This is part 2 of the "AI for Recruitment: What's Actually Changing" series. In part 1, we looked at the tools that are reshaping how recruiters work. Now we're getting into the harder question: if AI can save us 75% of our screening time, what are we actually trading for that speed?
AI resume screening is everywhere. Over 82% of large corporations now use it to shortlist candidates, and the productivity numbers are real—companies report saving $2,300 to $3,000 monthly and cutting screening time from 60 hours down to 8. But I want to be honest with you: the technology that's making recruiting faster is also making it easier to systematically exclude people without knowing why.
Let me walk you through how this actually works, why it fails, and what you need to do before deploying it.
How AI Resume Screening Actually Works
When you hear "AI resume screening," you're hearing three separate processes working together.
First comes parsing. The system ingests a resume—PDF, Word doc, whatever—and uses natural language processing (NLP) to convert unstructured text into organized information: job titles, employment dates, technical abilities, academic background, professional credentials. It's essentially automating what a person would manually note while reading through the document carefully.
Second is feature extraction. The AI system examines the parsed data and determines what factors are relevant to the position. Does the candidate possess the terminology mentioned in the job posting? Do their previous roles align with what you're looking for? How many years have they spent in comparable positions? This process creates a candidate summary based on criteria the system identifies as important.
Third is ranking via machine learning. The system takes that candidate summary and compares it against patterns discovered in historical hiring records—"candidates meeting these qualifications tended to succeed here"—and assigns it a score. Usually you get a percentile ranking indicating how the resume stacks up against others.
In theory, this delivers speed and consistency superior to a person evaluating hundreds of resumes quickly. AI resume screening can process documents rapidly and consistently apply the same evaluation criteria, with reported reductions in screening time of up to 75%—dropping from 60 hours to approximately 8 hours per hiring batch.
But here's where it gets thorny: that third step, the ranking, is where bias lives.
The Amazon Case: Why AI Hiring Failed
In 2018, Amazon famously scrapped its internal AI recruiting tool after discovering it was systematically penalizing resumes containing the word "women's."
This wasn't a glitch. It was the predictable result of how the system learned.
Amazon trained the tool on 10 years of historical hiring decisions—predominantly from a male-dominated engineering department. The AI absorbed the statistical characteristics of "resumes that led to hires," which in this case reflected male-skewed language patterns and typical male career paths. When resumes mentioned "women's chess club" or graduation from women's colleges, the system didn't explicitly blacklist the word "women's." Instead, it had learned that word clusters associated with these experiences correlated with rejection in the historical record, since women possessing these backgrounds hadn't been hired previously.
Amazon attempted fixes. They tried removing "women's" from problematic word lists. But that approach missed the fundamental issue: you cannot filter your way out of prejudice embedded in your source data. The problem isn't in the software logic—it's in the learned patterns themselves. Multiple hidden factors could replicate the same discriminatory effect.
As Amazon eventually recognized, if you instruct an algorithm to find resumes resembling those in your historical dataset, you'll inevitably recreate the composition of your existing workforce. Solving this would have demanded reconstructing the training dataset with representative hiring records.
They decided to discontinue the project instead.
The Regulatory World Is Catching Up
Three separate regulatory frameworks now treat AI resume screening as a significant compliance and civil rights matter.
NYC Local Law 144 (effective since July 5, 2023) mandates that any employer using AI to evaluate job applicants in New York City must:
- Obtain an independent third-party bias assessment each year (specifically testing for racial and gender-based disparities)
- Make assessment findings available to the public on their corporate website
- Inform prospective applicants before submission that an AI system will evaluate their materials
- Provide candidates with the ability to request human evaluation instead
That final requirement is significant: The legislation recognizes that algorithmic screening creates legitimate concerns for applicants.
The EEOC has shifted its stance from treating AI hiring as a theoretical future concern to actively addressing documented violations. In August 2023, it resolved an age-discrimination case against iTutorGroup, whose screening technology automatically rejected women aged 55+ and men aged 60+, eliminating over 200 candidates from consideration. The resolution included a $365,000 payment to affected parties.
More significantly, in May 2025, a federal judge in California granted preliminary approval for a class action suit (Mobley v. Workday) representing all applicants aged 40 and older rejected by Workday's screening algorithm since 2020. This case underscores how automated hiring systems can perpetuate age-based discrimination.
The EU AI Act, rolling out enforcement in stages through 2026, designates recruitment technology as high-risk and establishes substantial financial penalties for non-compliance.
In summary: government agencies have moved beyond viewing AI hiring as neutral technology. They're now treating it as a documented civil rights hazard.
What the Data Says About Real Bias
A significant concern companies raise is whether AI screening might eliminate qualified applicants. This worry reflects actual patterns identified through bias testing and legal cases.
Bias in AI resume screening frequently emerges as proxy variable bias. When systems optimize for specific terminology or job titles, those same words or titles may correlate strongly with demographic characteristics. A resume full of quantifiable achievements ("delivered $2M in cost savings") might rank higher than one emphasizing collaborative impact—and writing styles and vocabulary choice frequently track with demographic categories. You're not explicitly screening by demographic group; you're screening according to communication patterns that happen to track with demographic groups.
The system absorbs these associations from your company's historical hiring records. If your previous hires concentrated in particular universities or sectors, the AI will learn to favor those backgrounds. If your hired workforce has consistently reflected certain characteristics, the algorithm will learn to repeat that composition.
Before You Deploy: What You Need to Do
If you're considering AI resume screening, here's a practical checklist:
- Examine your historical hiring data. Look at who your company hired previously. What is the demographic makeup? Document this clearly and specifically.
- Screen for disparate impact. Use sample batches to assess your tool's performance: does it reject female applicants at higher rates? Older workers versus younger ones? This represents the NYC Law 144 requirement, and it's prudent to conduct this analysis before implementation, not afterwards.
- Establish confidence intervals rather than binary decisions. Don't let the system make simple pass/fail determinations. Work with confidence ranges. "Ranked in top 10% of applicants" provides more nuance than "qualified" or "not qualified."
- Incorporate manual evaluation throughout. Design the process to include human judgment from the beginning, not as an afterthought.
- Be transparent with candidates. Inform applicants that AI will be used in evaluation before they submit materials. Provide alternatives or opt-out options. This is required by law in New York and represents best practice everywhere.
- Review performance on an ongoing basis. Patterns can shift over time as your hiring data evolves. Regular evaluation should become standard practice rather than a one-time exercise.
The Honest Takeaway
AI resume screening delivers real results. The time savings are documented and substantial. The bias risks, however, are equally real—and they're not exceptions. They're inherent to how machine learning systems learn from past decisions.
This technology isn't disappearing. Implementing it without examining your underlying data, establishing protective measures, and communicating transparently with applicants introduces genuine legal and reputational risk. Put in the groundwork beforehand. Your legal department will appreciate it, and your applicants will too.
Next in this series: we're looking at what happens after screening—how AI is transforming interviews, skill assessments, and candidate evaluation. Things get even more nuanced from there.
References
[1] Reuters. "Amazon scraps secret AI recruiting tool that showed bias against women." October 2018. https://www.reuters.com/article/idUSKCN1MK08X/
[2] U.S. Equal Employment Opportunity Commission. "iTutorGroup, Inc. Agrees to Pay $365,000 Settlement for Alleged Discriminatory Hiring Practices Based on Age." August 2023. https://www.eeoc.gov/newsroom/iTutorGroup-inc-pay-settlement-allegedly-discriminatory-hiring-practices
[3] Law360. "Workday Faces First AI-Hiring Age-Bias Class Cert Grant." May 2025. https://www.law360.com/articles/1924762/workday-ageism-class-certification-granted
[4] McKinsey & Company. "The State of AI in 2025." January 2025. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-in-2025
[5] New York City Council. "Local Law 144 of 2021 (Bias Audit Law)." Effective July 5, 2023. https://legistar.council.nyc.gov/LegislationDetail.aspx?ID=4599046
[6] European Commission. "EU AI Act." Regulation (EU) 2024/1689. https://eur-lex.europa.eu/eli/reg/2024/1689/oj