- The published ranges disagree because each source measures a different population. Government data, self reported aggregators and job posting scrapes are not three estimates of one number.
- The most cited government figure is a median of 140,910 dollars per year as of May 2024 for computer and information research scientists, an occupation of only 40,300 jobs.
- Most people with the title work in an occupation of 1,895,500 jobs with a 2024 median of 131,450 dollars, which is the number the headlines skip.
- Self reported data reads far higher: an average base of 184,757 dollars with a range from 80,000 to 338,000 and an extra 26,486 dollars in cash on top.
- Median and average are not the same statistic. A right skewed pay distribution puts the average above the median by construction, before anybody exaggerates.
- Location and level move pay more than the AI label does. City spread is about 1.5 times and experience spread about 1.8 times within the same dataset.
A backend engineer looking at AI roles sees ranges from 80,000 to 338,000 dollars quoted with equal confidence, and reasonably concludes that somebody is lying. Nobody is. The numbers describe different groups of people, collected in different ways, and the disagreement is the most informative thing about them.
This piece explains what each source is actually counting, then gives a defensible number and the reasoning behind it. If you only read one section, read the one on what the title covers, because four different jobs are wearing it.
Why do published ranges disagree so violently?
Four reasons, and they compound.
Different statistics. Government publications report a median: the middle person. Aggregators usually report a mean. In a right skewed distribution, which every technology pay distribution is, the mean sits above the median automatically. That gap is arithmetic, not spin.
Different populations. A survey of people who chose to enter their salary into a website is not a sample of the workforce. It skews toward large employers, coastal cities and people who feel good about their number.
Different definitions of pay. Base salary, base plus bonus, and total compensation including equity are three numbers that can differ by a third or more for the same person.
Different occupation codes. This is the big one, and it is the subject of the next section.
What does each source actually measure?
| Source type | Figure it reports | Population it describes | Systematic bias |
|---|---|---|---|
| Government occupational data | 140,910 dollars median, May 2024, research scientists | Everyone in a defined occupation code, from employer payroll records | Lags the market, and the code may not match the job title |
| Government occupational data, software | 131,450 dollars median, 2024, software developers | 1,895,500 people, the code most AI engineers actually sit in | Mixes AI work with all other software work |
| Self reported aggregator | 184,757 dollars average base, plus 26,486 dollars cash | People who chose to submit a salary to a jobs site | Volunteer sample, skewed to larger employers and high cost cities |
| Employer review site by experience | 103,015 dollars at 0 to 1 years, 185,709 dollars at 15 or more | Self reported base pay bucketed by tenure | Tenure is a weak proxy for level and reported inconsistently |
| Job posting data by city | 206,706 dollars in San Jose down to 136,347 dollars in Seattle | Advertised roles, not accepted offers | Posted ranges are negotiating positions, and vacancies skew to growth areas |
The government figures come from the Occupational Outlook Handbook entry for computer and information research scientists, which reports median pay of 140,910 dollars per year as of May 2024, a typical entry requirement of a master's degree, 40,300 jobs in 2024, projected growth of 20 percent from 2024 to 2034 and about 3,200 openings a year.
The self reported figures come from Built In's AI engineer salary page, which reports an average base of 184,757 dollars, a range from 80,000 to 338,000, and average additional cash compensation of 26,486 dollars, based on responses gathered from anonymous employees in the US.
The experience and city breakdowns come from Coursera's guide, which aggregates Glassdoor and Indeed figures accessed on 5 February 2025, alongside a mean across all US occupations of 65,470 dollars for context.
Two sources cite the same agency and disagree. What now?
It happens, and it is instructive. Coursera attributes an annual median of 145,080 dollars to the Bureau of Labor Statistics as accessed on 5 February 2025. The Occupational Outlook Handbook page for the research scientist occupation states 140,910 dollars as of May 2024. Both point at the same agency.
The lesson is not that one is wrong. It is that a citation names an organisation, not a series, a date or an occupation code. When a figure matters to a decision, follow it to the page that publishes it and note which occupation and which reference period you are looking at.
Which occupation is an AI engineer, really?
Almost never the one the headlines quote. This is the single most useful thing in this article.
The research scientist occupation has 40,300 jobs and about 3,200 openings a year. The software developer occupation has 1,895,500 jobs, a 2024 median of 131,450 dollars, 15 percent projected growth and about 129,200 openings a year. That is roughly 47 times the headcount and 40 times the annual openings.
If you are a backend engineer moving into AI work, the job you will get is overwhelmingly in the second group. You will build systems that call models, evaluate outputs, manage data pipelines and ship product. The research scientist code describes people doing research, typically with a graduate degree, in a labour market a fortieth of the size.
This is why the 20 percent growth figure gets quoted so often and helps so little. Twenty percent of 40,300 is roughly 8,000 jobs added over a decade. Fifteen percent of 1,895,500 is a far larger absolute number. Percentage growth on a small base is a weak signal about where the work is.
What are the four jobs sharing this title?
Pay differs between them more than it differs between companies of the same size, so identifying which one you are applying for matters more than memorising a range.
| Role behind the title | What the day looks like | Where pay sits | What actually moves it |
|---|---|---|---|
| LLM application engineer | Prompts, retrieval, evaluation harnesses, product integration | Near the software developer median, higher at large employers | Shipping something that survives real traffic |
| Applied machine learning engineer | Training and fine tuning models against a business metric | Above the software median, below research | Owning a metric end to end rather than a model |
| Research engineer | Experiments at scale, infrastructure for training runs | The top of the published ranges | Publication record, distributed training experience |
| Data or platform engineer supporting ML | Pipelines, feature stores, serving infrastructure, cost control | Software developer band, often undervalued | Reliability and unit economics, both easy to evidence |
The fourth row deserves a note. Platform work supporting machine learning is consistently paid as ordinary infrastructure work while being the thing that decides whether any of the model work reaches production. If you already do backend and data engineering, this is the shortest path into an AI team and the one with the least competition.
How much does the AI label add?
Less than the location and the level, and this is measurable inside the sources themselves.
Take the city figures from the same dataset. San Jose at 206,706 dollars against Seattle at 136,347 dollars is a spread of about 1.5 times for the same job title. Take the experience figures from the other dataset: 103,015 dollars at 0 to 1 years against 185,709 dollars at 15 or more is about 1.8 times.
Now compare the AI premium. The self reported AI engineer average base of 184,757 dollars against the software developer median of 131,450 dollars looks like 1.4 times, except those two numbers are not comparable: one is a mean from volunteers, the other a median from payroll data. Correct for that and the label premium shrinks toward the noise.
The blunt version: moving from Seattle to San Jose does more for your number than adding AI to your title, and moving from mid level to senior does more than both. That is not a statement about the value of the skills. It is a statement about how compensation bands are actually built, which is by company tier, geography and level, with the specialism adjusting inside the band.
Does that mean learning this is not worth it?
No. It means the return arrives as access rather than as a premium. AI capability changes which companies will interview you and which projects you get, and those are what move you up a level. The level is what pays.
How should you read a range on a job posting?
As a negotiating position with legal constraints attached, not as a description of what people there earn.
Pay transparency rules in several states require a posted range, and employers responded rationally by posting wide ones. A band running from the low six figures to well over two hundred thousand is usually covering several internal levels at once, which means the range tells you the ladder exists and almost nothing about where you would land on it.
Three questions convert a posted range into information. Which levels does this band cover? Where in the band do people typically start at my level? Is the top of the band reachable without a promotion? Recruiters answer all three more often than candidates expect, because refusing to answer costs them a pipeline.
There is a matching asymmetry in the aggregate data. Job posting averages reflect vacancies, and vacancies cluster where hiring is hardest and pay is highest. That is why city rankings built from postings tend to sit above the same city's realised pay: you are looking at the roles that were hard to fill.
What about remote roles?
Remote pay is set by policy rather than by geography, and the policies differ enough to swamp the AI premium. Some employers pay a single national rate, some band by the employee's location, and some band by the nearest office. The same person doing the same job can see a difference of tens of thousands of dollars depending on which of those three a company chose, and none of it is visible from the job title.
What does the growth data say about the next few years?
That the work is expanding and the label is not where the volume is. Both government occupations are projected to grow much faster than the average for all occupations: 20 percent for research scientists and 15 percent for software developers, each over 2024 to 2034.
Put the openings side by side and the picture is clear. About 3,200 research scientist openings a year against about 129,200 software developer openings a year. Even if only a modest share of the second group involves model based systems, that share is a far larger absolute number of jobs than the entire first category produces.
For a candidate, the practical reading is that you should be searching by the work rather than by the title. Roles described as backend, platform or data engineering that mention retrieval, evaluation or inference cost are the same job as many postings labelled AI engineer, and they attract fewer applicants because the keyword is missing.
So what is a defensible number?
For a competent backend or data engineer moving into applied AI work in the United States, without a graduate research background: expect an offer in the range of the software developer median to roughly 40 percent above it, with location and company tier explaining most of where you land inside that band.
Concretely, that anchors on 131,450 dollars as the middle of the broad occupation and reaches toward the self reported averages in high cost metros at larger employers. Numbers near the top of published ranges, above 300,000 dollars, exist and belong to research roles at a small number of employers, usually with equity doing most of the work.
Two habits make that estimate usable. Always convert a quoted figure to total compensation before comparing it to anything, since base plus bonus plus equity is the only comparable unit. And always ask which occupation a statistic covers before treating it as a benchmark for your role.
One caution about that band. It describes a market at a moment, and the sources behind it carry reference dates: May 2024 for the government medians, and February 2025 for the aggregated site figures. Pay data always describes the past, and in a market moving this quickly a figure eighteen months old is a floor rather than a forecast.
What should you actually verify before accepting an offer?
- The level, in the company's own ladder. Title inflation is free. The level determines the band and the next promotion.
- The equity terms. Vesting schedule, refresh policy and, for private companies, the last valuation date. An equity number without those is a decoration.
- The location band. Many employers pay by location even for remote roles. Ask which band applies to you specifically.
- What the role actually is. Use the four row table above. If the job is platform work labelled AI, expect platform pay and negotiate on scope instead.
Is the demand real, or is it a bubble in job titles?
Both things are true at once, which is why the debate never resolves. Demand for people who can make model based systems work in production is real and visible in openings. Demand for the title as a status marker is also real, and it is what produces roles where the AI content is a wrapper around ordinary software work.
The useful evidence is what the systems can and cannot do yet, since that determines whether the hiring persists. We looked at the measured version of that question in how far AI agents get on real remote work and in what the research says about AI agent coworkers. The short version: capable on bounded tasks, unreliable on long ones, which is precisely the profile that creates engineering jobs rather than eliminating them. The engineering is the reliability.
There is a mirror image of this on the buying side worth naming. If you are a founder rather than a candidate, the same arithmetic runs backwards: the cost of getting software built has fallen faster than the cost of hiring someone to build it, which is why comparing an annual salary against what a build with AI tooling actually costs per month is now a real decision rather than a rhetorical one. It does not replace the hire. It changes when you need to make it.
One last suggestion for anyone weighing the move. Do not optimise for the title. Optimise for a project where you own an outcome that can be measured, because that is what converts into a level, and the level is what the money follows. The interesting version of this is not new to AI either, as the Ford story about bringing back 350 experienced engineers illustrates: an organisation pays for judgement it can point at, and judgement is demonstrated on work that shipped.