Skill.reDiversity and Bias in Sourcing: How AI Can Help and Harm
Overview
Lecture URL: https://skill.re/learn/recruiting/diversity-and-bias-in-sourcing-how-ai-can-help-and-harm.php
TRANSCRIPT: Diversity and Bias in Sourcing: How AI Can Help and Harm
Course: AI for Recruiters - Professional Credential
Module: Level 3: Independent Practice
Section: Chapter 11 -- Sourcing Support And Research
Theme: sourcing-support-and-research
Lecture: 11.2
Duration: 75 min
Format: Workshop + Case Studies
Audience: Experienced recruiters applying AI independently
Prerequisites: L2 Certification
What you will learn: Identify how AI sourcing strategies can inadvertently create demographic bias and implement safeguards for inclusive recruiting.
AI can expand your sourcing dramatically. It can find candidates in places you wouldn't look. It can identify skills across non-traditional backgrounds.
But AI can also encode and amplify bias. If your training data reflects historical hiring bias, AI will learn and scale it. A sourcing algorithm trained on your past hires (which were probably less diverse than your talent pool) will source people like your past hires.
This session focuses on understanding bias vectors in AI sourcing and implementing safeguards.
- *Understanding Sourcing Bias**
Bias in sourcing isn't malicious. It's embedded in:
- *Training Data:** If you train an algorithm on your past hires, and your past hires skew toward certain demographics, the algorithm learns that bias and scales it.
- *Platform Bias:** LinkedIn over-represents employed people in certain industries and geographies. GitHub over-represents certain coding languages and male developers. When you source from a single platform, you inherit its demographic skew.
- *Keyword Bias:** If you search for "Stanford graduate" or "5 years startup experience," you're creating demographic filters that correlate with privilege.
- *Network Bias:** Referral-based sourcing amplifies homogeneity. Your people refer people like them.
- *How AI Amplifies These**
AI doesn't create bias. But it can scale existing bias at remarkable speed.
Example: You build a sourcing algorithm trained on your past hires. Your past hires are 60% male. The algorithm learns to prioritize signals associated with hiring males. It scales this across thousands of candidates. You've just created a system that proportionally hires more men than women, without anyone explicitly coding that.
- *Safeguards for Inclusive AI Sourcing**
- *1. Audit Training Data**
Before using AI sourcing, examine what data you're training on. If you're training on your past hires: ask whether those hires are representative of your talent pool or reflect historical bias.
If your past hires skew toward certain demographics, either:
- Don't use them as training data, or
- Explicitly acknowledge the bias and correct for it in your model
- *2. Diversify Sources**
Don't rely on a single sourcing channel. If LinkedIn is 80% of your sourcing:
- Add GitHub for developers
- Add industry-specific forums (Dev.to, HashNode, etc.)
- Add university recruiting
- Add diversity job boards
- Add community groups
Different channels have different demographic distributions. Diversifying channels increases diversity of sourcing.
- *3. Be Explicit About Demographic Goals**
If you want a more diverse candidate pool, say so. Make it an explicit goal.
"Our goal is to source candidates from underrepresented backgrounds in tech. We'll source from diverse job boards, specifically recruit from historically Black colleges, and partner with diversity-focused communities."
That's fair. Candidates know you're prioritizing diversity. You're being intentional.
What's less fair: claiming your sourcing is "blind to demographics" while using algorithms that statistically produce homogeneous candidate pools.
- *4. Monitor Sourcing Demographics**
What demographics are you actually sourcing? If you source 1,000 candidates and they're 80% male, something is off.
Regularly check: what's the gender distribution of sourced candidates? Racial/ethnic distribution? Age distribution? Education background?
If sourced candidate pool demographics don't match your talent pool demographics, something in your sourcing is filtering.
- *5. Validate Keyword Choices**
When you search for specific keywords ("Stanford graduate," "startup experience," "executive presence"), ask: am I filtering intentionally or filtering bias?
If a role genuinely requires Stanford education, that's a choice you're making. More often, "Stanford" is a proxy for other things (selectiveness, networks, pedigree) that you actually need.
Better: identify what you actually need. Selectiveness? Look for strong performers from any background. Networks? Multiple backgrounds have networks. Pedigree? Define what that means and measure it. Changing from "Stanford graduate" to "demonstrated strong problem-solving ability with evidence of growth" opens your sourcing dramatically.
- *6. Include Non-Traditional Paths**
Bootcamp graduates, career changers, people with non-traditional backgrounds. These populations often bring diverse perspectives.
If your sourcing algorithm filters for "computer science degree," you're excluding bootcamp grads and self-taught developers. That's a choice. Make sure it's intentional.
- *7. Test for Bias**
Take two identical resumes. Only change the name--one traditionally male, one traditionally female. Or one Anglo-American name, one ethnic-minority name. Run both through your sourcing process.
Do they get different scores? Different likelihood of advancing? If yes, you have bias in your sourcing logic.
- *8. Document Your Sourcing Logic**
Why are you using the criteria you're using? What's your justification? Could you defend it to someone who accused you of bias?
If you can't defend it, change it.
ANTI-PATTERNS
- *Anti-Pattern 1: The Blind Bias**
- Description:* Claiming sourcing is "blind to demographics" while using algorithms that statistically exclude certain groups. *Why:* You believe the algorithm is neutral because it doesn't explicitly code race or gender. *What goes wrong:* Demographic filtering still happens, just covertly. *How to avoid:* Monitor what you actually source. If sourced pool lacks diversity, something is filtering.
- *Anti-Pattern 2: The Proxy Proliferation**
- Description:* Using multiple demographic proxies in sourcing without realizing it. Ivy League + startup experience + major cities = significant demographic filtering. *Why:* Each seems reasonable individually. Combined, they're restrictive. *What goes wrong:* You source an extremely homogeneous pool. *How to avoid:* Audit your sourcing logic. How many of your proxies correlate with demographics?
- *Anti-Pattern 3: The Training Data Bias**
- Description:* Training sourcing algorithms on your past hires, which contain historical bias. *Why:* Your past hires are available data. The algorithm performs well on historical data. *What goes wrong:* The algorithm learns historical bias and scales it. *How to avoid:* Audit training data for bias. Don't train on populations you know are demographically skewed.
PRACTICE PROMPTS
- Sourcing Audit: Analyze your last 100 sourced candidates. What's their demographic distribution? How does it compare to your talent pool? What's filtering?
- Algorithm Audit: If you use AI sourcing, examine the training data. What population was it trained on? Does it reflect your talent pool?
- Keyword Analysis: List your top sourcing keywords. For each, ask: am I filtering intentionally or filtering bias? What's my actual requirement?
- Resume Test: Take two identical resumes. Change only the name. Run through your sourcing process. Do they score differently?
- Channel Diversity: Analyze your sourcing channels. Are you over-reliant on one? What populations does each channel over- and under-represent?
KEY TAKEAWAYS
- AI scales existing bias at speed. If your training data is biased, your algorithm will be biased.
- Platform choice is demographic choice. LinkedIn sources differently than GitHub. Diversify channels.
- Keywords are filters with demographic consequences. "Startup experience" filters differently by demographic.
- Monitor what you actually source. Claim to be diverse while sourcing homogeneous pools. Measure demographics.
- Document and defend your sourcing logic. Could you defend it to someone accusing you of bias?
- Include non-traditional paths. Bootcamp grads, career changers, and diverse educational backgrounds bring different perspectives.
GLOSSARY
- *Sourcing Bias:** Demographic filtering in sourcing process.
- *Training Data Bias:** Using biased historical data to train algorithms.
- *Platform Bias:** Demographic distribution skew inherent in different sourcing platforms.
- *Demographic Proxy:** A criterion that correlates with and filters by demographics without explicitly stating it.
- *Disparate Impact:** Demographic filtering that occurs through seemingly neutral criteria.
- *Sourcing Diversity:** Intentional use of multiple channels and approaches to reach diverse candidate populations.
[SYNTHESIS AND APPLICATION]
AI sourcing can expand your reach dramatically. But it amplifies bias unless you build safeguards.
The practice: audit training data, diversify sourcing channels, monitor what you actually source, document and defend your sourcing logic.
Diverse sourcing isn't about political correctness. It's about accessing a larger talent pool and bringing different perspectives into your organization.
[REFLECTION EXERCISE]
- What demographic patterns exist in your sourced candidates?
- Are there sourcing channels you're not using that could expand your reach?
- If you audited your sourcing keywords, which ones would you eliminate?
- What non-traditional paths are you missing in sourcing?
- Could you defend your current sourcing logic to someone accusing you of bias?
[CLOSING REMARKS]
Inclusive sourcing is efficient sourcing. Access a larger pool of talent.
AI for Recruiters Certification Program
Level 3: Independent Practice | Sourcing Support And Research | Lecture 11.2
A SkillsClinic initiative.
Duration: ~75 minutes | Word Count: ~2200