An AI Tool Drafted the Job Ad. The Employer Still Owned the Legal Risk.

An employer used an AI tool to generate job advertisements that restricted consideration to applicants with H-1B, Optional Practical Training, or H-4 status. The Justice Department’s response was direct: responsibility did not depend on whether the unlawful language came from an employee, a recruiter, or an AI system.

The matter ended in a settlement, not a court judgment. But the agreement with Elegant Enterprise-Wide Solutions offers a concrete employment-AI lesson. A company that automates drafting still owns the resulting advertisement, the approval process, and the records needed to show compliance.

That lesson reaches beyond citizenship-status restrictions. Recruiting teams increasingly use generative tools to write job descriptions, suggest qualifications, tailor outreach, screen language, and republish postings across platforms. Each step can introduce restrictions or preferences that nobody deliberately chose. The employer remains the legal actor.

What DOJ Found

The Justice Department’s Immigrant and Employee Rights Section investigated a charge filed in December 2025. According to the settlement agreement, the section determined there was reasonable cause to believe Elegant Enterprise-Wide Solutions had posted discriminatory advertisements in violation of 8 U.S.C. § 1324b.

DOJ said the advertisements were generated by an AI tool and included citizenship-status restrictions that were not authorized by law. The language limited consideration to people with specified visa or work-authorization categories, including H-1B, OPT, or H-4 status.

The parties resolved the investigation without litigation. The company agreed to pay a $9,460 civil penalty and accept three years of compliance obligations. Those obligations include policy review and revision, training for personnel involved in recruiting or approving job advertisements, posting worker-rights notices, preserving training records, and responding to DOJ compliance inquiries.

The settlement does not establish judicial liability, and it should not be described as a court ruling about generative AI. It records DOJ’s reasonable-cause determination and the company’s agreement to resolve the investigation.

That distinction matters. So does the remedy. DOJ did not treat the use of AI as a reason to narrow the employer’s responsibility.

Drafting Automation Is Still Part of the Hiring Process

Many organizations treat job-ad drafting as low risk compared with applicant scoring or automated interviews. The output is public-facing text, and a recruiter may review it before publication.

The Elegant settlement shows why that assumption is too comfortable. A job advertisement defines who is invited to apply. A restriction in the posting can exclude workers before any résumé reaches a screening system.

Generative tools can create that problem in several ways. A prompt may contain an unlawful preference. A model may infer one from prior advertisements. A recruiter may ask the tool to make a posting “more targeted” without understanding the resulting legal line. A platform may reuse language from a template designed for a different role, location, or contractual requirement.

None of those paths changes who published the advertisement.

Human Review Has to Be Specific

It is easy to add a policy saying that a person must review AI-generated content. That control works only if the reviewer knows what to look for and has enough context to recognize a problem.

A useful review should test the proposed posting against the actual role. Does a law, contract, export-control rule, security requirement, or customer obligation support the restriction? Is the language broader than the requirement? Does it refer to citizenship when work authorization would be the relevant question? Has a legal or HR owner approved any exception to the normal nondiscrimination language?

The reviewer should also see what the AI changed. If a system silently rewrites a previously approved template, a final read may miss a new limitation buried among ordinary qualifications. Redline review or structured exception flags can make the control more reliable.

The point is not to require legal review of every sentence in every posting. It is to design the workflow so that legally sensitive terms receive attention before publication.

Employers Need Provenance for Job-Ad Text

When a challenged advertisement appears, the company should be able to reconstruct how it was created.

That record may include:

the original approved template;
the prompt or structured inputs supplied to the AI tool;
the generated draft and later revisions;
the person who reviewed and approved the posting;
any stated reason for citizenship, work-authorization, location, credential, or experience restrictions;
the platforms and dates on which the advertisement appeared; and
the model, vendor, or system version used to generate the text.

Without that provenance, the company may know what was published but not why the language appeared or who could have caught it.

Retention also matters. Recruiting platforms and generative tools may overwrite drafts, retain only the final version, or store prompts outside the employer’s normal personnel systems. Legal, HR, and technical teams should decide which records must be preserved and where they will live.

Vendor Controls Do Not Replace Employer Controls

A vendor may advertise safeguards for biased or discriminatory language. Those safeguards can help, but they do not transfer the employer’s legal responsibility.

Contracts should address access to prompts, outputs, edit histories, model-change notices, and incident support. Employers should know whether the vendor trains on customer job advertisements, reuses prior postings as examples, or applies automatic targeting language. They should also know whether administrators can prohibit sensitive terms or require an approval step when those terms appear.

The company still needs its own testing. A recruiting team can run representative prompts through the tool and inspect whether it introduces citizenship preferences, age-coded language, disability-related assumptions, gendered requirements, or other unjustified restrictions. Testing should include attempts to make a posting more selective, urgent, local, senior, or “culture fit” oriented, because vague optimization requests can produce legally risky shortcuts.

A Practical Review for Recruiting Teams

Organizations using AI to draft or revise job advertisements should review five parts of the workflow.

First, identify every tool that can generate or alter recruiting text. Include writing assistants embedded in applicant-tracking systems, browser extensions, general-purpose chatbots, and outside recruiting platforms.

Second, establish approved source templates and prohibited language. Sensitive restrictions should require a documented, role-specific basis rather than free-form drafting.

Third, make review visible. Preserve the generated version, the final version, the approver, and any exception rationale.

Fourth, test the system with realistic prompts. Look for restrictions that appear even when nobody directly asks for them.

Fifth, train the people who draft and approve postings. The Elegant agreement specifically covers personnel involved in recruiting, drafting or approving job advertisements, screening applications, and hiring. Training should connect nondiscrimination rules to the tools people actually use.

The Enforcement Lesson

The Elegant settlement is modest in dollar terms, but its compliance structure is more instructive than the penalty. DOJ required policy changes, training, records, notices, and three years of oversight after an AI-generated advertisement allegedly crossed a legal line.

The practical rule is simple: automated drafting does not create automated immunity. If a company publishes a job advertisement, it should be ready to explain the source of any restriction, show who approved it, and prove that its review process was designed to catch unlawful language.

An AI tool can draft the text. The employer still owns the decision to use it.