AI Job Description Generators and Discrimination Risk 2026
Bias enforcement has concentrated on the screening layer — the tools that rank and reject candidates. But the first automated step in most hiring funnels now happens earlier, in the job description itself, and almost nobody reviews it for the same exposure.
The Gap in the Compliance Story
Ask an HR team how they manage AI hiring risk and you'll usually hear about the assessment vendor: the bias audit, the accommodation notice, the disparate-impact statistics. All of that starts at the point a candidate has already applied. The job description is upstream of it — it determines who ever enters the funnel — and it is now routinely drafted by a generative tool in under a minute, edited lightly, and published.
Federal employment law has restricted discriminatory job advertising for decades. A posting that signals a preference or limitation based on age, sex, disability, national origin, or other protected characteristics is a problem on its own terms — no screening algorithm required. The awkward part is that generative models are trained on a corpus of real job ads, including decades of ads written before anyone reviewed them for this.
What Generated Job Copy Tends to Reproduce
Age-coded phrasing
"Digital native," "recent grad," "young and hungry," "high-energy team," "0–3 years experience" as a hard ceiling. These read as culture copy but function as age signals, and the ADEA protects applicants 40 and over.
Physical requirements that aren't essential functions
Models pad descriptions with boilerplate — lifting weights, standing for full shifts, driving requirements — that has nothing to do with the actual role. Under the ADA, a listed requirement you can't defend as essential invites a failure-to-accommodate claim and may deter qualified disabled applicants from applying at all.
Gender-coded adjectives
Clusters like "aggressive," "dominant," "competitive" versus "supportive," "nurturing," "collaborative" have been linked in research to skewed applicant pools. A generator optimizing for punchy copy leans on exactly these words.
Invented qualifications
Asked for a job description with thin input, a model fills the gap: degree requirements, years of experience, certifications nobody asked for. Inflated screens that don't track job performance are the textbook setup for a disparate-impact claim.
Proxy culture language
"Work hard, play hard," late-night team events, "no clock-watchers." Read as a signal about caregiving status, disability, or religious observance, this narrows who self-selects in.
Why "The Tool Wrote It" Is Not a Defense
Employment discrimination liability generally follows the employer that published the ad. The vendor's terms of service almost certainly disclaim responsibility for output, and the employer is the party that reviewed and approved the text. The same principle has been playing out in vendor-liability litigation over screening tools: adopting a third-party system does not move the employer's obligations onto the vendor, it just adds a second potential defendant.
There is a scale problem too. A recruiter who writes discriminatory phrasing does it in one posting. A prompt template that produces it does it in every posting, across every requisition, until someone notices — which is precisely the pattern that turns an individual claim into a pattern-or-practice theory.
The Essential-Functions Trap
The job description is not just marketing copy — it is the document an employer leans on later to define a role's essential functions in an accommodation dispute. Courts give weight to a written description prepared before the position was advertised. That cuts both ways: a description assembled by a model from generic boilerplate is a weak foundation for arguing a function was essential, and any junk requirement it invented is now an on-the-record employer statement about the job. Confirm essential functions with the hiring manager before publishing, not after a request for accommodation arrives.
Review Checklist Before You Publish
A five-minute review per requisition closes most of the exposure.
Your careers page is part of the hiring funnel too
A compliant job description on an inaccessible application form still excludes candidates. RatedWithAI scans your public pages for the accessibility and compliance gaps that turn into complaints. Start with a free scan.
Scan Your Site for Free →Frequently Asked Questions
Is it legal to use AI to write job descriptions at all?
Yes. No US law prohibits drafting a job posting with a generative tool. The obligations attach to the content that gets published and to the employment decisions that follow, so the practical question is what your review step looks like, not whether the tool is allowed.
Do bias-audit laws like NYC Local Law 144 apply to a description generator?
Generally no. Those statutes are aimed at tools that substantially assist in screening, scoring, or ranking candidates. A drafting assistant producing text a human approves usually sits outside the definition — but that is a scope limit on one statute, not a safe harbor from discrimination law generally.
What if the vendor advertises that its output is bias-checked?
Treat it as a feature claim, not a legal opinion. Ask what specifically is checked, get it in writing, and keep your own review step. Employer liability for a published ad does not transfer because a vendor asserted its filter works.
How would this surface as a claim in practice?
Most often as evidence rather than a standalone case. A rejected applicant's counsel pulls the archived postings, and age-coded language across every requisition supports an inference of intent that would be hard to build from a single hiring decision. Job ads are public and easy to preserve, which makes them attractive early discovery.
Should we disclose that a job post was AI-generated?
No general US requirement compels it for a job advertisement. The disclosure obligations that do exist in hiring focus on the use of automated tools in evaluating candidates, which is a different step in the process.