GB541 Unit 6 Seminar Reflection Example

Keeping a Screening Tool on Purpose: A Unit 6 Seminar Reflection on Algorithmic Hiring Risk

[Student Name]

Purdue University Global

GB541: Strategic HR Law and Risk Management

Unit 6 Seminar Reflection

[Instructor Name]

[Date]

The veterinary group, its vendor tool and the seminar classmates are composite; the selection figures are illustrative, and the federal rules described are real and summarized as of the course term.

The Position I Brought

The numbers seemed to settle the matter before the seminar started. Over twelve months, the tool scored 6,400 applicants for client service representative and veterinary assistant roles. It advanced 429 of 1,480 Black applicants, a rate of 29 percent, and 1,316 of 3,210 white applicants, a rate of 41 percent. Set against each other, 29 over 41 comes to 0.71. The federal selection guidelines treat a ratio below four-fifths as a sign of adverse impact that calls for scrutiny (29 C.F.R. § 1607.4, 2024), and once a disparity like that is shown, Title VII puts the burden on the employer to show the practice is related to the job and consistent with business necessity (Griggs v. Duke Power Co., 1971). I did not see how a vendor score built on resume features could carry that burden. My recommendation was simple: drop the tool, return screening to the hospitals and close the exposure.

The Question That Broke It

A classmate playing a practice manager waited until I was done and then put a single question to me: what did screening look like before the tool? I knew the answer from the case file, and I had not thought about it. Before the vendor contract, 58 hospital managers read resumes on their own, each with different habits and none keeping a record of who advanced and who did not. Some called back everyone who lived nearby, some favored applicants with prior veterinary work, and at least one admitted to skipping resumes with gaps. No one had ever calculated a selection rate because there was nothing to calculate from.

What That Meant

I spent the rest of the hour, and most of the evening, working out what that question implied. Switching off the tool would not remove the disparity. It would remove the measurement of it. The manual process almost certainly produced its own gap, possibly larger, but with no data the company could neither see it nor fix it, and an applicant who challenged it would face the same lack of records the company did. Raghavan et al. (2020) make a related point about vendor tools: the value of auditing an algorithmic screen depends on whether the employer can see how it performs across groups and adjust it, which is easier with a logged system than with dozens of individual judgments.

Eliminating the tool trades a disparity the company can measure for one it cannot, and the law does not reward an employer for choosing not to know.

That did not make the 0.71 acceptable. It meant the right response was to find what drove the gap and change it, not to return to a process whose bias was simply invisible.

Where I Ended Up

My revised position keeps the screen with three changes. First, the commute-distance feature comes out. Our review of the vendor's documentation showed that the tool scored applicants higher the closer they lived to the hospital, and in several of our metropolitan markets distance tracks neighborhood, which tracks race. Removing one feature is something the vendor can test before deployment, and it attacks the most likely cause of the gap. Second, HR runs a selection-rate review every quarter by role and by region, comparing advancement rates across groups and documenting what it finds, so that a ratio drifting below four-fifths triggers a review instead of sitting unnoticed for a year. Third, a recruiter reviews a sample of rejected applications every week, so a human looks at the people the tool screened out, not only at the ones it advanced. If the ratio stays below the benchmark after the feature is removed, the company should be ready to suspend the tool and validate it properly before using it again.

This position is harder to defend than the one I came in with. "We removed the biased tool" is easy to say to a board. "We kept the tool on purpose and are measuring it" invites the question of why the company accepts any risk at all. My answer is that measured risk under active control is lower than unmeasured risk, even when it does not look that way.

What Remains Uncertain

I am relying on the selection guidelines and Title VII doctrine as they stood during this course, along with Raghavan et al. (2020), whose review of vendor practices dates from 2019. Federal agency guidance on artificial intelligence in hiring has changed more than once in recent years, and several states and cities have passed or proposed rules on automated employment decisions, so the plan should be checked against the rules in force in each of Quillmere's seven states before the next contract renewal. I also do not yet know whether removing commute distance will close the gap. That is an empirical question the quarterly review will answer, and I would rather wait for the data than predict it.

References

29 C.F.R. § 1607.4 (2024). https://www.law.cornell.edu/cfr/text/29/1607.4

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

Raghavan, M., Barocas, S., Kleinberg, J., & Levy, K. (2020). Mitigating bias in algorithmic hiring. 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