Can ChatGPT Build a Salary Range? Why AI Numbers Fail Compliance

The Friday Afternoon Posting Problem

It is Thursday afternoon. You have a job posting due tomorrow — a hybrid Marketing Coordinator role that will be visible to applicants in California, New York, and Colorado. Your state requires a salary range in the listing. You open a browser tab, type a job title into an AI chatbot, and within ten seconds you have a number: "$58,000–$72,000."

Fast. Frictionless. Done — or so it feels.

Now imagine your employment attorney sends you an email six weeks later. A job seeker has filed a complaint under California's SB 1162. The civil penalty is $100–$10,000 per violation, with each non-compliant posting potentially treated as a separate violation (California Legislative Information, 2022 — verify the current rule at the California DIR before acting). The attorney's first question: "What was your methodology for the range you posted, and what data source supports it?"

You open the chat log. The AI gave you a number. It cited nothing. It listed no dataset, no reference year, no geography, no percentile anchor. There is nothing to hand to counsel, nothing to submit to a regulator, and no way to demonstrate that the number reflects what the market actually pays for that role in those three states.

That is the compliance case against using AI to build salary ranges — and by the end of this article you will understand exactly why AI-generated wage figures fail, what a defensible range requires instead, and how to build one without starting from scratch.

---

How AI Language Models Actually Generate Salary Figures

To understand why AI salary figures are unreliable for compliance purposes, it helps to understand what an AI language model is actually doing when it answers a compensation question.

A large language model (LLM) — the technology underlying tools like ChatGPT and Claude — is trained on a broad corpus of text from the internet, books, and other sources. It learns statistical patterns in language. When you ask "what is the salary range for a Marketing Manager in Denver?", the model does not query a live database. It generates a response that is statistically consistent with text it has seen during training.

That means the number it produces is, in effect, a confident-sounding synthesis of whatever compensation-adjacent text happened to be in its training data — job postings, salary-survey summaries, news articles, forum discussions, blog posts — with no disclosed weighting, no documented vintage, no geographic precision beyond what you typed, and no methodology a human can audit.

Three compliance-critical problems follow from this architecture.

First: hallucination risk on wage figures. LLMs can produce plausible-sounding numbers that are simply wrong. Unlike a structured database query, there is no lookup against a canonical record. The model has no mechanism to distinguish a wage figure from a 2019 regional survey from a current national median. It will present both with equal confidence.

Second: no data vintage. A defensible salary range documents when the market data was captured. The BLS Occupational Employment and Wage Statistics (OEWS) program — the standard authoritative source for this purpose — publishes annual releases with a specific reference year (the May 2024 national, state, and metro estimates, for example). An AI response carries no date. You cannot tell whether the figure reflects pre-pandemic labor markets or the most recent survey cycle.

Third: no audit trail. An audit trail for compensation purposes means a documented, reproducible record of the data source, the reference year, the geographic scope, the occupation taxonomy used, and the methodology by which you arrived at a min, midpoint, and max. An AI chat log is none of those things. It cannot be submitted to a regulator as evidence of methodology. It cannot be cited in a response to a demand letter. It does not satisfy the record-retention requirements that several transparency laws now impose.

Illinois, for example, requires employers to retain pay-scale information and the posting for each position for five years (Greenberg Traurig / Illinois DOL, 2024 — verify the current requirement at the Illinois IDOL before acting). Ontario's rules, effective January 1, 2026, require employers to retain each publicly advertised posting for three years after it is taken down (HRPA, 2026 — verify with the Ontario Ministry of Labour before acting). A chat log that you cannot reproduce and that contains no source citation is not a record.

---

What "Grounded in BLS Data" Actually Means

The phrase comes up often in compensation work, and it is worth being precise about it.

The BLS OEWS program is the U.S. Bureau of Labor Statistics' Occupational Employment and Wage Statistics survey. It produces wage estimates annually for more than 800 occupations, constructed from a sample of approximately 1.1 million establishments (BLS, May 2025). Estimates are reported at five percentile points: 10th, 25th, 50th (the median), 75th, and 90th.

A percentile is the wage level below which a given share of workers in that occupation and geography earn. The 50th percentile (median) means half of workers in that role and area earn less, half earn more. The 25th percentile means three-quarters earn more. This structure matters because a defensible salary range is anchored to a specific point in the market distribution — not a vague "typical" figure.

Consider a concrete example using library figures. For Marketing Managers (SOC 11-2021), the BLS OOH May 2024 national figures show a median of $161,030, a 10th percentile of $81,900, and a 90th percentile of $239,200 (BLS OOH, May 2024 — confirm the current live release at bls.gov/oes). Those are not estimates from blog posts or forum discussions. They are survey-derived statistics from a defined reference period, with a publicly documented methodology, available for download and citation.

When you build a range from BLS OEWS data, you can write down: "The midpoint of this range is set at the 50th percentile for SOC 11-2021 (Marketing Managers) in the [state/metro] area, per BLS OEWS May 2024. The range spread — the width of the band expressed as a percentage of the midpoint — is 50%, producing a minimum of $X and a maximum of $Y." That sentence can be handed to an attorney, submitted to a regulator, or stored in an audit file.

An AI-generated number cannot produce that sentence. There is no SOC code to cite. There is no percentile. There is no reference year. There is no methodology.

For a deeper look at how BLS data compares to proprietary survey sources, see our breakdown of BLS vs. Payscale data and Glassdoor vs. BLS salary data.

---

Can ChatGPT Build a Salary Range? A Practical Test

The honest answer is: it can produce a number that looks like a salary range. Whether that number is defensible, accurate, or jurisdiction-appropriate is a different question — and the answer to each is generally no.

Here is what a compliance-ready range needs that an AI response will not contain:

Requirement	AI chatbot response	BLS-grounded range
Named data source	✗ None	✓ BLS OEWS (dataset named)
Reference year	✗ None	✓ May 2024 (or current release)
Geographic specificity	✗ Inferred from prompt	✓ National / state / metro
Occupation taxonomy (SOC)	✗ None	✓ SOC code documented
Percentile anchor	✗ None	✓ 50th pct / 25th–75th / etc.
Range spread methodology	✗ None	✓ Documented % of midpoint
Reproducible	✗ No	✓ Yes (same query = same data)
Submittable to regulator	✗ No	✓ Yes

The compliance exposure is real. Colorado's Equal Pay for Equal Work Act carries fines of $500–$10,000 per violation, with each non-compliant posting treated as a separate violation; as of July 1, 2024, 1,634 complaints had been filed and $238,000 in fines assessed (Trusaic citing Colorado CDLE, 2024 — verify the current enforcement posture with the Colorado CDLE before acting). New York City's pay transparency law carries civil penalties of up to $250,000 per violation, enforced by the NYC Commission on Human Rights (Trusaic, 2025 — verify with the NYC Commission on Human Rights before acting).

The point is not to alarm you. It is to be precise: the penalty for a non-compliant posting in a covered jurisdiction is a documented, jurisdiction-specific figure — not hypothetical. An AI-generated range that you cannot defend does not reduce that exposure.

---

The Status Quo Most SMBs Are Actually Using

Before we talk about the right tool, it is worth naming where most small HR teams actually are: a Google Sheet, a BLS.gov tab, and forty-five minutes of manual lookups per role.

That approach uses the right data source. BLS OEWS is publicly available and free. The problem is what happens around the data: no audit trail documents the vintage, the percentile chosen, or the spread methodology; the calculation lives in a spreadsheet that may be overwritten before the next audit; and the output is a number in a document, not a formatted, sourced, compliance-ready range.

Replacing that workflow with an AI chatbot does not solve any of those problems. It replaces a free but laborious process that uses authoritative data with a fast but undocumented process that does not. You trade hours for accuracy — and you are trading in the wrong direction.

For a structured walkthrough of how to build a range the right way, see our guide on how to build a salary range and the companion piece on building a compliance audit trail for salary ranges.

---

What a Defensible Range Requires (and What Makes Building One Faster)

A compliance-ready salary range has four documented components:

1. Market anchor — the BLS OEWS percentile (typically the 50th, sometimes the 25th or 75th depending on your labor market positioning) for the relevant SOC code and geography, with the reference year stated.
2. Range spread — how wide the band is, expressed as a percentage of the midpoint. A typical individual-contributor range runs 40–60%; a senior or management role may run wider. The spread should be documented and consistently applied across your job architecture.
3. Geographic adjustment — if the role is in a high- or low-cost metro relative to the national median, the range should reflect local market data (BLS publishes state and metro-area estimates) or a documented geographic differential.
4. Record — a file that captures all of the above, dated, and stored for the retention period your jurisdiction requires.

If you are building ranges across ten or twenty roles, the manual version of that process is sustainable but slow. If you are building across fifty roles, or if you operate in multiple states with different compliance thresholds, the manual version becomes a material operational risk — not because the data is hard to find, but because the documentation discipline is hard to maintain at scale.

That is the gap Salary Range Builder is built for: the same BLS OEWS data your spreadsheet pulls, assembled into a structured, sourced, state-formatted PDF with the data vintage watermarked and the methodology documented — in the time it would take to run the lookups manually.

If you want to start with a structured template before committing to software, our Salary Range Builder Workbook is an Excel-based tool that walks you through the BLS lookup, the spread methodology, and the documentation fields — and produces an output you can actually hand to counsel.

---

The Compliance Line AI Cannot Cross

The question "can ChatGPT build a salary range?" has a narrow technical answer: yes, it can produce a number formatted like a salary range. The compliance answer is different: no, it cannot produce a range that is grounded in a named authoritative source, anchored to a documented percentile, adjusted to a specific geography, and reproducible for an enforcement review.

Pay-transparency laws are not asking you to post a number. They are asking you to post a defensible number — one that reflects what the market actually pays and that you can explain if asked. The standard is documentation, not speed.

AI is a useful tool for many tasks. Generating compensation data that can be cited in enforcement documentation is not one of them.

If you are ready to build ranges that will hold up, start a 14-day free trial of Salary Range Builder — no free tier, no enterprise sales call required, and same-day setup so the range due Friday actually gets done.

For a broader look at how the tool category compares, see our best compensation software for small business guide.