The Resumes AI Won't Touch — and the People Who Write Them
The Resumes AI Won’t Touch — and the People Who Write Them
Infantry team leader. Paramedic. Psych nurse. Crisis counselor. Millions of people have careers whose honest description — weapons, trauma, overdoses, self-harm — reads as “borderline content” to a default AI safety filter. We built a test for exactly those resumes. Here’s why, and what we found.
Every AI resume tool demos the same way: a marketer, a software engineer, a project manager. Clean bullets about dashboards and launches. Nobody demos the combat medic.
That’s not an accident. An honest military, EMS, clinical, or mental-health resume is full of words that AI content filters were trained to flinch at. The same sentence that wins you an interview — “triaged three casualties under direct fire” — pattern-matches, to a default-configured model, as violent content. The filter can’t tell a war story from a service record. And when it flinches, one of two things happens: the tool refuses outright, or worse, it quietly sanitizes — your naloxone reversals become “assisted individuals in challenging situations,” and your resume walks into the interview weaker than you are.
Pick your career below and see exactly what a default filter flags — and what our pipeline did with the same bullet:
The Refusal Index, v1
We assembled one deep, deliberately hard fixture per category — a full resume plus a matching job posting, written the way the work actually happened — and ran each through Bloom’s complete production pipeline: parse, tailor, and the line-by-line verification pass we run on every resume.
The result across all four categories: zero refusals, zero content blocks — 4/4 fixtures tailored normally, end to end, with every bullet verified against the source.
| Category | What the fixture describes | Result |
|---|---|---|
| Military | combat deployments, weapons systems, casualty response | ✓ 0 refusals |
| First responder | overdose reversals, fatality scenes, vehicle extrication | ✓ 0 refusals |
| Clinical | controlled substances, restraint protocols, code events | ✓ 0 refusals |
| Mental health | suicide-risk assessment, self-harm intervention | ✓ 0 refusals |
That’s the Bloom Refusal Index v1, from our 2026-07 test cycle. One full synthetic fixture per category — no user data — and the published index grows a category per revision. Have a career category you suspect AI tools mishandle? Suggest it: hello@bloomcareer.io.
Why this test exists at all
Because we got burned. On one mainstream model we’ve run in production, the default safety configuration treated exactly these categories as borderline — real careers, honestly described, sitting one classifier-shrug away from a refusal. We had to explicitly configure the filter to its most permissive documented tier to keep legitimate resumes flowing, and we’ve tested every model we ship against this corpus since. Our current production model passed 4/4 with zero special configuration.
The uncomfortable part: a tool that hasn’t run this test doesn’t know how it fails. The refusal isn’t logged as a bug. The sanitized bullet doesn’t throw an error. The veteran just gets a worse resume than the marketer, silently.
AI filters are the second wall. The ATS was the first.
If this pattern feels familiar to veterans and clinicians, it should — automated screening was already filtering these careers out before AI writing tools existed, just at a different layer. Roughly 99% of Fortune 500 companies run applicant-screening software that winnows resumes before a human reads them (Harvard Gazette), and those systems judge by rigid keyword and credential matching. Here’s how often that goes wrong, from the two best primary datasets on it:
- 88% of employers admit their own screening systems vet out qualified candidates — rising to 94% for middle-skills roles, the bracket most healthcare-support hiring sits in (Harvard Business School × Accenture, “Hidden Workers: Untapped Talent,” 2021).
- Veterans are 15.6% more likely to be underemployed than nonveterans — the downstream outcome of resumes that screening systems never surface (LinkedIn Veteran Opportunity Report, 2019).
The Harvard study names veterans explicitly: they go hidden because military skills — and the way military resumes describe them — don’t match the civilian keyword taxonomies employers configure. Licensed clinical fields hit the same wall from the other side: the credential acronyms and protocol names that make a nursing resume credible are exactly what a rigid keyword screen mis-scores.
Note what those numbers measure, honestly: keyword-and-criteria screening, not AI content filters. The ATS wall is at least studied — researchers can quantify it. The AI-refusal wall is newer, and to our knowledge nobody publishes data on it at all. That’s the gap the Refusal Index exists to fill, from the tool side, with our own results on the table.
If an AI tool waters down your service record
- Watch for euphemism swaps. “Challenging situations” for combat, “supported individuals in crisis” for suicide intervention, “medication management” for controlled-substance custody. If the output is vaguer than what you typed, the filter is editing you.
- Keep the concrete nouns. The 9-line MEDEVAC, the C-SSRS, the Pyxis count — specific systems and protocols are what make a licensed-field resume credible, and they’re exactly what sanitization deletes first.
- Demand receipts. A tool that verifies its output line-by-line against your source resume can show you what changed and why. A tool that can’t show you that has no answer to “where did my tourniquet bullet go?”
The standard we’d like to see
Every AI resume tool should publish its refusal testing: the career categories it tested, the configuration it ships, and the results. We’ll keep publishing ours — including the categories where we find problems. If your career keeps reading as “borderline content” to the tools that are supposed to help you, you deserve to know which ones actually engage with it.
This refusal corpus is one gate in the harness every model must clear before it ships under Bloom. In We Ran Resumes With Gemini vs. Anthropic. The Hallucinations Changed Shape. we put raw Gemini and raw Claude against our trained guardrail prompt and published the fabrication numbers — including the ones that made us keep the verification pass.
FAQ
Will an AI resume tool refuse a military resume?
Some will, depending on the model’s default safety configuration — military, first-responder, clinical, and mental-health language carries structural refusal risk. In Bloom’s 2026 testing, our production pipeline completed 4/4 categories with zero refusals; we test every model against this corpus before shipping it.
What does “sanitizing” a resume mean?
When a filter won’t refuse outright, it may quietly replace concrete, credential-bearing language (naloxone, restraint protocol, casualty care) with vague euphemisms. The resume still generates — it’s just weaker, and nobody tells you.
Do ATS systems reject military and medical resumes?
Not because of content filters — but rigid keyword screening does disproportionately miss them. 88% of employers admit their screening systems vet out qualified candidates (94% for middle-skills roles), the Harvard “Hidden Workers” study names veterans’ skills-translation mismatch specifically, and veterans run 15.6% more likely to be underemployed than nonveterans. AI content filters are a second, newer, unmeasured wall on top of that one.
Your service record deserves a resume that keeps it. Bloom tailors without sanitizing — and shows you a line-by-line verification of every bullet.