A Job Posting Is a Role and a Route In—Verify Both

Use human review to verify job facts, challenge unsupported requirements, and test accessibility and application routes before publishing AI-assisted copy.

A Fluent Posting Can Still Describe the Wrong Job

Generative AI can turn scattered hiring notes into a polished job advertisement quickly. It can also smooth over disagreements, fill blank fields with plausible details, and turn a manager’s preference into a supposed requirement. The result may read professionally while misdescribing the work, narrowing the candidate pool, overstating pay or benefits, or sending applicants into an inaccessible process.

A job posting describes a real opportunity and the route into it. Every factual term needs an approved source, every qualification should connect to the work, and every link and form should function. AI can organize and phrase that material. It cannot decide what the job is.

This article provides an editorial and operational review workflow, not legal advice or a universal compliance checklist. It is anchored in U.S. federal guidance from the Equal Employment Opportunity Commission (EEOC), with useful practice guidance from the United Kingdom and voluntary risk-management guidance from NIST. State, local, sector-specific, union, contract, public-service, and non-U.S. rules may differ or add requirements. The workflow does not guarantee a bias-free or compliant result. Qualified reviewers should identify what applies to the actual employer, role, location, and publication channel.

1. Lock the Requisition Facts AI May Not Change

Start with a one-page posting brief that AI is not allowed to expand. Freeze the jurisdiction, employing entity, role, work location, employment type, approved pay information, approved benefits language, and application route. Include the publication channels and intended closing date. If any field is unresolved, mark it VERIFY rather than asking the model to make the copy feel complete.

• Jurisdiction: every place where the employee may work and any location-specific posting rules identified by the employer’s reviewers.
• Employer: the legal employing entity and the public-facing name approved for the advertisement.
• Role: approved title, reporting relationship, team, level, and position identifier.
• Location and type: onsite, hybrid, or remote boundaries; full-time, part-time, temporary, fixed-term, or another approved status.
• Terms: approved pay range or rate, schedule, travel, benefits description, and any contingencies.
• Route in: the exact application URL, alternative accommodation contact, deadline, and expected next step.

Treat a changed field as a review trigger. Clearance for one state does not automatically cover another, and “remote” does not mean “work from anywhere.” The public brand, employing entity, location, and eligible benefits may differ. Freeze the frame so each sentence has a source.

2. Build a Role-Source Packet

Have the people who own the facts assemble an access-controlled, versioned source packet. Human reviewers may consult the current role analysis, organization chart, manager’s duty inventory, HR classification record, compensation approval, benefits source, work-location policy, travel and schedule expectations, application instructions, accommodation route, and approved equal-opportunity language. Give the drafting system only a compact derivative containing the minimum approved, non-personal, non-confidential facts cleared for use with that tool. Do not upload the underlying HR records merely because the system is approved.

Record the owner, approval date, effective date, and recheck date for every source. Resolve conflicts before drafting. AI must not choose between a manager’s notes and a different current policy, and an old advertisement does not prove that a requirement remains valid.

Create a source ledger with one row per proposed claim. Map the role title, each duty, each qualification, pay, schedule, location, travel, benefits, application instruction, and accommodation statement to its approved source. Give each row a status of VERIFIED, VERIFY, or REMOVE. A model-generated sentence cannot be its own source, and a recruiter’s memory should not silently become a public employment term.

3. Separate Functions from Credentials

Describe what the person must accomplish before deciding which credential supposedly proves they can do it. For each duty, record the outcome, frequency, context, tools, dependencies, and consequence of nonperformance. Then have the hiring manager and HR distinguish essential functions from marginal or reassigned tasks.

The EEOC’s ADA employer guidance explains that a written job description prepared before advertising or interviewing can be evidence of essential functions. It is not conclusive by itself. Other evidence can include the employer’s judgment, the amount of time spent on a function, the consequences if it is not performed, the terms of a collective bargaining agreement, and the experience of current or past employees in the job. The actual work matters; polished wording does not turn a marginal task into an essential one.

Next, separate required qualifications from desirable ones. A required item should have a documented relationship to the role and a named owner willing to defend why it is necessary. Do not let AI add a degree, years-of-experience floor, driver’s license, lifting threshold, language requirement, software brand, or unrestricted schedule because those details are common in similar advertisements. Common is not the same as supported.

Write desirable criteria as genuine preferences, not disguised gates. Test whether scoring, screening questions, and recruiter instructions use the same distinction. A posting that says “preferred” while the applicant system automatically rejects anyone without it has a route that contradicts its words.

4. Verify Every Employment Term

Run a factual pass independent of the style edit. Check the approved title, entity, workplace, remote boundaries, schedule, travel, duration, pay, incentives, benefits, contingencies, deadline, and requisition number. This is a verification list, not a claim that every advertisement must legally state every field. Remove unapproved fields rather than filling them with industry norms.

Be especially literal with location. Distinguish “remote within named locations” from “fully remote,” and distinguish a current hybrid schedule from a permanent promise. Quantify travel only when the source packet supports it. If evening or weekend work is occasional, explain the operational context instead of presenting open availability as a personality trait.

AI must not invent salary, benefits, leave, accommodations, equal-opportunity promises, or a hiring timeline. It must not borrow terms from another role or turn conditional eligibility into a guarantee. Any linked detail must match the posting’s classification and location.

5. Look Beyond a Bias-Word Scan

Check more than a blacklist of words. The EEOC’s guidance on prohibited employment policies and practices notes that job advertisements may not show a preference or discourage applicants because of protected characteristics under the federal laws it enforces. Review the title, duties, qualifications, schedule, physical demands, images, targeting, distribution, and application questions together.

A phrase scanner can flag terms for human attention, but it cannot prove compliance. “Energetic” may be an ordinary adjective in one context and part of an age-coded pattern in another. Removing it does not cure an unsupported experience cap, an exclusionary ad audience, or a screening rule. Likewise, swapping “walk” for “move” is not enough if the team never examined what the function actually requires.

Ask reviewers to examine both express and implied messages:

• Does a requirement describe the outcome, or does it prescribe one way of performing it without a documented need?
• Could a cultural, educational, language, mobility, schedule, or location preference have been promoted into a gate?
• Do images and channel targeting imply a narrower audience than the written invitation?
• Does “culture fit” substitute for specific, job-related behavior?
• Do automated screening questions match only the verified required criteria?

The U.K. government’s 2026 guidance on making job descriptions inclusive, published for England, Scotland and Wales, is not U.S. law, but it offers useful editorial practice: use neutral language, review long “must-have” lists, describe behaviors instead of traits, distinguish essential from desirable criteria, use accessible formats, and offer a route to request adjustments. Apply it as writing guidance while qualified reviewers determine the employer’s obligations.

6. Verify the Accommodation Route and Final Rendering

The route into the role deserves the same scrutiny as the description. For employers covered by Title I of the ADA, the EEOC’s guidance for job applicants and the ADA describes reasonable accommodation in the application process, explains how an applicant may request a change, and notes that some arrangements need advance notice. As an operational practice, publish a usable employer-approved request route and make sure the responsible team receives and handles it.

Do not let AI promise a specific accommodation. Publish the employer-approved request channel without asking every applicant to disclose a diagnosis or medical details. Preserve an individualized process, verify who monitors and securely routes requests, and arrange coverage. A contact sentence does not by itself make the process accessible or responsive. Keep this route distinct from general recruiting support.

As an operational QA pass—not as an exhaustive statement of any legal or technical standard—render the actual posting in every channel. Check headings, focus, labels, contrast, zoom, link purpose, keyboard access, error recovery, and alternative text. Confirm that a screen reader encounters the role, key terms, apply action, deadline, and accommodation route sensibly. Test any third-party applicant system too; accessible prose cannot repair an inaccessible form.

7. Protect Applicant and Employee Data

Do not paste résumés, interview notes, employee examples, accommodation information, demographic data, or identifiable records into a general-purpose or unapproved AI system. Even with an approved tool, use only authorized, minimum-necessary inputs under the organization’s privacy, security, access, and retention controls. Draft this posting from the role-source packet without candidate or employee data; make examples fictional and non-identifying.

The U.K. government’s responsible AI in recruitment guide is also practice guidance rather than a universal rule. Its risk-management themes are useful: understand where AI is used, assign accountability, consider fairness and accessibility, manage data, test systems, provide information to affected people, and monitor performance. A team should apply the privacy, employment, procurement, and records requirements that govern its actual system and jurisdiction.

Inventory every AI-assisted step. This article covers publication review, not validation of automated employment decisions. If AI influences who sees the role or advances, that use needs separate qualified governance and testing.

8. Test the Real Application Journey

In a clean browser and on representative devices, follow the route from public card through apply link, form, upload, errors, confirmation, and accommodation contact. Test desktop, mobile, keyboard, zoom, and a screen reader. Confirm the deadline and job identifier survive every handoff.

Submit only authorized test data in a test-safe environment. Confirm required and optional fields behave as represented, errors explain recovery without erasing input, and expired sessions, unsupported files, duplicates, and confirmations behave predictably. Broken routing is a publication defect.

Review distribution copies too. Job boards may truncate location, drop the desirable/required distinction, or cache old pay. Record each rendered destination; correct it or remove it from release.

9. Worked Example: A Polished Draft with Unsupported Terms

Consider the fictional employer Harborline Civic Tools and a fictional Implementation Support Coordinator role. The approved packet says the employee will guide customer onboarding, maintain implementation records, and coordinate issue handoffs. It identifies two essential outcomes—running scheduled onboarding sessions and maintaining accurate records—and one marginal task: occasionally moving lightweight demo materials. The role is hybrid in Chicago on Tuesdays and Thursdays, includes up to 10% approved travel, and has an approved salary range of $62,000 to $68,000.

An AI draft renames it “Implementation Success Manager,” calls it fully remote, adds five years of experience and a 50-pound lifting requirement, advertises $70,000 to $80,000, and promises “industry-leading benefits.” None appears in the packet. It also asks for a “digital native” who can “work hard at any hour,” while the apply button points to a closed requisition.

The review does not merely soften those phrases. HR restores the approved title and salary. The hiring manager removes the invented experience and lifting gates, describes the actual outcomes, and replaces the years threshold with a source-backed desirable criterion: familiarity coordinating multi-step customer work. The location owner restores the Chicago hybrid schedule and travel limit. The benefits owner inserts only the approved summary. Accessibility testing catches an unlabeled upload button, and recruiting replaces the closed link. Legal reviews the complete rendered posting and distribution plan rather than a text fragment.

The corrected advertisement tells applicants what they would do, what is required or preferred, where the work occurs, which terms are approved, and how to apply or request an accommodation. The RELEASE record links claims to the packet and preserves the tested route.

10. Use a Bounded Drafting Prompt

A bounded prompt can expose missing fields and reduce silent completion; it cannot prevent invention or approve the output. For example:

Draft a clear job posting using only the approved model-ready facts below. Preserve the exact employing entity, title, location limits, employment type, pay, schedule, travel, benefits language, required criteria, desirable criteria, application URL, deadline, and accommodation contact. Do not add or infer duties, credentials, years of experience, physical demands, salary, benefits, equal-opportunity promises, accommodations, or remote-work terms. Mark conflicts or missing fields as [VERIFY] and list them after the draft. Keep required and desirable criteria in separate sections. Do not use candidate or employee data.

After generation, compare the draft with the ledger mechanically and manually. A constrained prompt reduces risk; it does not approve the result. Keep an access-controlled audit record that links to controlled sources and records the approved model-ready facts, prompt, tool version when policy requires it, generated draft, edits, and final rendered artifact. Do not co-locate raw HR files with the prompt or model output merely for convenience.

11. Name the People Who Can Stop Publication

Use named roles. The hiring manager owns duties and role requirements; HR owns classification, process, and approved terms; legal or another qualified employment reviewer evaluates the applicable context; accessibility owns the posting and application journey; and privacy and security own tool and data-use approval when protected or confidential information is in scope. Other specialists join when their fields are material.

The record should say RELEASE or HOLD, identify the version and destinations, name each approver, list conditions, and include tested application and accommodation routes. HOLD when a source is missing, owners disagree, a fact is unverified, a route fails, or the final rendering lacks approval.

The voluntary NIST AI Risk Management Framework 1.0 Core offers a useful lens for this structure: govern responsibilities, map context, measure identified risks, and manage them over time. It is not an employment-law safe harbor or certification. Its value here is operational: people define authority and tolerances, evidence is recorded, testing matches the context, and a human release decision remains visible.

12. Put the Posting on a Clock

Give the posting a version, source-packet version, publication time, owner, and recheck date. Expire it automatically at the approved close or review date. Trigger a new review when the duties, manager, location, classification, pay, benefits, schedule, travel, qualifications, legal language, application system, accommodation contact, or distribution channels change.

Monitor live destinations for stale copies and broken links. Give recruiters, employees, applicants, and job-board partners a clear correction channel. If a material error appears, place affected distributions on HOLD where practical, correct the canonical posting, update downstream copies, preserve the before-and-after record, and tell the necessary owners. Do not allow a corrected source page to leave an inaccurate cached advertisement recruiting people elsewhere.

Nine Questions Before This Posting Ships

• Are jurisdiction, employer, role, location, employment type, pay, benefits, and route in frozen?
• Does every duty, qualification, and employment term map to an approved source?
• Are essential and marginal functions, and required and desirable criteria, clearly separated?
• Were express and implied exclusion risks reviewed beyond a word scan?
• Does the accommodation route work and reach an accountable team?
• Was candidate and employee data kept out of unapproved drafting systems?
• Do the real links, forms, mobile flow, keyboard path, and screen-reader sequence work?
• Did the hiring manager, HR, legal or qualified reviewer, and accessibility owner record RELEASE?
• Are version, expiry, monitoring, and correction paths in place?

If any answer is no, polished copy is not a reason to publish. Keep the posting on HOLD, resolve the source or route, and test again. AI can help turn verified material into clearer language. It cannot define the role, decide that a requirement is job-related or lawful, authorize employment terms, or approve the candidate journey.

Scope Notes and Official Sources

• EEOC: The ADA—Your Responsibilities as an Employer
• EEOC: Prohibited Employment Policies and Practices
• EEOC: Job Applicants and the ADA
• GOV.UK: Make Job Descriptions Inclusive
• GOV.UK: Responsible AI in Recruitment Guide
• NIST: AI Risk Management Framework Core