You Cannot Detect It. You Can Still Be Sued For Acting Like You Can.
Application volume is up, recruiters believe most of it is machine-written, and the reflex is to buy a detector. The legal problem is not the preference for human writing. It is that the instrument enforcing the preference has a false-positive rate that lands hardest on applicants who write English as a second language — which converts a style preference into a selection procedure with a measurable adverse impact.
The Screen Is a Selection Procedure
Employers describe AI-detection screening as an integrity check, an authenticity filter, or a signal about effort. None of those characterizations change what it does mechanically: it takes a population of applicants and removes some of them from consideration. The Uniform Guidelines on Employee Selection Procedures define a selection procedure by function, not by label, and Title VII disparate impact analysis follows the selection rate. If the flagged-and- rejected group is disproportionately composed of applicants sharing a protected characteristic, the employer carries the burden of showing the screen is job related and consistent with business necessity.
That burden is unusually hard here, because the employer would have to establish that the detector measures something the job requires. It does not measure writing ability, domain knowledge, or honesty. It estimates the statistical likelihood that a passage resembles model output. Validating that against a job description is not a documentation exercise the employer has skipped — it is a claim most employers could not substantiate if asked.
Where the False Positives Land
Detectors infer machine authorship largely from statistical regularity — predictable word choice, low sentence-length variance, conventional structure. Those are also the signatures of competent second-language writing, of writing produced with grammar-correction tools, and of writing by people who learned professional English from templates. The result is a screen whose errors correlate with national origin and, through assistive technology use, with disability. An error rate that correlates with a protected characteristic is not a quality problem to be tuned away; it is the legal defect itself.
Non-native English applicants
The best-documented skew. Flagging is driven by the same low-perplexity features that characterize second-language professional writing, creating a national-origin-correlated rejection rate the employer never intended and usually never measures.
Applicants using assistive writing tools
Grammar, dictation, and drafting assistants used as disability accommodations produce text with the same regularity signature. A blanket screen removes these applicants before any accommodation conversation can occur.
Applicants coached by a career service
Workforce development programs, veteran transition services, and university career centers teach templated professional formats. Screening on regularity screens on having been coached, which tracks socioeconomic access rather than capability.
Older applicants using boilerplate
Long-tenured candidates often reuse a decades-old letter structure. Structural conventionality reads to a detector much like generation, adding an age-correlated dimension to the same error.
Anyone the recruiter simply suspected
The informal version of this policy — no tool, just a hunch about em dashes and tidy prose — produces the same exclusions with no audit trail at all, which is worse evidentially than an instrumented process.
The State AI Hiring Layer
Beyond Title VII, an AI-detection screen may fall within the state and local automated employment decision regimes. NYC Local Law 144 requires an annual independent bias audit and candidate notice for tools that substantially assist or replace discretionary screening decisions. Illinois HB 3773 brought AI use in recruitment and hiring within the Human Rights Act, including a prohibition on proxies for protected classes. Colorado's AI Act imposes developer and deployer duties around consequential employment decisions. Employers adopting a detector typically evaluate it as a vendor security purchase, not as an automated employment decision tool — which means the audit, the notice, and the impact assessment those regimes require are simply never scoped.
A Defensible Policy Instead of a Detector
- ☐Separate 'assisted drafting' from 'misrepresentation of experience' — only the second is a real integrity concern
- ☐Write the policy against fabricated credentials and unverifiable claims, which you can check, not against tooling, which you cannot
- ☐State the position publicly in the job posting so applicants can comply rather than guess
- ☐Replace the cover letter's screening weight with a structured work sample tied to the job description
- ☐Use live or synchronous exercises where authorship is observable rather than inferred
- ☐Validate the assessment against job performance and keep the validation evidence on file
- ☐Never let a detector score auto-reject; require a documented human review of every flag
- ☐Track selection rates by protected group before and after the screen and act on the gap
- ☐Ask the vendor for false-positive rates disaggregated by writer first language, in writing
- ☐Give every flagged applicant a route to contest, and treat that route as part of the process, not a courtesy
- ☐Determine whether NYC, Illinois, or Colorado obligations attach before deployment, not after a complaint
- ☐Provide candidate notice where required, describing the tool and the characteristics assessed
- ☐Retain application records and screening decisions for the full applicable retention period
- ☐Route disability-related exception requests to a named accommodation owner, not to the recruiter
Check what your careers pages tell applicants
Candidate notice obligations are satisfied on a page, and most careers sites were built before any of these rules existed. RatedWithAI scans your public pages and reports what an applicant actually sees.
Scan Your Site for Free →Frequently Asked Questions
Can we just say in the posting that AI-assisted applications will be rejected?
Announcing the policy helps on notice and on fairness perception, but it does not cure disparate impact. A clearly disclosed screen that excludes protected-group applicants at a higher rate is still a screen the employer must justify as job related. Disclosure changes what applicants can anticipate; it does not change the selection rate the analysis measures.
Our applicant tracking system vendor added AI detection as a feature. Does that shift the liability?
Not for the employment decision. The employer is the one making the hiring decision and generally bears Title VII exposure for the tools it deploys. Vendor litigation has also begun testing whether a screening vendor can be liable in its own right as an agent of the employer, which means the vendor's involvement can add a defendant rather than substitute for one. Contractual indemnity allocates cost between you and the vendor; it does not remove you from the claim.
What if we only use the detection score as one factor among many?
Weighting reduces but does not eliminate the issue. If the score measurably moves outcomes, it contributes to the selection rate and is part of the procedure under analysis. Courts examining multi-factor processes have looked at the components, particularly where a single component is the one carrying protected-characteristic correlation. 'One factor among many' is a mitigation argument, not an exemption.
Is there any version of this that is clearly safe?
Assessing the work rather than the writing. A structured, job-validated exercise administered to every candidate at the same stage is defensible on exactly the grounds the disparate impact framework asks about, and it makes the authorship of the cover letter irrelevant. Every safe version of this policy ends up being a better assessment rather than a better detector.