cleanFiles/bookmap_Putting_AI_Ethics_into_Practice_at_SAP_en-US_learning-sap-com_lsc-prod.zip (MMCP-DROPZONE) - 2024-11-07T08:50:40.000Z

Ensuring Fairness in HR AI Systems - Discrimination-Free Decisions

Lesson Content

Putting AI Ethics into Practice at SAP

Explore the ethical considerations and strategies for ensuring fairness in HR AI systems. Learn how to mitigate biases and discrimination in decision-making processes. Enhance your understanding of AI ethics governance.<br><br># This description was auto-generated #

SAP_MOOC_aie2_Lesson_6_FAIR_Video.mp4

Welcome to the lesson, Ensuring Fairness in HR AI Systems, Discrimination-Free Decisions. My name is Khawla Mallat and I'm a member of the Corporate Sustainability Team, where I'm contributing to SAP's AI ethics governance. And I'll present the content of this lesson to you. We'll be first revisiting the same use case we discussed in the previous lesson. To quickly recap, our focus is on an HR-related use case involving a resume assessment tool tool that is designed to analyze candidates' resumes, comparing them to the job requirements in order to highlight the most suitable candidates. During the AI ethics impact assessment, this use case was classified as high risk, prompting a thorough review by the AI ethics steering committee. The committee then provided key recommendations highlighting the importance of transparency and explainability, which we covered in the previous lesson, but also the need for rigorous assessment and mitigation of potential biases. To provide you with some context, when the use case was first presented to the committee, it was initially framed more as a ranking task rather than a qualitative assessment. Also, the ranking was intended to be based on an aggregated score combined in various factors such as text similarity between the resume's text and the job description, work experience and skill matching scores. Given the risks of bias and discrimination inherent in this use case, a comprehensive analysis was conducted to examine how the system would rank similar profiles in terms of qualifications but with different demographic attributes. To address this, assessments were carried out with a focus on counterfactual fairness, which states that a decision is fair if a candidate would receive the same outcome even Even if hypothetically, only their demographic attributes were different, with all the other factors unchanged. This approach helps ensure that the AI system does not unfairly disadvantage any group based on demographic attributes. To conduct this analysis, synthetic resume profiles were generated using a Bayesian network, which allowed to model the relationships between different pieces of resume information. These profiles were then placed into a standard resume template to ensure that factors like, for instance, writing style didn't affect the results. Once these synthetic resumes were prepared, multiple versions were created of each profile, where the only difference was the demographic attributes. For instance, we take the exact same resume and use it in one scenario to represent a male candidate and in another scenario to represent a female candidate. Similarly, other versions were crafted to reflect different nationalities, religious beliefs, and sexual orientations. So here's how our standard resume template is structured. The fields highlighted in grey are filled with information related to education, skills, and work experience, which are the core qualifications that remain consistent across all versions of the resumes, and then the colored fields on the other hand are filled with information that infer demographic properties. For this analysis, the synthetic resumes were evaluated using the resume assessment scoring against different job descriptions. For simplicity, we'll only focus on results related to gender bias, specifically representing the two majority gender groups. The selected results presented here reveal some interesting patterns. In the box plot, we see that the scores received by female candidates for positions such as midwife, nurse, and professor are higher than those of male candidates. Ideally, if the model were unbiased, the scores would be equal. Additionally, when looking at the rankings, we observe that for roles like nurse, nine out of the top ten candidates are female, while ideally it should be even listed. Conversely, in sectors traditionally dominated by males, such as IT, we observe the opposite pattern. Male candidates receive higher score than females, and in terms of rankings, all top ten candidates for a database administrator role, for instance, are all male candidates, when ideally it should again be evenly split. These results suggest that the model may have inadvertently learned and reinforced gender stereotypes and job associations, leading to unfair outcomes in the resume assessment process. Based on these findings, a further examination of the different scoring components was conducted and as a result it was decided to only use the skills matching score to rank the candidates, as it was identified as the most biased neutral score. A more recent update has shifted from a ranking-based selection to a fit-based evaluation approach. This change was made recognizing that minor differences in scores can unfairly impact the hiring manager's decision, whereas grouping candidates in two bands such as anything above 80% being considered a great match for instance provides a fairer and more balanced hiring process. The evaluation experiments were rerun considering the changes focusing only on the skills matching scores. As a result we now see that the scores for male and female candidates are equal and the top 10 rankings are fairly split for jobs where females were previously advantage. The same improvement is observed for jobs that are stereotypically associated with males as well, indicating a more balanced and fair assessment. To sum up, in this lesson we explored how ethical risks like bias and discrimination in a high-risk use case such as SAP Fieldglass Resumease Assessment were thoroughly assessed and effectively addressed. These efforts in compliance with our AI ethics policy and the recommendations of the AI ethics steering committee ensure that the HR processes are fairer and more balanced for all individuals. This is the end of our lesson. Thank you for your attention and goodbye.