknowable, even to the screener. An algorithm, by contrast, has an explicit objective, explicit training data and inputs that can be examined. If an employer is required to keep records of these, a regulator or court can test whether the algorithm treats groups differently and why, which is much harder to do with a human interviewer. On this view, refusing algorithmic scoring may leave candidates with the less accountable alternative.

The argument is strong, and it applies to many uses of algorithms in hiring. A structured scoring rubric applied by software to candidates' written answers, for example, would gain from exactly the transparency it describes.

Why the Rival Does Not Rescue Facial and Vocal Scoring

The rival argument depends on the inputs being open to scrutiny and justification. Facial and vocal analysis fails that condition in a specific way. Even if every input is recorded, there is no account of why the movement of a candidate's eyebrows or the variation in her pitch should bear on whether she can resolve a customer's billing problem. The algorithm may be transparent about what it measured, but what it measured cannot be justified as job-related, because the empirical premise shows that these features do not reliably indicate the traits they are taken to signal. Transparency makes it possible to detect that an input is being used. It cannot make an unjustifiable input acceptable.

There is also a moral point independent of accuracy. A candidate can prepare her answers, practice explaining her experience and show what she knows. She cannot control the involuntary features of her face and voice. Scoring her on them treats her as a set of signals to be read rather than as a person making a case for herself.

What the Position Gives Up

The position has costs. If the vendor's claim is accurate, declining facial and vocal scoring means hiring some agents who will leave sooner or satisfy customers less, and screening candidates more slowly. Halverson's hiring managers estimate that human screening of 6,000 hires a year, from about 40,000 applicants, costs about $1.2 million in recruiter time. The position also forgoes the consistency an algorithm offers. These costs can be reduced without facial and vocal scoring: by using structured interviews scored against a rubric, which research has long found more predictive than unstructured ones, and by letting software score the content of candidates' answers while ignoring their faces and voices.

Conclusion

What the literature supports should be kept apart from what this paper proposes. The literature supports the claim that facial movements are not reliable signs of inner states, and it documents that vendors disclose little about how their bias safeguards work. It also supports the view that regulated algorithms can be more accountable than human judgment. The paper's own contribution is the argument that transparency cannot justify an input that has no defensible connection to the job, and the moral claim that candidates should not be scored on what they cannot control. Halverson should decline the facial and vocal scoring, require any algorithmic tool it adopts to disclose the features that drive its scores and limit scoring to the content of what candidates say.

References

Barrett, L. F., Adolphs, R., Marsella, S., Martinez, A. M., & Pollak, S. D. (2019). Emotional expressions reconsidered: Challenges to inferring emotion from human facial movements. Psychological Science in the Public Interest, 20(1), 1-68. https://doi.org/10.1177/1529100619832930

Kleinberg, J., Ludwig, J., Mullainathan, S., & Sunstein, C. R. (2018). Discrimination in the age of algorithms. Journal of Legal Analysis, 10, 113-174. https://doi.org/10.1093/jla/laz001

Raghavan, M., Barocas, S., Kleinberg, J., & Levy, K. (2020). Mitigating bias in algorithmic hiring: Evaluating claims and practices. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (pp. 469-481). Association for Computing Machinery. https://doi.org/10.1145/3351095.3372828