Your AI Screener Is Probably Downranking Veterans — Without Ever Reading the Word
Employers who actively want to hire veterans are running screening tools that quietly work against them. Not through a rule about military service, but through the ordinary machinery of resume scoring: gaps penalized, unfamiliar titles unmatched, frequent moves read as instability. The company's careers page says "veteran friendly." The funnel data says something else, and protected-status law looks at the funnel.
A Military Resume Looks "Risky" to a Model Trained on Civilian Ones
Screening models learn what a strong candidate looks like from historical data, and that data is overwhelmingly civilian. Against that baseline, a service record reads as a series of negatives. Duty stations change every two or three years, which scores as job hopping. Deployment and transition produce gaps, which score as instability. An occupational specialty maps to no title in the taxonomy, so skill-matching returns a weak result for someone who ran a maintenance operation with a seven-figure equipment inventory. Training that took eighteen months and produced no civilian certificate simply does not register.
None of that requires a discriminatory rule. It requires only a scoring function that treats the ordinary shape of a military career as a defect. That is precisely the pattern disparate-impact analysis was built to detect, and it is invisible from inside the recruiting team, because the recruiter never sees the candidates the tool filtered out.
Where the Exposure Concentrates
- •Continuous-employment scoring that treats deployment as unexplained absence
- •Average-tenure features that punish standard rotation cycles
- •Recency rules that discount pre-separation experience
- •Reserve and Guard drill obligations read as availability risk
- •Transition-period gaps scored identically to voluntary unemployment
- •Occupational specialty codes unmapped to civilian job families
- •Military schools and certifications absent from the skills ontology
- •Leadership scope understated because rank is not parsed
- •Clearance status treated as noise rather than a qualification
- •Keyword matching that requires civilian tool names for identical work
- •Timed or gamified tests with no accommodation route at application
- •Video interviews scored on affect, penalizing service-connected conditions
- •Automated rejections issued before any accommodation request is possible
- •Vendor assessments never validated for this applicant population
- •No human review path for candidates who flag a barrier
- •Vendor cannot produce pass rates by stage for protected veterans
- •Self-identification data collected but never joined to funnel outcomes
- •Benchmark comparisons run on hires only, ignoring automated rejections
- •Retention periods shorter than the records obligation
- •Outreach commitments on the careers page unmatched by funnel data
The Marketing Claim Is Part of the Record
Veteran-hiring commitments are unusually public. Careers pages carry pledges, job listings carry badges, press releases announce partnerships, and procurement responses repeat all of it. Those statements are evidence. When funnel data shows protected veterans passing an automated stage at a materially lower rate than other applicants, the gap between the published commitment and the measured outcome is the first thing an investigator or plaintiff quotes back.
The fix is not to remove the commitment. It is to be able to show, with stage-level data, that the automated part of the process is not the thing undercutting it.
How to Actually Test For It
Two tests, run together. The first is an impact comparison: for every automated stage — parsing, ranking, knockout questions, assessment, scheduling — compute the pass rate for self-identified protected veterans against other applicants and apply the same ratio threshold you use for any other group. The second is a counterfactual audit: build matched resume pairs identical in substance but differing in whether the experience is written in military or translated civilian language, and measure the score delta. The first test tells you whether a problem exists; the second tells you which feature is causing it, which is the part you can fix.
Remediation Checklist
In the Tool
- ☐Map occupational specialty codes and military schools into the skills taxonomy
- ☐Remove or neutralize continuous-employment and average-tenure penalties
- ☐Parse rank and command scope as leadership evidence, not decoration
- ☐Disable knockout questions that no service history can satisfy
- ☐Require the vendor to expose pass rates by stage and by self-identified status
Around the Tool
- ☐Publish an accommodation route visible before the first automated assessment
- ☐Keep a human review lane for candidates the parser scores as unmatched
- ☐Join self-identification data to funnel outcomes, not only to hires
- ☐Retain stage-level applicant records for the full required period
- ☐Re-audit after every model update, not once at procurement
Frequently Asked Questions
We do not ask about military service. Does that end the analysis?
No. Service histories carry strong proxies — specialty codes, base locations, rotation patterns, deployment gaps — and a model can act on those without any status field. What matters legally is the effect on protected applicants, not whether a checkbox existed.
Is this only a federal contractor problem?
No. Contractor status adds outreach, benchmark and recordkeeping duties, but the underlying prohibition on denying employment because of uniformed service applies broadly, and many states protect military status independently at much lower employer thresholds.
Our vendor says the model is bias-tested. Is that enough?
Ask which groups were tested. Vendor bias testing frequently covers race and sex and stops there, leaving veteran status untested entirely. Request the tested categories in writing and run your own stage-level comparison on your own applicant data.
How do we handle service-connected disability in automated assessments?
Treat it as a separate obligation with its own path. Make an accommodation route visible before the first timed or scored assessment, and ensure no automated rejection can fire before a candidate has a realistic opportunity to use it.
What records will we be asked for?
Stage-by-stage disposition data for applicants, tied to self-identification where collected, retained for the applicable period. If the screening vendor cannot export outcomes per stage, you cannot produce the record, and that gap is itself the finding.
We publicly committed to veteran hiring. Does that increase our risk?
The commitment is not the risk; the unmeasured funnel behind it is. Public pledges become evidence when measured outcomes contradict them, so the sensible response is instrumentation rather than quieter marketing.
Your Careers Pages Are Part of the Evidence
Veteran-hiring pledges, job listings, application forms and accommodation notices are the documents a candidate and an investigator read first. If the application flow has no visible accommodation route, or the pledge is not reflected in the process it describes, that is discoverable long before anyone examines the model.
See what your site currently says. Run a free scan and review every page in your hiring funnel.