Latest / AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence / When Your AI Recruiter Ignores Veteran Candidates
Transcript
- Lucas: So there's this quietly persistent problem in AI hiring that doesn't get nearly as much attention as gender or racial bias, and it's the way resume-scanning systems treat military veterans. Luna: You mean something like the AI literally filtering out people who served? That sounds like a huge oversight. Lucas: Exactly. And there's a solid 2025 study from the University of Maryland that quantified it: veteran applicants received twenty-three percent fewer callbacks when their resumes were processed by an algorithmic screener compared to a human reviewer. Same resume, just routed differently. Luna: That's a massive gap. What's going on under the hood that causes that? Lucas: Well, part of it is what I call the 'skill translation gap.' A lot of military job titles and responsibilities don't map neatly onto the keywords that commercial AI screeners are trained to look for. Things like 'platoon leader' or 'logistics coordinator in a combat zone' — the model might not have seen enough civilian equivalents to score them highly. Lucas: Another factor is career trajectory. Veterans often have non-linear paths: multiple deployments, gaps for training, lateral moves. The AI is typically trained on smooth upward progression, so anything that deviates looks like a risk signal. Luna: Right, so the model penalizes the very structure of a military career without understanding it. And if you're a veteran trying to break into tech or finance, you might not even know your resume is being ranked lower. Lucas: And that's the part that worries me. The screening happens before any human ever sees the application. So the bias is invisible — you never get feedback explaining why you weren't called. Luna: Speaking of invisible — this show stays ad-free because listeners chip in when they find value here. If today's tech conversation gave you something usable, the link is buy me a coffee dot com slash fexingo. No pressure, just a way to keep these episodes going. Lucas: Yeah, it really does make a difference. And speaking of making a difference — some companies are starting to audit their models specifically for veteran bias. Microsoft, for instance, reportedly retrained its screening tool after internal tests showed a similar pattern. Luna: What did they change? Did they just add more military keywords to the training data? Lucas: That's part of it, but there's a risk there. If you just keyword-stuff the model, you might get surface-level improvement without actually solving the structural bias. What Microsoft apparently did was include veteran career paths as positive examples in the training set, so the model learned that non-linear experience isn't a red flag. Luna: So it's not about adding keywords, it's about changing the model's understanding of what a good candidate looks like. Lucas: Right. And Amazon had a similar discovery a few years ago with their recruiting tool. They found that their model was penalizing resumes that included the word 'women's' — like 'women's chess club captain.' That's a different bias, but the fix was the same: you have to retrain the model on a more representative set of what success looks like. Luna: But here's the thing: veterans aren't a monolith. You have different branches, different roles, different eras of service. An Army infantry veteran from 2010 has a very different resume than a Navy IT specialist from 2020. Can a single model handle that diversity? Lucas: That's a really good point. And the short answer is: not well, unless you design the system with that in mind. Some companies are moving toward 'skills-based' screening that focuses on competencies rather than job titles. So instead of looking for 'project manager,' the AI looks for evidence of budgeting, stakeholder communication, deadline management — things that appear across industries. Lucas: But that's harder to implement because it requires more nuanced natural language processing. And most vendors are selling a one-size-fits-all solution that's optimized for the most common applicant profiles. Luna: So the burden ends up on the veteran to 'translate' their resume into corporate-speak. Which is doable if you know the system is biased, but many people don't. Lucas: Exactly. And that's where transparency comes in. If a company uses AI screening, should they be required to disclose what signals the model weights most heavily? Some advocates are pushing for that. Luna: Who's pushing? Are there any regulatory moves on this? Lucas: There's been some movement at the state level. Illinois passed a law in 2020 requiring employers to notify candidates if they use AI to evaluate video interviews. New York City's Local Law 144 mandated bias audits for automated hiring tools starting in 2023. But neither specifically addresses veteran status. Luna: So it's covered under generic protected classes maybe, but not called out. Lucas: Right. And that's part of the reason it flies under the radar. Veteran status is a protected class under federal employment law, but the EEOC hasn't issued specific guidance on AI and veterans the way it has on race and gender. Luna: What about the vendors themselves? Are any of the major HR tech companies — like Workday, or SAP SuccessFactors — doing anything proactively? Lucas: Workday, to their credit, published a white paper in 2024 outlining fairness principles and they've started offering bias detection reports to customers. But whether those reports drill down to veteran-specific metrics is unclear. I've seen their documentation mention 'military experience' as a potential bias attribute, but it's not always included in standard audits. Luna: So it's optional. Which means the companies that care about it will check it, and the ones that don't, won't. Lucas: Exactly. And that's the core problem. Without external pressure — either from regulators or from customers — there's little incentive for vendors to prioritize this. And for a veteran applying to a hundred jobs, the bias compounds across each system. Luna: I want to go back to the skill translation piece. You mentioned natural language processing limitations. Can you give me a concrete example of a military skill that an NLP model might miss? Lucas: Sure. Take 'convoy commander.' That involves logistics planning, risk assessment, personnel management, real-time decision-making under pressure. But if the model has never seen 'convoy commander' in a civilian context, it might assign a low relevance score. Meanwhile, a civilian 'logistics coordinator' with warehouse experience might score higher, even if the veteran's responsibilities were far more complex. Luna: That's a massive loss of talent for the employer too. They're filtering out people who've led teams in high-stakes environments. Lucas: Exactly. And there's research showing that veterans often outperform their peers in areas like leadership, adaptability, and integrity. So the bias isn't just unfair to the applicant — it's a bad business decision. Luna: So what would a better system look like? If you could design an AI screener from scratch, what would you do differently? Lucas: First, I'd include a diverse training set that explicitly includes military career paths — not just a few examples, but hundreds. Second, I'd use skills extraction rather than keyword matching. Third, I'd build in an explainability layer so that a recruiter can see why a candidate was ranked a certain way. And fourth, I'd run regular bias audits that include veteran status as a metric. Lucas: And honestly, I'd also consider a human-in-the-loop for the top ten percent of candidates that the AI almost rejected. That's where you catch the false negatives. Luna: So the AI becomes a triage tool, not a gatekeeper. That seems like a healthier design pattern in general. Lucas: It does. And it's the direction a lot of ethicists are pushing. The technology is good at ranking and filtering, but it's bad at making nuanced judgments about unconventional backgrounds. So we should use it where it's strong and keep humans where we need judgment. Luna: Is there any promising startup or nonprofit working specifically on this veteran ai hiring gap? Lucas: There's a nonprofit called 'Hire Heroes USA' that offers resume translation services — they help veterans convert their military experience into civilian language. But that's a human service, not an AI fix. On the tech side, a company called 'Eightfold AI' claims to do skills-based matching that's less prone to keyword bias, but I haven't seen independent audits of their veteran outcomes. Luna: So the landscape is still pretty fragmented. A few good intentions, a few regulations starting to appear, but nothing systematic yet. Lucas: That's exactly right. And given that we're heading into a tight labor market in the second half of 2026, employers who ignore this are leaving talent on the table. The veterans I've spoken to say they'd welcome the chance to be judged on what they can do, not on how well their resume matches a keyword list. Luna: I think that's a fair ask. And maybe the next step is for more companies to demand that their HR tech vendors include veteran bias in the standard audit package. Lucas: If listeners take one thing from this episode, I hope it's that. If you work in HR or procurement, ask your vendor: 'Does your bias audit include veteran status?' If they say no, that's a starting point for a conversation. Luna: And if you're a veteran applying for jobs, maybe consider having a civilian friend look at your resume and suggest translations for military terms. It shouldn't be on you to fix the system, but it might help in the short term. Lucas: Fair point. Hopefully in a few years, the AI will be smart enough to do that translation on its own.