TL;DR — Our index held 708 live remote AI and machine-learning engineering roles on 20 September 2026. Three titles compete for the same searches and mean different work: a data scientist answers questions with models, an ML engineer keeps models running in production, and an "AI engineer" builds product features on models somebody else trained. Read the responsibilities, not the title.
The numbers first
Counted across live remote postings in our index on 20 September 2026, from companies' own careers pages.
| What we counted (20 September 2026) | Number |
| Live remote AI / ML engineering roles | 708 |
| Of those, roles that publish a salary range | 16.5% |
| Roles that name the United States among their hiring regions | 23.0% |
| Live remote postings in the index overall | 37,811 |
| Postings that publish a salary range (all roles) | 15.6% |
| Postings restricted to the United States only | 18.7% (7,089) |
Two things stand out. Pay transparency in this family is a shade better than the market average but still means that five roles in six tell you nothing about money. And 23.0% naming the US is lower than you might expect for a field whose centre of gravity is San Francisco — a good share of these roles are open to people outside the US, or say nothing about geography at all.
Recognisable employers hiring remotely in this family include Reddit. Most of the rest are companies you will not have heard of until you read the posting, which is normal: model work is now inside logistics firms, health insurers and B2B SaaS, not only at AI labs.
Three titles, three jobs
The titles are used loosely, but there is a real division of labour behind them, and it has sharpened since generative models arrived.
| Title | What the work is | A typical week |
| Data scientist | Answering business questions with data and models. Framing the problem, choosing features, validating that a result is real. | Exploration, a notebook, an experiment design, a readout to stakeholders |
| Machine learning engineer | Making models run reliably in production: serving, retraining pipelines, monitoring, drift, cost and latency. | A feature pipeline, a deployment, an alert that fired at 3am, a regression in a metric |
| AI engineer | Building product features on top of pretrained models: retrieval, prompts, agents, tool calls, guardrails, evaluation. | A retrieval pipeline, an eval set, a prompt regression, a latency budget |
The honest summary: a data scientist decides what the model should do, an ML engineer makes it keep doing it, and an AI engineer wires a model into a product. Most companies need two of the three and write one job posting.
How to tell which one a posting means
Ignore the title in the header. The responsibilities list tells you within four bullets.
- "Design experiments", "A/B test", "statistical significance", "stakeholders", "dashboard" → data science. The output is a decision, and you will spend real time explaining results to people who do not read code. Our guide to remote data and analytics jobs covers this branch.
- "Training pipeline", "feature store", "model serving", "retraining", "drift", "latency", "Kubernetes", "Airflow" → ML engineering. This is a software job with statistics in it. If you are coming from backend work, our remote software engineer guide maps the overlap.
- "RAG", "embeddings", "vector database", "agents", "LangChain", "prompt", "evals", "hallucination" → AI engineering. Almost no model training; a great deal of retrieval, evaluation and product judgement.
- "Publications", "novel architectures", "SOTA", "PhD preferred" → research. A different hiring bar and a much smaller number of remote openings.
A posting that mixes all four is a company that has not decided what it needs. That is not automatically a bad job, but ask in the first call which of the four the first six months are actually about.
What the market is doing
When we checked in September 2026, LinkedIn's Jobs on the Rise report for 2026 ranked AI engineer as the fastest-growing job title in the United States, and listed LangChain, retrieval-augmented generation and PyTorch among the skills most common on the profiles of people in the role. That combination is the whole story in one line: the fastest-growing title in the field is mostly about using models, with classic training work sitting beside it rather than under it.
For your search that means two things. Searching only "machine learning engineer" misses a large and growing share of the openings. And a resume built entirely around model training reads as over-qualified and under-equipped for half of what is being hired.
What interviews ask in 2026
The loops have shifted. Reading across the 2026 interview guides published by training platforms and interview-prep sites when we checked in September 2026, the recurring themes were consistent enough to plan around.
| Round | What it actually probes |
| Coding | Ordinary software engineering — data structures, a small data-manipulation task, clean code. Not Kaggle tricks. |
| ML fundamentals | Metrics and why you chose one, class imbalance, data leakage, train/serve skew, overfitting. Still asked, still failed. |
| LLM fundamentals | Tokenisation, embeddings, context windows, what fine-tuning changes and what it does not, LoRA/QLoRA and when parameter-efficient tuning beats prompting. |
| RAG and system design | Design a retrieval pipeline end to end: chunking, index choice, re-ranking, what you cache, what you do when retrieval returns nothing useful. |
| Evaluation | The round that separates candidates. How do you evaluate a non-deterministic system? Retrieval and generation measured separately, labelled eval sets, rubric or model-as-judge scoring, regression tracking across prompt and model changes. |
| Production failure | "Your agent fails 30% of the time in production — what do you do?" Observability, retrieval drift, runaway tool calls, prompt injection, rollback. |
| Cost and latency | Tokens per request, caching, model routing, the trade you made and why. |
Two pieces of advice that follow directly from that table.
Bring an evaluation story. Not "we used RAG", but: what you measured, what the baseline was, what regressed when you changed the prompt, and how you caught it. Interviewers in this field are tired of demos and short of evidence.
Bring a number. Latency, cost per request, retrieval precision, tokens saved. One concrete number does more for you than a list of frameworks, and it is the thing an interviewer repeats to the hiring manager afterwards. Our guides to remote interview questions and to what a salary range in a job posting really tells you cover the rest of the loop.
What to put on the resume
The failure mode we see most is a resume that lists models and libraries and never says what shipped. Three fixes, in order of effect:
- Name the production system, not the notebook. "Served a ranking model to 2m requests a day" beats "experience with XGBoost".
- Say what you measured. Every bullet with a metric in it survives the screen; every bullet with a framework in it competes with a thousand identical bullets.
- Match the vocabulary of the branch you want. The same three years of work can be written as data science, ML engineering or AI engineering. Pick the one the posting is written in — not to game the filter, but because the reader is scanning for evidence of their job, and generic wording gives them none.
Where these roles are, and how to search
- Search all three titles. "Machine learning engineer", "AI engineer", "applied scientist", "ML platform engineer" and "MLOps engineer" surface overlapping but different sets.
- Check the geography line before you invest. 23.0% of these roles name the US among their hiring regions, and across the whole index 18.7% of remote postings are US-only. If you are outside the US, that is the filter that saves the most time.
- Expect seniority to skew up. Across our index, senior-labelled postings outnumber junior ones roughly five to one (11,039 against 2,018). In model work the gap is wider still, and the realistic entry path is usually a software or data role at a company that is building AI features.
- Browse the role. Our remote machine learning engineer page lists live openings from employers' own careers pages, and the machine learning engineer salary page collects the ranges the postings that publish them actually state.
A note on pay
With only 16.5% of these postings publishing a range, most of your negotiating will be done without a public anchor. The usual asymmetry applies: companies know their band, you do not. Ask two questions on the first call — what the band is for the level, and whether it is adjusted for location — and you will be better informed than most candidates in the loop.
See which AI and ML roles actually fit your background, scored against your own resume. Try it free.
FAQ
What is the difference between an ML engineer and a data scientist?
A data scientist frames a question, builds a model to answer it, validates the result and explains it to the business. An ML engineer takes models into production and keeps them there: serving, retraining, monitoring, drift, latency and cost. The first role's output is a decision; the second role's output is a running system.
What is an "AI engineer", and is it a real job title?
It is real, and in 2026 it usually means building product features on top of pretrained models rather than training models. The work is retrieval, prompts, agents, guardrails and evaluation. When we checked in September 2026, LinkedIn's Jobs on the Rise report for 2026 ranked AI engineer as the fastest-growing job title in the United States.
Do I need a PhD for remote AI and ML engineering jobs?
For research roles, usually yes. For the majority of the 708 remote roles we counted, no — they ask for production engineering ability, evaluation discipline and judgement about model behaviour. "PhD preferred" in a posting that otherwise describes pipelines and deployments is generally a nice-to-have written by a recruiter.
What do AI and ML engineering interviews ask about in 2026?
Across the 2026 interview guides we read in September 2026, the recurring rounds were ordinary coding, ML fundamentals such as metrics and data leakage, LLM fundamentals including fine-tuning and LoRA, end-to-end RAG system design, evaluation of non-deterministic systems, production failure modes such as prompt injection and retrieval drift, and cost and latency trade-offs.
How many remote AI and ML engineering jobs publish a salary range?
16.5% of the 708 roles we counted on 20 September 2026 published a range, against 15.6% across all remote postings in our index. That is slightly better than the market and still means most roles tell you nothing about pay until you ask.