AI INTERVIEW SCORING: DISCRIMINATION RISKS

Introduction

Artificial Intelligence (AI) is increasingly used in recruitment to conduct video interviews, analyse speech and facial expressions, evaluate answers, assign numerical scores, and rank candidates. AI interview-scoring systems may assess factors such as vocabulary, speech patterns, response time, facial expressions, confidence indicators, or predicted job performance.

Although these systems are designed to make recruitment faster and more consistent, they can create significant discrimination risks. An apparently neutral algorithm may reproduce historical discrimination, rely on inappropriate proxy variables, or disadvantage applicants whose communication style, disability, age, sex, race, or other protected characteristic differs from the data used to train the system.

The central legal issue is therefore not merely whether the employer intended to discriminate, but whether the design, validation and use of the AI scoring system produces unlawful discriminatory effects.

1. Meaning of AI Interview Scoring

AI interview scoring refers to the use of algorithmic or machine-learning systems to evaluate candidates during an interview and generate a score, ranking, recommendation, or hiring decision.

For example, an AI system may:

record a video interview;

convert speech into text;

analyse the candidate's answers;

evaluate speech characteristics;

analyse facial or behavioural signals;

compare the candidate with historical hiring data;

assign a numerical score; and

recommend whether the candidate should proceed.

The legal risk increases when the AI score is treated as an objective measure of merit without adequate testing or human review.

2. Historical Bias in Training Data

One of the principal risks is historical bias.

If an AI system is trained using previous hiring decisions, the historical data may reflect discriminatory patterns. The algorithm can learn that candidates possessing characteristics associated with previously successful employees received higher scores.

Consequently, the system may reproduce previous inequalities even though the protected characteristic is not expressly entered into the algorithm.

For example, if previous successful employees predominantly belonged to one demographic group, an algorithm trained on those outcomes may learn patterns associated with that group and give other candidates lower scores.

The current Mobley v. Workday, Inc. litigation illustrates this legal concern. The plaintiffs allege that algorithmic screening and scoring systems can reproduce historical disparities through training data, model design, input variables and evaluation criteria. The litigation involves allegations concerning race, sex, age and disability discrimination. These remain allegations rather than final findings of liability.

3. Proxy Discrimination

AI does not necessarily need to receive a person's race, sex, age, or disability information directly in order to produce discriminatory results.

Other variables may operate as proxies for protected characteristics.

Examples include:

geographical location;

educational institution;

employment history;

career gaps;

language patterns;

speech characteristics;

names;

availability patterns;

social or professional networks.

Therefore, removing protected characteristics from the algorithm does not automatically eliminate discrimination.

The Mobley allegations specifically raise the concern that protected characteristics can be inferred through correlated variables even where the protected characteristic itself is not explicitly supplied to the system.

4. Disability Discrimination

AI interview technology creates particular risks for applicants with disabilities.

Speech-recognition systems may disadvantage:

applicants with speech impairments;

applicants who stutter;

applicants with neurological conditions;

applicants using assistive technology.

Similarly, facial-analysis systems may inaccurately evaluate applicants whose facial expressions or movements differ from patterns expected by the software.

The U.S. Equal Employment Opportunity Commission (EEOC) has specifically warned that algorithmic decision-making tools may unintentionally screen out qualified individuals with disabilities. It explains that employers may need to provide reasonable accommodation, including an alternative testing format, where AI technology inaccurately measures an applicant's ability because of disability.

Thus, an AI score cannot automatically be treated as an accurate measure of ability where the scoring methodology is affected by a disability.

5. Disparate Impact Discrimination

AI interview scoring may produce disparate impact even without intentional discrimination.

Disparate impact occurs where an apparently neutral employment practice disproportionately disadvantages a protected group.

For AI recruitment, the relevant questions include:

Does one racial group receive systematically lower scores?

Are women disproportionately rejected?

Are older applicants disadvantaged?

Are applicants with disabilities receiving lower scores?

Is the scoring criterion actually related to job performance?

Is there a less discriminatory alternative?

The important point is that an employer may face legal scrutiny even where there is no evidence that the algorithm was deliberately programmed to discriminate.

The Mobley litigation is particularly significant because plaintiffs have asserted disparate-impact theories involving automated recruitment and screening tools under Title VII, the ADA and the ADEA.

6. Facial Expression and Emotion Analysis

Some AI interview systems attempt to evaluate:

facial expressions;

eye contact;

smiling;

body movements;

voice tone;

perceived confidence;

emotional responses.

These techniques raise reliability and discrimination concerns.

A candidate may naturally have a different communication style because of culture, disability, neurodiversity, anxiety, language background, or other factors unrelated to job performance.

If an employer treats a low "confidence" or "engagement" score as evidence of poor suitability, the employer may unintentionally disadvantage particular groups.

The legal question therefore becomes whether the AI measurement is job-related, reliable, consistently applied, and appropriately validated.

7. Race and Ethnic Discrimination

Speech and language analysis can create racial and ethnic discrimination risks.

An AI model may have been trained predominantly on particular accents or speech patterns. Consequently, candidates with different accents may receive inaccurate transcriptions or lower scores.

For example, an algorithm might incorrectly interpret pronunciation differences as poor communication ability.

The problem becomes especially serious where communication is not an essential requirement of the particular job.

An employer should therefore establish that the scoring criterion measures an actual occupational requirement rather than merely similarity to historical successful candidates.

8. Sex and Gender Discrimination

AI interview systems can also reproduce gender-related bias.

Historical hiring data may reflect male-dominated occupations or historical preferences for particular communication styles.

If an AI system learns from such data, it may reproduce those patterns.

The Mobley litigation includes allegations concerning gender discrimination in automated recruitment and scoring systems. The allegations have not yet been finally adjudicated.

9. Age Discrimination

AI scoring may disadvantage older applicants where the system associates certain characteristics with younger workers.

Potential risk factors include:

career length;

employment gaps;

older educational qualifications;

technology-related experience;

speech characteristics;

assumptions concerning adaptability.

Under U.S. federal law, the Age Discrimination in Employment Act (ADEA) protects workers aged 40 and above from specified forms of age discrimination.

The Mobley litigation also includes allegations that automated recruitment and scoring tools disproportionately disadvantage applicants over 40. Again, these are allegations in ongoing litigation rather than established findings of liability.

10. Lack of Transparency

Another major problem is algorithmic opacity.

A candidate may receive a score such as:

"Interview Score: 62/100"

without knowing:

which factors produced the score;

whether the score was generated by AI;

what data was used;

whether protected characteristics influenced the result;

whether the system has been tested for bias;

whether a human reviewed the decision.

Lack of transparency makes it difficult for applicants to identify discrimination and challenge an adverse decision.

Therefore, employers should maintain documentation concerning the system's design, validation, testing and use.

11. Human Oversight

AI should not necessarily be treated as an independent decision-maker.

Human oversight is particularly important where an AI system:

rejects candidates automatically;

produces a final hiring score;

detects personality characteristics;

evaluates disability-related characteristics;

makes recommendations concerning protected groups.

A human reviewer should be capable of identifying obvious algorithmic errors and considering relevant accommodation requests.

However, merely placing a human at the end of the process does not automatically eliminate liability if the human simply accepts the AI recommendation without meaningful review.

12. Important Case Laws

1. Mobley v. Workday, Inc. (N.D. Cal., 2026)

This is one of the most significant contemporary cases involving algorithmic employment discrimination.

The plaintiffs allege that Workday's automated recruitment, screening, scoring and ranking tools disproportionately disadvantage applicants based on race, sex, age and disability.

The case demonstrates how traditional employment-discrimination doctrines are being applied to modern algorithmic hiring systems. The court allowed significant portions of the litigation to proceed, while the ultimate merits remain unresolved.

Legal significance: AI vendors and employers may face discrimination claims where automated employment tools materially influence hiring decisions.

2. Griggs v. Duke Power Co., 401 U.S. 424 (1971)

The U.S. Supreme Court established the principle that an apparently neutral employment requirement can violate Title VII when it disproportionately excludes a protected group and is not sufficiently related to job performance and business necessity.

Relevance to AI: An AI interview score can similarly be examined as a neutral employment practice producing discriminatory effects.

3. Washington v. Davis, 426 U.S. 229 (1976)

The U.S. Supreme Court distinguished discriminatory impact from discriminatory intent in constitutional equal-protection litigation.

Relevance: The case helps explain why statistical disadvantage and intentional discrimination are legally distinct concepts.

4. Watson v. Fort Worth Bank & Trust, 487 U.S. 977 (1988)

The Supreme Court recognized that subjective employment-selection practices may be examined under disparate-impact principles.

Relevance to AI: The fact that a hiring process is based on apparently objective scoring does not necessarily make it immune from disparate-impact scrutiny.

5. Ricci v. DeStefano, 557 U.S. 557 (2009)

The Supreme Court considered the tension between avoiding disparate impact and avoiding intentional disparate treatment.

Relevance: Employers implementing AI fairness measures must carefully structure their practices because employment-discrimination law can involve competing legal considerations.

6. Nelson v. USAble Mutual Insurance Co. (8th Cir. 2019)

The Eighth Circuit considered a discrimination claim involving a hiring/promotion process where interview scoring was relevant to the employer's decision. The court upheld summary judgment for the employer because the plaintiff failed to demonstrate that the stated legitimate reason was pretextual.

Relevance to AI: Interview scores can become important evidence concerning the legitimate, non-discriminatory basis for a hiring decision. Employers should therefore be able to explain how scores were generated and used.

7. EEOC Guidance on Algorithmic Decision-Making and Disability

Although not a judicial case, EEOC guidance is highly relevant. The EEOC states that algorithmic tools may screen out individuals with disabilities and emphasizes reasonable accommodation where the technology inaccurately evaluates an applicant because of disability.

Legal significance: Employers should assess accessibility and accommodation requirements before relying upon AI assessments.

13. Employer Compliance Duties

Employers using AI interview scoring should consider the following safeguards:

A. Validate the System

The employer should determine whether the AI actually measures characteristics relevant to successful job performance.

B. Conduct Bias Audits

Regular statistical testing should identify disproportionate outcomes affecting protected groups.

C. Test for Disability Effects

The employer should determine whether speech, facial or behavioural analysis disadvantages disabled applicants.

D. Provide Reasonable Accommodation

Applicants should have an accessible alternative where AI technology does not accurately evaluate their abilities because of disability.

E. Maintain Human Review

Important employment decisions should not necessarily be based exclusively on automated scores.

F. Document the Decision-Making Process

Employers should maintain records regarding:

training data;

scoring criteria;

validation studies;

bias testing;

accommodation procedures;

human review;

final hiring decisions.

G. Monitor Vendors

An employer should not assume that responsibility disappears merely because the AI tool is supplied by a third-party technology company.

14. Legal Issues for AI Vendors

AI vendors may also become involved in litigation where their technology materially influences employment decisions.

Important questions include:

Who designed the algorithm?

Who selected the training data?

Who controls the scoring criteria?

Who determines how the score is used?

Does the vendor provide bias-testing information?

Does the vendor provide accommodation mechanisms?

Does the employer independently validate the technology?

The Mobley litigation is particularly important because it illustrates the emerging debate concerning the legal relationship between employers and third-party algorithmic hiring providers.

15. Remedies and Corrective Measures

Where unlawful discrimination is established, potential remedies may include:

reconsideration of an employment application;

compensation for economic loss;

damages where legally available;

injunctive relief;

modification of discriminatory hiring practices;

implementation of bias-audit procedures;

reasonable accommodation;

reconsideration of automatically rejected candidates.

The exact remedies depend on the applicable jurisdiction and statute.

Conclusion

AI interview scoring can improve recruitment efficiency, but it does not automatically make employment decisions objective or legally neutral. Discrimination may arise through biased training data, proxy variables, inaccurate speech or facial analysis, disability-related limitations, historical hiring patterns, inappropriate scoring criteria, and lack of meaningful human oversight.

The emerging legal approach is to apply established employment-discrimination principles to the technological process rather than treating AI as legally exempt because the decision is made by software. The Mobley v. Workday litigation is particularly significant in demonstrating how automated recruitment and scoring systems are becoming the subject of discrimination litigation.

Therefore, employers should ensure that AI interview systems are job-related, validated, accessible, regularly audited, transparent enough to permit meaningful review, and accompanied by appropriate human oversight and accommodation procedures.

In short, AI can automate the hiring process, but it cannot automatically remove the employer's legal responsibility to conduct non-discriminatory recruitment.