Your AI Wrote a Salary Range. That Range Is Now Your Position.
Pay transparency laws do not ask for a market estimate. They ask what you expect to pay, in good faith, for the role as posted. A generator that fills the field from benchmark data produces a number that reads identically and means something entirely different — and the posting is the record.
Where the AI Tool and the Statute Disagree
A job-description generator is optimising for a plausible, competitive-looking posting. It has access to scraped market data and to the words of your previous reqs. It does not have access to the approved band in your compensation system, the budget the finance team signed off, or the offer the last person in that role accepted. So it produces a range that is defensible as a market observation and unverified as a statement of intent.
The statute only cares about the second thing. Once published, the range is your representation of what you expect to pay. The tool has not made a mistake — it answered the question it was asked. The compliance gap is that nobody re-asked the right question before the posting went live.
The Failure Modes, Ranked by How Often They Ship
One posting, many jurisdictions. A remote listing is not one posting subject to one rule. Regimes that reach work performable in the state apply simultaneously, and the requirements are not identical — some want benefits described, some want the range in the posting itself rather than on request, some impose posting duties for internal promotional opportunities as well. Building to a single strictest standard is less work than maintaining a per-state matrix inside a generator's prompt.
The Discrimination Exposure Underneath the Posting Rule
Pay transparency statutes are the visible layer. The durable risk is that published ranges are now a comparable dataset. Anyone — a candidate, a current employee, a plaintiff's firm, an agency — can line up your postings for the same role across locations and dates and ask why they differ. Where an AI tool set the ranges from location-based market data, the answer is that the difference tracks local market rates, and that answer is only as good as whether the resulting pattern correlates with anything protected.
This is the same analysis pay equity law has always required, arriving through a new door. The relevant question for an automated posting workflow is not whether the model considered a protected characteristic — it did not — but whether the geographic or historical inputs it did use reproduce a disparity you would have to justify. Run the comparison across your own published ranges before someone else does; it is a spreadsheet, not an investigation.
A Workflow That Survives an Inquiry
- Make the band the input, not the output. The approved range enters the generator as a fixed field the model may format but not invent. If your compensation data cannot reach the drafting tool, the drafting tool should not be producing the range at all.
- Require a named approval before publish. One person, one timestamp, recorded against the req. This is the single artefact that answers most of what an inquiry asks.
- Put the range in a structured field. Not only in prose. Structured fields survive syndication; description bodies get truncated.
- Include the other required elements. Benefits and other compensation, a general description, and whatever your strictest applicable regime adds. Template them so they cannot be dropped.
- Re-approve on repost. Treat reopening a req as a new posting. Bands move; generated text does not know that.
- Audit live listings quarterly. Pull your own postings from the aggregators candidates use and check the range is present, current and consistent. Your careers page is not the evidence; the syndicated copy is.
- Compare ranges across locations. Same role, different geographies, same period. Ask whether the spread is explainable in a sentence you would be willing to write down.
Frequently Asked Questions
We are a small company with no employees in any pay transparency state. Are we exempt?
Only until you post a remote role. Employee-count thresholds vary and several regimes set them low, but the more common route in is the posting itself: statutes that reach work that could be performed in the state capture a fully remote listing regardless of where your company sits. Small employers also tend to have the weakest version of the underlying control, because there is no formal compensation band for the generator to be checked against — which makes the good-faith question harder to answer, not easier.
Can we post a range and then pay outside it?
Paying above the top is generally permissible if circumstances genuinely changed, and paying below the bottom is where the trouble is, because it directly contradicts what you said you expected to pay. The exposure is not usually a single offer; it is the pattern. If a meaningful share of offers land outside the posted range in the same direction, the range was not held in good faith at the time of posting, and every posting that used it shares the defect. Track the variance as a metric rather than resolving each case individually.
Our ATS vendor's AI writes the postings. Is compliance their responsibility?
No. The employer makes the posting and the employer owes the duty; a vendor's tool is an instrument you chose. What a vendor can usefully give you is the mechanism — a structured compensation field, an approval gate before publish, syndication that preserves the range, and an audit log of who approved what and when. Ask for those capabilities specifically in procurement. A tool that drafts a range with no way to bind it to an approved band is making your compliance harder while appearing to make your recruiting easier.
Do these rules cover internal promotions and transfers?
In several jurisdictions yes, and it is the duty most often missed because internal opportunities do not flow through the same posting machinery. Requirements to notify existing employees of promotional opportunities, or to disclose ranges for internal moves, sit alongside the external posting rule. If your AI drafting workflow only touches external reqs, map where internal openings are announced and apply the same range discipline there — the internal audience is also the one most likely to notice an inconsistency.
How long should we keep the posting and approval records?
Longer than the req is open, and to the longest retention period among the regimes that reach your postings — several impose multi-year duties covering job descriptions and wage-rate history. Keep the posting as published, the approved band with its effective date, the approver, and the syndicated copies you can capture. The reason to over-retain slightly is that these inquiries arrive well after the hire, and the artefact that resolves them is a dated approval nobody thought to save at the time.
Bind the Generator to the Band
There is no version of this problem where better prompting fixes it. A drafting tool cannot know what you intend to pay unless the approved band is an input it is not permitted to override.
Wire compensation data into the posting workflow, gate publication on a named approval, keep the range in a structured field, and pull your own live listings from the aggregators once a quarter. That is the whole control, and it is considerably cheaper than explaining a number nobody can account for.