Aftertalks #5 - Deep Dive into AI Agents, Hiring Bias & Gender Equity | AI-swers

Can an AI recruitment system be objective when it is trained on a biased history? In this AI-swers Aftertalks episode, we analyze how ChatGPT, Claude, Gemini, DeepSeek, Llama, and Grok responded when forced to choose between identically qualified job candidates whose only differences were demographic. The experiment produced two competing definitions of fairness. One group treated fairness as active correction. These models intentionally favored female candidates to counter historical inequality, improve representation, and support workplace diversity. The second group treated fairness as strict neutrality. These models removed gender and other demographic attributes from the decision boundary and used random selection because no merit-based difference remained. Neither approach is inherently neutral. Every recruitment system reflects choices made by its developers: which data is included, which outcomes are rewarded, how fairness is defined, and whether historical imbalance should be corrected or ignored. CHAPTERS 00:00 Amazon’s biased recruitment algorithm 01:30 Why algorithms are not automatically objective 03:00 The identically qualified candidate experiment 04:30 The six models split into two groups 05:00 Fairness as active correction 06:00 RLHF, training data and diversity policies 07:00 Role-specific arguments for female candidates 09:00 Fairness as strict neutrality 10:30 Removing demographics from the decision 12:00 Why random selection can be mathematically fair 13:30 What this means for real recruitment systems 15:00 The same candidate, two different algorithms 16:30 Active equity vs. procedural neutrality 18:00 Should AI balance the scales or flip a coin?