Published AI salary figures disagree by factors of two or three, and almost all of the disagreement comes from three things: who filled in the survey, what counts as compensation, and what the title means. This page will not give you a number. It will let you read anyone else’s and tell whether it is worth anything.
Why the figures disagree so wildly
A salary figure has to answer four questions before it means anything: which population, which geography, which components of pay, and which seniority. Most published figures answer none of them explicitly, and the spread that results is not measurement noise — it is four different quantities being reported under one label.
“AI engineer” makes this worse than the equivalent problem for, say, plumbers. It is not a regulated occupation, there is no licensing body, and it has no code in any official occupational classification. In the US Standard Occupational Classification, people doing this work are scattered across software developers, data scientists and computer and information research scientists. So there is no official statistic about the job as such, only official statistics about buckets that contain it along with a great deal else.
The sources, ranked by what they can support
| Source | Description |
|---|---|
| BLS OEWS (US) | The Bureau of Labor Statistics Occupational Employment and Wage Statistics survey. A probability sample of employers, published annually with percentiles by occupation and metro area. Authoritative, but by occupation code, so it cannot isolate AI roles and it lags by about a year. |
| ONS ASHE (UK) | The Annual Survey of Hours and Earnings, drawn from PAYE records rather than self-report. Same strength and same limitation as OEWS: real sampling, occupational codes too coarse for this job title. |
| DOL LCA disclosures (US) | Employer-filed wage data for visa applications, published quarterly as downloadable files. Not a survey at all. Covered in detail below. |
| Stack Overflow Developer Survey | Very large, self-selected, publishes its methodology and its raw anonymised responses. Excellent for relative comparisons within the sample, unsafe as a population estimate. |
| levels.fyi | Self-reported and verification-encouraged, with company-specific levelling, which is its real contribution. Heavily skewed towards large US technology firms, so it describes the top of the distribution well and the rest badly. |
| Glassdoor / Indeed / job-board aggregates | Mixtures of self-report and posted ranges. Posted ranges are increasingly mandated by pay-transparency laws in several US states and in the EU, which makes them more useful than they were, but a posted band is an intention, not a payment. |
| Recruiter and consultancy salary guides | Published by firms that profit from salary movement, usually with no stated sample. Treat as marketing unless a method section exists. |
The source almost nobody uses
The US Department of Labor’s Office of Foreign Labor Certification publishes disclosure files for Labor Condition Applications — the filings an employer must make to hire on an H-1B visa. They are released as downloadable spreadsheets, quarterly, free.
For this particular question they are unusually good, because of what they are rather than what they were intended for:
- The employer files them, not the employee. No self-report bias, no survey response bias, no rounding up in a bar.
- They are legally binding. The wage stated is the wage the employer is committing to pay, and misstating it has consequences.
- They carry the job title, the worksite city and the wage, so you can filter by literal title text and by location rather than by an occupational code.
- They are a census of that population, not a sample. Every application appears.
The limitations are equally clear and must be stated whenever you quote from them. The population is only workers on these visas, which skews towards large employers and towards certain roles and origins. The wage field is base salary and does not include equity or bonus, so it understates total compensation at firms where equity dominates. And it is US-only.
Filing volumes and wage levels move every quarter, and the file format and column names have changed between releases. Download the current quarter’s file from the OFLC performance data pages rather than relying on any third-party mirror, and check the column layout before parsing.
The equity arithmetic that breaks comparisons
At the firms that pay the most, base salary is a minority of compensation, and there are at least three defensible ways to value the rest. Two surveys can report honestly and differ by 40 per cent purely on this choice.
Total compensation, annualised:
TC = base + target_bonus + (equity_grant / vest_years)
Worked on one offer:
base = $200,000
target bonus = 15% of base = $30,000
equity grant = $600,000 over 4 years = $150,000 per year
TC at grant-date value = 200,000 + 30,000 + 150,000 = $380,000
Now value the same equity at a share price that has doubled since grant:
equity value = $1,200,000 over 4 years = $300,000 per year
TC at current value = 200,000 + 30,000 + 300,000 = $530,000
Same person, same offer letter, same year.
$380,000 and $530,000 are both correct. They differ by 39%.
A third convention reports only what actually vested and was sold in
the year, which for somebody in year one of a four-year grant with a
one-year cliff is:
TC realised = 200,000 + 30,000 + 0 = $230,000
Three conventions, one offer, a range from $230,000 to $530,000. Before comparing any two compensation figures, establish which convention each used. If a source does not say, it is not comparable to one that does.
Levelling, and the title problem
The second-largest source of spread is seniority, and job titles do not encode it consistently. “Senior” means five years at one company and fifteen at another; a “staff” engineer at a large firm may out-earn a “director” at a small one.
- Match on level, not on title. The only reliable cross-company mapping is via published levelling ladders, which is what levels.fyi exists to provide and why its data is more useful than its sampling deserves.
- Watch the title drift. The set of titles under which this work is advertised has changed repeatedly — machine learning engineer, data scientist, prompt engineer, AI engineer, research engineer — and a time series built on one title measures fashion as much as pay. See how the roles actually differ and what happened to the prompt engineer job.
- Geography dominates almost everything else. The same level at the same company in two countries can differ by more than the gap between two adjacent levels in one of them. Never quote a global figure.
- Survivorship in the sample. People who volunteer a salary to a public database are disproportionately pleased with it. The direction of that bias is knowable and its magnitude is not.
Getting a defensible number for yourself
- Define the population precisely before looking at any data: country, metro area, level, company size band, and whether the number is base or total.
- Take base salary from a source with real sampling — OEWS or ASHE for the occupational floor, or the LCA disclosure files filtered on literal job titles for something much closer to the actual role.
- Take the equity and bonus structure from levels.fyi or from posted ranges under pay-transparency laws, and state the valuation convention you applied.
- Report a median and an interquartile range. Compensation distributions are strongly right-skewed and a mean is dragged upward by a handful of outliers.
- Attach the date. Everything in this area moved sharply in both directions over the last three years, and a two-year-old figure is a historical note rather than a benchmark.
- Sanity-check against demand signals rather than against other salary articles. Indeed’s Hiring Lab and Lightcast both publish job posting counts and shares by skill, which move before pay does.
Top comments (0)