Build the systems that learn from data, and keep them running once real people use them.
Click through it
Pay, the path, pros and cons and the facts on one printable page. Counsellors: send it to a student from your own email in one click.
The median pay is about $140,300 a year. Across everyone in the job it runs from about $82,200 at the lower end to about $230,630 at the upper end. The source is listed under Sources at the end of this page.
4 to 6 years after high school, on the usual route.
$1,500 to $90,000 all in for tuition, fees, exams and kit. Living costs are not counted.
No. There is no license to hold, though many employers still ask for training or a certificate.
Very high demand nationally, from government outlook data. Demand varies by region, so check your area.
Or read the whole thing here
What you earn and what it costs to get there.
Median
$140,300
Bottom tenth
$82,200
Top tenth
$230,630
The federal wage survey has no code for this title, so these are computer and information research scientists: median $140,300, bottom tenth $82,200, top tenth $230,630. That code leans toward research roles with advanced degrees. A lot of working machine learning engineers are actually counted as software developers instead, where the 2025 median is $135,980. Neither figure is a typical first salary.
$82,200 to $103,570
The bottom tenth and bottom quarter of computer and information research scientists, 2025
This is the bottom quarter of a group that leans toward research roles. Pay also varies by city, and so does the cost of living.
$103,570 to $188,700
Half of all computer and information research scientists earn between these two figures, 2025
Published salary figures usually miss stock. At big technology firms a meaningful part of total pay is company shares that vest over several years, which is also what keeps people from leaving.
Roughly, all in
$1,500 to $90,000
Tuition, fees, exams and kit. Not rent, food or travel
The low end is the self taught route, course material at no cost, open datasets and a few hundred dollars of cloud computing time, and people really do get hired that way. Be honest about it though, because that door almost always opens for someone already working as a software developer who is moving across, not for someone starting from nothing. The high end is a four year public university degree at in state rates plus a master's, which is the common path. Two years at a community college then transferring cuts the degree cost sharply, and paid internships can cover part of a year's tuition.
File the FAFSA (Free Application for Federal Student Aid). Pell Grants are not repaid, and federal loans for a dependent undergraduate cap at $5,500 the first year and $31,000 across the degree. Paid internships in this field can cover part of a year's tuition.
Remote and hybrid work
Common in this field.
Conference and training budgets
Employers pay for courses and conferences because the field changes fast.
Bonuses
Common at larger technology companies, on top of salary.
Skills that transfer across industries
Hospitals, banks, farms, insurers and governments all hire for this now, not just technology companies.
The exact path from high school to qualified.
A degree plus proof you can actually build things. Both halves are required.
High school
Grades 9 to 12
Take the hardest math available and start writing code, ideally Python. Everything after this assumes you already enjoy both.
Bachelor's degree
4 years
Computer science, mathematics, statistics or a related engineering field. Community college for two years then transferring to a public university is the cheapest version of the same degree.
Build and ship things
Throughout
Personal projects, open source contributions, competition entries. A working project someone else uses beats a course certificate every time.
Internships
Summers
Apply from your second year. This is how most people get their first full-time offer, and many companies hire almost exclusively from their own interns.
First job, then decide on a master's
0 to 2 more years
Engineering roles rarely need one. Research roles usually want a master's or a doctorate. Working first tells you which you actually want.
Calculus
Model training is built on derivatives. You will meet them again on day one.
Statistics and probability
Arguably more important than calculus, and far more often skipped.
Computer science
Any programming at all. Python is the language this field runs on.
Physics
Good practice at turning a messy real situation into a model you can actually compute with.
There is no board, no exam and no legal title. Anyone can call themselves a machine learning engineer, and hiring is done almost entirely by technical interview and portfolio. That cuts both ways: it means there is no gatekeeper keeping you out, and it means the interview is genuinely hard. Rules are starting to arrive on the use side rather than the practitioner side. The National Institute of Standards and Technology publishes a voluntary framework for managing risk in artificial intelligence systems, and some states and sectors regulate specific uses like hiring tools or medical devices. None of it licenses you personally.
Large public universities
In-state tuition at a state flagship is the cheapest strong route into this field, and employers hire from them.
Specialist departments
A handful of universities have dedicated machine learning departments, for example Carnegie Mellon University.
Online master's programs
Georgia Tech's online computer science master's is the best known and costs a fraction of an on-campus degree.
Self-taught plus a portfolio
It happens, mostly for people already working as software engineers. Breaking in this way with no degree and no job history is rare.
Things you don't need, but that help.
Large language models
Fine-tuning, evaluation and building products on top of them. The busiest corner of the field right now.
Machine learning infrastructure
Serving models reliably at scale. Less glamorous, and essential.
Computer vision
Medical imaging, manufacturing inspection, self-driving. Deep and durable.
Applied research
The route that usually wants a doctorate.
Solid software engineering
Testing, version control, code review. Models that cannot be deployed are worth nothing, and this is the most common gap in new graduates.
Cloud experience
The major cloud platforms all have no-cost tiers and cheap certifications. Employers do read them.
Writing clearly
You spend a lot of time explaining to non-technical people why the model did something strange.
A public portfolio
Code someone can read and a project someone can try, on a public repository.
At most companies this role is software engineering with a modeling component: you take a model someone else trained, adapt it, wire it into a product and keep it working. At a small number of research labs it means inventing new methods and publishing papers, and those jobs want a doctorate. Both are called machine learning engineer. Read the actual job description rather than the title, because the day-to-day work and the qualifications expected are very different.
Day to day, pros and cons, where you'd work.
Start with whether last night's training run finished and whether the numbers moved. Most of the morning is data: finding it, cleaning it, working out why a chunk of it is wrong. Then a code review, a meeting with the product team about what the model is supposed to do, and an argument about how to measure that. Afternoon might be writing an evaluation, or chasing why the deployed model got slower this week. Actual model architecture work is a small slice of the whole thing.
The good
The hard parts
Things people don't tell you.
The first artificial neuron model was published in 1943, and Frank Rosenblatt built the perceptron, a machine that learned from examples, in the late 1950s. The theory sat mostly unused for decades because the hardware and the data to make it work did not exist yet.
Training a large model costs millions of dollars in computing time, so the overwhelming majority of engineers build on top of models trained by someone else. The valuable skill is not building one from nothing, it is knowing how to adapt and evaluate one.
Every list is built from the same numbers shown above, so you can see where this career sits next to the others.
Data and AI first, then the rest of the field. See every career in tech.
Last checked October 2026. Pay figures are typical full-time annual amounts in United States dollars, based on Bureau of Labor Statistics wage data and published pay scales. They vary by state, employer and experience. Tuition is for in-state students at public schools unless the card says otherwise. How every figure on this site is sourced and checked.
Tell us about this page. A person reads every message.
No email app on this device? Use Gmail instead: report a mistake in Gmail or say what you like in Gmail.