Which AI Employers Publish a Salary: 52% Do, and the ATS Predicts It

Roughly half of the AI job market will not tell you what it pays. On 7 September 2026 our salary audit covered 12,907 active postings at the companies we track, of which 6,757 carried a pay band we could parse: 52.4%. That universe is every active posting at those companies, non-AI roles included. Narrowed to the roles we classify as AI, the shape is the same: 2,852 live AI rolesdisclosing at 49%. Two universes, and every figure below says which one it belongs to.

This is the first monthly edition of the disclosure index, frozen at publication, and it carries a correction to every disclosure percentage this site published before 2 September 2026. Live counts are at /api/public/stats.

The applicant tracking system predicts disclosure, not the company

The strongest predictor of whether an AI employer publishes a band is not its size, its funding stage or its stated values. It is the software its recruiters open in the morning. Across all 12,907 active postings at tracked companies, non-AI included:

• Ashby - 4,138 disclosed of 7,382 active (56%)
• Greenhouse - 2,566 of 4,846 (53%)
• Lever - 52 of 280 (19%)
• Workday - 1 of 343
• Comeet - 0 of 56

Ashby and Greenhouse put a structured pay-range field in front of the recruiter while the requisition is being created, so the path of least resistance is to fill it in. On Lever, Workday and Comeet the number lives in prose, if at all, and the path of least resistance is silence. One posting in 343 from the Workday tenants we read carried a parsable band, and those tenants are the large enterprises: the part of the market with the most formal pay structure publishes least.

The bands that do get published are enormous

Verbatim from listings live that day, in each employer's own currency:

• Anthropic, Research Engineer, Model Evaluations - $500,000 - $850,000
• Anthropic, Research Scientist, Interpretability - $350,000 - $850,000
• OpenAI, Researcher, Alignment Interpretability - $295K - $500K
• Roblox, Senior Engineering Manager, AI Safety Platform - $345,040 - $399,420
• Shield AI, Senior Staff Engineer, ML Ops (remote) - $280,000 - $420,000
• Cohere, Senior Research Scientist, Model Evaluation (Toronto) - CA$250K - CA$535K
• Retell AI, Forward Deployed Engineer, New Grad - $150K - $260K
• Waymo, Senior Safety Researcher (London) - £113,000 - £119,000
• Cohere Health, Sr. Forward Deployed Automation Strategist (remote) - $100,000 - $115,000
• Modal, ML Research Intern - $15K per month

The Anthropic evaluations band spans $350,000 from bottom to top. That is not a price, it is a range wide enough to hold several levels, and where you land inside it is settled by a levelling conversation the posting does not describe. At the other extreme Waymo publishes a London band £6,000 wide and Roblox publishes to the dollar, which is what a posting looks like when it comes out of a compensation grid. A wide band and a narrow band are not the same disclosure, though both count as one with us. More on these functions in our AI safety, model evaluation and AI research hubs.

Note what the disclosed pool is made of. Eight of the ten bands above are US employers, and US state pay-transparency law is why many of these numbers exist at all. Published bands skew senior, US and well funded: an upper view, never a median. The same caveat applies to the bands in our remote and entry-level editions.

A correction: our own disclosure numbers were wrong until 2 September

This is the part of the report that matters most, because we got it wrong first.

On 2 September 2026 we found a defect in our own salary parser. Greenhouse stores each job description HTML-entity-escaped, so the tags inside its pay-range widget reached us as literal text: a reader sees a clean band, our parser saw <span> sitting between the two numbers. Our extractor stripped real tags only, so the range pattern never matched and every posting using that widget was recorded as "salary not disclosed". We fixed the decoder and re-ran the board.

Board-wide disclosure went from 39.7% to 52.8%. Greenhouse disclosure went from 878 to 2,531 of 4,784 active postings.

That jump is a measurement fix, not employers becoming more transparent. Nothing changed in the market between one day and the next. Thousands of postings had been publishing a band all along and we were reporting them as silent. Any disclosure percentage published by this site before 2 September 2026 is superseded by the numbers in this edition. We would rather say that plainly than quietly re-baseline and let a parser bug look like a trend. If you are tracking AI pay transparency over time, our series starts here.

What our parser still misses

The fix did not close the gap completely, and the residual belongs in the open. In the same 7 September audit, 134 active postings still carried a Greenhouse pay-range widget with no band we could parse, and 206 mentioned "base salary range" without a range we could read. The largest holders were Lightmatter 59, Pinterest 25, Datadog 17, Block 12 and SambaNova 9.

These are limits of our heuristics, not the escaping defect. The extractor looks for an anchor phrase and then for numbers within a short window of it, and the biggest residual holder puts other text between the anchor and the numbers, so the window closes first. Each of those postings is a real disclosed band counted as undisclosed, which makes 52.4% a floor, not a ceiling.

One more thing the postings are not saying

Pay is not the only field that goes quiet. In the same audit, 5,077 of the 12,907 active postings had been open more than 90 days. That is not proof of jobs that do not exist, but a posting open a quarter with no price on it asks for an afternoon of work while disclosing neither the money nor the urgency. Both are on the market dashboard.

What to do with this

• Read the apply URL before the posting. jobs.ashbyhq.com and greenhouse.io mean a band is more likely than not. jobs.lever.co, myworkdayjobs.com and Comeet mean you should assume no number is coming and ask on the first call.
• Treat a wide band as a level range. A $350,000 spread tells you the ladder, not the offer. Ask which level the requisition is approved at.
• Do not benchmark against the disclosed pool. It skews senior, US and well funded, and the half of the market missing from it is not a random sample.
• Ask for the band at the screening call. Half of employers already publish one, which makes it an ordinary request.
• Employers: fill in the field. The gap between Ashby and Workday is mostly a difference in defaults, not in policy. Company lists are at companies hiring in AI.

Method and caveats

Two universes appear above and are labelled every time. The first is our salary audit read on 7 September 2026: all 12,907 active postings at the companies we track, non-AI roles included, 6,757 of them with a parsed band (52.4%). The ATS breakdown, the 134 and 206 residuals and the 90-day figure belong to it. The second is /api/public/stats the same day: 2,852 live AI roles, 49% disclosing and 13% remote. The two are not interchangeable and should never be quoted as one series.

"Disclosed" means the extractor found a parsable range in the structured pay field or the description, so pay described only in prose we cannot parse counts as undisclosed. Postings are read hourly from each company's own applicant tracking system (Greenhouse, Lever, Ashby, Workday and Comeet); enterprise Workday feeds are capped per tenant, which under-counts large multinationals and is part of why the Workday row is so small. It is a large sample of companies that publish an ATS feed, not a census, so treat every count as a floor. Pay bands are quoted exactly as published, in the employer's own currency. Every figure here is free to quote with attribution to AI Hiring Board.