AI research scientist job description templates: 6 variants for LLM, applied, safety, research engineer, and intern roles, with pay, FLSA, and IP notes.
6 templates for the teams hiring research talent without a research HR function: general, foundation model, applied, safety, research engineer, and paid intern, each with the classification, IP, and export control notes the generic versions skip. Download as DOCX.
The first AI research scientist posting I ever reviewed for a founder had one line about the research problem and nine bullets about programming languages. It read like an engineering job with a doctorate requirement stapled to the front, which is exactly how the candidates read it too. Nobody good applied.
Research hiring breaks the usual job description rules. The candidates you want are evaluating you as much as you are evaluating them, and they screen on two things most postings never mention: what compute they will have, and whether they are allowed to publish. Leave those out and your posting looks like every other one.
At FirstHR we write hiring templates for companies without a dedicated HR function, and this role is a hard one for them. The six templates below cover a general research scientist, a foundation model scientist, an applied scientist, a safety and evaluation scientist, a research engineer, and a paid research intern, each with the classification, intellectual property, and export control notes the generic versions leave out.
TL;DR
An AI research scientist creates new capability rather than applying existing capability, and the job description has to state the research problem, the compute budget, and the publication policy to attract anyone serious. The nearest federal benchmark, computer and information research scientists, had a median wage of $140,300 (BLS OEWS, May 2025). Six templates below, downloadable as DOCX.
What an AI Research Scientist Does
An AI research scientist owns a question, not a task. The work is formulating a research problem, designing controlled experiments, developing or adapting methods, evaluating results against real baselines, and converting what survives into a publication, a prototype, or a shipped capability.
That ownership is the whole distinction. An engineer executes an approach that has already been chosen; a scientist decides what is worth trying and accepts that most attempts will fail. Budget for the failure rate honestly, because a research function measured on shipping velocity stops doing research within a quarter.
The role sits inside a small federal occupation. Computer and information research scientists numbered about 40,300 employees in 2024, with roughly 3,200 job openings projected per year through 2034 and growth rated much faster than average by the Bureau of Labor Statistics. That is a tiny pool competing against very well-funded employers.
Scientist, Engineer, or Something Else Entirely
Four adjacent roles get posted under research titles, and picking the wrong one is the most expensive mistake in this hire. The dividing line is whether your problem needs a new method invented or an existing method applied well.
Research scientist
Owns the question
Sets a research direction, formulates the problem, and decides what to try. Measured by new capability or new knowledge, usually holds a doctorate, and expects a compute budget and a publication policy in writing before accepting.
Research engineer
Owns the machinery
Builds the training and evaluation infrastructure the scientists run on, implements methods from papers, and makes experiments fast and reproducible. Often the higher-value first hire at a small company.
Machine learning engineer
Owns the system in production
Takes known methods and ships them: serving, monitoring, retraining, cost. If your problem is solvable with methods that already exist, this is the role you actually need.
Data scientist
Owns the answer from data
Analyzes data to inform decisions and builds models where a model is warranted. Overlaps with research at the edges, but the job is insight and measurement rather than novel method development.
Write the Problem Before You Write the Posting
Draft one paragraph describing the problem the hire will own, then ask a blunt question: could a strong engineer solve this with published methods and available models? If the answer is yes, you need a machine learning engineer or an AI engineer, and you will fill that role faster and for less money. Post a research role only when the capability you need does not exist yet and your business depends on creating it.
The comparison also runs the other way. If your problem is measurement and insight rather than new capability, a data scientist fits better, and if the work is mostly designing and testing model inputs, look at the prompt engineer templates instead.
What Belongs in the Posting
A research job description does four jobs at once: it states a problem worth solving, it filters people who cannot do research, it protects your intellectual property, and it closes the candidate. Most postings do only the fourth, badly. Here is the full inventory.
The parts candidates read first
The research problem, stated concretely
Subfield and how it connects to the product
Compute budget, in real terms
Publication policy, stated before the first call
The parts that filter applicants
Degree expectation marked required or preferred
Research output you expect to see
Split between open research and product work
Engineering depth the role genuinely needs
The parts that protect you
FLSA classification stated on the posting
Invention assignment and confidentiality at offer
Export control screening where it applies
Equal opportunity statement
The parts that win the hire
Base salary or a good-faith range
Equity, refresh policy, and what it is worth
Who the scientist reports to and works with
A named person and a real decision timeline
The single most common omission is the compute budget. Serious candidates ask about it in the first fifteen minutes because it determines what research is possible, and a vague answer signals a small budget you are hiding. State it plainly, even when it is modest. Our guide to writing a job description covers the general structure in more depth.
6 AI Research Scientist Job Description Templates to Download
Download all six as one file or copy them individually. Each follows the same structure: company overview, position summary, key responsibilities, required and preferred qualifications, a classification and compliance note, an equal opportunity statement, and how to apply. The bracketed fields are the only parts you need to change.
Download All 6 AI Research Scientist Job Description Templates
General, foundation model, applied, safety, research engineer, and paid intern. All in one download.
AI Research Scientist (General)
Field-neutral base
The adaptable core: research agenda, experiment design, publication and product transfer, with the compute and publication policy fields candidates ask about first.
Foundation Model / LLM
Pretraining and post-training
For modeling work on large models: training and adaptation methods, evaluation that predicts real performance, and the export control note that comes with frontier compute.
Applied Research Scientist
Research that ships
For a role measured by shipped capability rather than papers: problem translation, fast prototyping, honest evaluation, and the handoff to engineering.
AI Safety / Alignment
Evaluation and mitigation
For understanding how your models fail and proving the fixes work, with a written disclosure policy so nobody discovers the disagreement after a bad finding.
Research Engineer (AI)
Infrastructure for research
For the person who makes experiments fast, cheap, and reproducible. Often the hire a small team actually needs before it needs a scientist.
Research Intern / Early Career
Paid, scoped, mentored
For a research internship or first postdoctoral hire, with the primary beneficiary warning and the non-exempt classification stated plainly.
Template 1: AI Research Scientist (General)
The adaptable core, with a research environment section covering compute, data, publication policy, and the split between open research and product work.
AI Research Scientist Job Description (General)
AI RESEARCH SCIENTIST JOB DESCRIPTION
Company: __ ([City, State] / Remote)
Reports to: [Head of Research / Director of AI / CTO]
Employment type: Full-time, W-2
FLSA status: Exempt (learned professional; see classification note)
Compensation: $_____ per year, plus [bonus / equity]
ABOUT [COMPANY NAME]
[Company Name] is a [stage, size, industry] company working on [the specific
problem your research addresses]. Our research group is [number] people and
works [independently / embedded with the product engineering team].
POSITION SUMMARY
The AI Research Scientist sets and pursues a research agenda in [subfield],
formulates problems, designs and runs experiments, develops new methods, and
turns results into publications, prototypes, or product capabilities.
KEY RESPONSIBILITIES
•Formulate research problems in [subfield] and propose approaches to them
•Design, run, and analyze experiments with proper baselines and ablations
•Develop, train, and evaluate models; report results honestly, including
negative ones
•Write reproducible research code and maintain experiment records
•Track the literature and bring relevant methods into our work
•Publish, present at [venues], or transfer findings into the product
•Collaborate with engineering, product, and data teams on what ships
•Experience with large-scale training or distributed compute
•A record of moving research results into production systems
•Open-source contributions or reviewing for [venues]
RESEARCH ENVIRONMENT (be specific, candidates compare this)
•Compute budget: [describe the actual cluster or cloud allocation]
•Publication policy: [free to publish / internal review / restricted]
•Data access: [what data the scientist will work with]
•Split between open research and product work: [rough percentage]
CLASSIFICATION, IP, AND COMPLIANCE NOTE (read before posting)
This role is almost always exempt under the FLSA learned professional
exemption, which covers work requiring advanced knowledge in a field of science
or learning customarily acquired by a prolonged course of specialized
intellectual instruction. Classify on the actual primary duty, not the title,
and confirm the current federal and state salary thresholds. Pair the offer with
an invention assignment and confidentiality agreement signed before the first
day. Screen for export control exposure if the work touches controlled
technology. This is general information, not legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_____ per year, [bonus], [equity], [benefits summary]
To apply, send your resume and [publication list / portfolio / code samples] to
__ by _.
Template 2: Foundation Model / LLM Research Scientist
For modeling work on large models: training and adaptation methods, evaluation that predicts real task performance, and the export control note that comes with frontier compute.
Foundation Model / LLM Research Scientist Job Description
FOUNDATION MODEL / LLM RESEARCH SCIENTIST JOB DESCRIPTION
Company: __ ([City, State] / Remote)
Reports to: [Head of Research / Research Lead, Modeling]
Employment type: Full-time, W-2
FLSA status: Exempt (learned professional; see classification note)
Compensation: $_____ per year, plus [bonus / equity]
ABOUT THIS ROLE
[Company Name] trains and adapts large models for [domain]. We are hiring a
Research Scientist to work on [pretraining / post-training / evaluation /
retrieval / efficiency] and to own a research direction end to end.
POSITION SUMMARY
The Foundation Model Research Scientist advances our modeling work: designing
training and adaptation methods, running controlled experiments at scale,
building evaluation that reflects real use, and publishing or shipping results.
KEY RESPONSIBILITIES
•Design and run [pretraining / fine-tuning / preference optimization]
experiments and interpret the results
•Build evaluation sets and metrics that predict real task performance rather
than benchmark scores alone
•Investigate failure modes: hallucination, brittleness, distribution shift,
and prompt sensitivity
•Improve data quality, curation, and mixing strategy for [domain]
•Work with infrastructure on training throughput, checkpointing, and cost
•Document methods and results so another scientist can reproduce them
•Partner with product on what a research result means for users
REQUIRED QUALIFICATIONS
•PhD or equivalent research experience in machine learning or a related field
•Hands-on experience training or adapting large neural models
•Strong experimental discipline: baselines, ablations, and honest reporting
•Proficiency in Python, a modern deep learning framework, and distributed
training tooling
•Working knowledge of evaluation design and its limits
PREFERRED QUALIFICATIONS
•Publications at [venues] on language modeling, alignment, or evaluation
•Experience with [retrieval / multimodal / long-context / quantization] work
•Familiarity with inference cost and latency constraints in production
•Contributions to open model or evaluation ecosystems
COMPUTE AND DATA (state this honestly)
•Available compute: [describe it; candidates will ask in the first call]
•Data: [licensed / proprietary / public], with [describe access controls]
•Publication policy: [free to publish / internal review / restricted]
CLASSIFICATION, IP, AND COMPLIANCE NOTE
Exempt under the learned professional exemption in nearly all cases; classify
on primary duty and confirm current thresholds. Use an invention assignment and
confidentiality agreement signed before day one, and be explicit in writing
about what the scientist may publish. Advanced computing hardware and some model
weights sit inside export control rules, and releasing controlled technology to
a foreign national inside the United States can count as an export, so run the
screen before granting cluster access. Document model provenance, training data
sources, and evaluation results as you go. This is general information, not
legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_____ per year, [bonus], [equity], [benefits summary]
To apply, send your resume and [publication list / technical writeup] to
__ by _.
Still Using Spreadsheets for Onboarding?
Automate documents, training assignments, task management, and track onboarding progress in real time.
For a research role measured by shipped capability rather than publication count, with problem translation, fast prototyping, and the handoff to engineering written in.
Applied AI Research Scientist Job Description
APPLIED AI RESEARCH SCIENTIST JOB DESCRIPTION
Company: __ ([City, State] / Remote)
Reports to: [Head of AI / VP Engineering / CTO]
Employment type: Full-time, W-2
FLSA status: Exempt (learned professional or computer employee; see note)
Compensation: $_____ per year, plus [bonus / equity]
ABOUT THIS ROLE
[Company Name] is hiring an Applied Research Scientist to close the gap between
what the literature can do and what our product does. This is a research role
measured by shipped capability, not by publication count.
POSITION SUMMARY
The Applied AI Research Scientist takes ambiguous product problems, decides
whether a research approach is warranted, prototypes solutions, proves them with
offline and online evaluation, and hands them to engineering to run.
KEY RESPONSIBILITIES
•Translate product problems into measurable modeling problems
•Survey the literature and pick approaches worth the cost of trying
•Prototype quickly, then evaluate rigorously offline and in [A/B tests]
•Define success metrics with product before building anything
•Work with engineering on the production path: latency, cost, and monitoring
•Write up what worked, what failed, and why, so the team does not repeat it
•Advise on build versus buy versus fine-tune for each capability
REQUIRED QUALIFICATIONS
•PhD or master's degree in a quantitative field, or equivalent applied research
experience
•A track record of research work that reached real users
•Strong software engineering fundamentals alongside the research skills
•Comfort with incomplete data, shifting requirements, and negative results
•Clear written communication with product and executive audiences
PREFERRED QUALIFICATIONS
•Experience in [our domain: healthcare, fintech, logistics, security]
•Familiarity with online experimentation and its statistical pitfalls
•Experience mentoring engineers on applied modeling work
CLASSIFICATION, IP, AND COMPLIANCE NOTE
Applied research roles are exempt in nearly all cases, usually under the learned
professional exemption and sometimes the computer employee exemption. Classify
on primary duty and confirm the current thresholds. Use an invention assignment
and confidentiality agreement at hire. If model output touches hiring, lending,
housing, insurance, or health decisions, treat the resulting system as regulated
and document how it was evaluated for disparate impact before launch. This is
general information, not legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_____ per year, [bonus], [equity], [benefits summary]
To apply, send your resume and a short writeup of a research result you shipped
to __ by _.
Template 4: AI Safety / Alignment Research Scientist
For understanding how your models fail and proving the mitigations work, with a written disclosure policy so a difficult finding does not turn into a difficult conversation.
AI Safety / Alignment Research Scientist Job Description
AI SAFETY / ALIGNMENT RESEARCH SCIENTIST JOB DESCRIPTION
Company: __ ([City, State] / Remote)
Reports to: [Head of Safety / Head of Research]
Employment type: Full-time, W-2
FLSA status: Exempt (learned professional; see classification note)
Compensation: $_____ per year, plus [bonus / equity]
ABOUT THIS ROLE
[Company Name] builds [describe the system]. We are hiring a Research Scientist
focused on safety, alignment, and evaluation: understanding how our models fail,
measuring it, and reducing it before customers find it.
POSITION SUMMARY
The AI Safety Research Scientist studies model behavior under adversarial and
realistic conditions, designs evaluations and mitigations, and gives the company
an honest read on what the system does and does not do reliably.
KEY RESPONSIBILITIES
•Design and run evaluations for [robustness / bias / misuse / jailbreaks /
factual reliability] in our specific deployment context
•Red team our own systems and document reproducible failure cases
•Develop mitigations and measure whether they actually work
•Build monitoring that catches regressions after release
•Advise product and legal on residual risk in plain language
•Map our practices against a recognized risk management framework
•Publish or share findings externally where policy allows
REQUIRED QUALIFICATIONS
•PhD or equivalent research experience in machine learning, security, or a
related field
•Experience designing evaluations rather than only consuming benchmarks
•Statistical literacy: sampling, confidence, and the limits of a small eval set
•Ability to write findings that a non-technical executive can act on
•Judgment about what is a real risk in our context and what is not
PREFERRED QUALIFICATIONS
•Publications on evaluation, interpretability, robustness, or alignment
•Security or adversarial machine learning background
•Familiarity with sector rules that apply to our customers
CLASSIFICATION, IP, AND COMPLIANCE NOTE
Exempt under the learned professional exemption in nearly all cases; classify on
primary duty and confirm current thresholds. Safety findings are sensitive:
define in writing what the scientist may disclose, to whom, and on what
timeline, and put that in the offer rather than discovering the disagreement
later. A voluntary risk management framework is not law, but documenting your
process against one is the cheapest evidence of diligence a small company can
produce. This is general information, not legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_____ per year, [bonus], [equity], [benefits summary]
To apply, send your resume and an example evaluation or red team writeup to
__ by _.
Template 5: Research Engineer (AI)
For the person who makes experiments fast, cheap, and reproducible. If you are choosing between this and a scientist as a first research hire, most small teams get more from this one.
Research Engineer (AI) Job Description
RESEARCH ENGINEER (AI) JOB DESCRIPTION
Company: __ ([City, State] / Remote)
Reports to: [Research Lead / Head of AI Infrastructure]
Employment type: Full-time, W-2
FLSA status: Exempt (computer employee or learned professional; see note)
Compensation: $_____ per year, plus [bonus / equity]
ABOUT THIS ROLE
[Company Name] is hiring a Research Engineer to build the systems our research
runs on. If you like making experiments fast, reproducible, and cheap more than
you like writing papers, this is the role, not the scientist posting.
POSITION SUMMARY
The Research Engineer builds and maintains training and evaluation
infrastructure, implements methods from papers, runs experiments at scale, and
turns research prototypes into code the team can rely on.
KEY RESPONSIBILITIES
•Build and maintain training, evaluation, and experiment tracking pipelines
•Implement methods from the literature correctly and verify them against
reported results
•Profile and optimize training throughput, memory, and cost
•Manage data pipelines: ingestion, cleaning, versioning, and access controls
•Support scientists by removing infrastructure friction from their experiments
•Harden promising prototypes into maintainable code
•Keep experiment records and artifacts organized and reproducible
REQUIRED QUALIFICATIONS
•Bachelor's or master's degree in computer science or equivalent experience
•Strong Python and systems engineering skills
•Hands-on experience with distributed training and GPU or accelerator workloads
•Comfort reading papers and reproducing their results
•Pragmatism about when good enough is good enough
PREFERRED QUALIFICATIONS
•Experience with [cluster scheduling / cloud accelerators / inference serving]
•Contributions to open-source machine learning tooling
•Prior work inside a research group
CLASSIFICATION, IP, AND COMPLIANCE NOTE
Exempt in nearly all cases, usually under the computer employee exemption and
sometimes the learned professional exemption. The computer employee exemption
carries its own salary or hourly rate condition, so check the current federal
and state thresholds rather than assuming. Use an invention assignment and
confidentiality agreement at hire. Because this role administers access to
compute, data, and model weights, include it in your export control screen and
your access review. This is general information, not legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_____ per year, [bonus], [equity], [benefits summary]
To apply, send your resume and links to code you have written to
__ by _.
Template 6: AI Research Intern / Early Career Scientist
For a paid research internship or a first early-career hire, with the primary beneficiary warning and the non-exempt classification stated plainly. More templates for adjacent roles sit in our hiring templates library.
AI Research Intern / Early Career Scientist Job Description
AI RESEARCH INTERN / EARLY CAREER SCIENTIST JOB DESCRIPTION
FLSA status: Non-exempt for most paid interns; see classification note
Compensation: $_ per [hour / month]
ABOUT THIS ROLE
[Company Name] runs a [number] week paid research program for [PhD students /
recent graduates]. Each participant owns one scoped project with a named mentor
and presents results at the end.
POSITION SUMMARY
The Research Intern works on a defined research question under the guidance of a
mentor: reviewing prior work, running experiments, analyzing results, and
producing a writeup or prototype the team can build on.
KEY RESPONSIBILITIES
•Own one scoped research project from question to writeup
•Review relevant prior work and summarize it for the team
•Run and document experiments with proper baselines
•Meet weekly with the mentor and report progress honestly
•Present findings to the research group at the end of the term
•Follow all data handling, confidentiality, and access rules
REQUIRED QUALIFICATIONS
•[Enrolled in a PhD or master's program / within [number] years of completing
one] in a quantitative field
•Coursework or project experience in machine learning
•Working Python and one deep learning framework
•Ability to work independently between mentor check-ins
CLASSIFICATION AND COMPLIANCE NOTE (read this one carefully)
Unpaid internships at for-profit companies are risky. The Department of Labor
applies a primary beneficiary test to decide whether an intern is an employee
entitled to minimum wage and overtime, and a research intern doing productive
work that benefits the company will usually be an employee. The practical answer
for a small company is simple: pay the intern, classify them as non-exempt
unless they clearly meet an exemption, and track hours. A short-term intern
rarely satisfies the learned professional test on their own. Confirm work
authorization and any student visa training rules before the start date, and
include interns in your export control screen if they will touch controlled
technology. This is general information, not legal advice.
EEO STATEMENT
[Company Name] is an equal opportunity employer and provides reasonable
accommodations for the essential functions of this role.
COMPENSATION AND HOW TO APPLY
Compensation: $_ per [hour / month], [housing or relocation stipend]
To apply, send your resume, transcript or advisor reference, and a short
statement of research interest to __ by _.
Classification, IP, and Export Controls
Three compliance items attach to a research hire, and only the first one is familiar to most employers. Classification is straightforward, intellectual property is the reason the hire exists, and export control screening is the requirement small AI teams miss most often.
On classification, the federal regulation on learned professionals covers work requiring advanced knowledge in a field of science or learning customarily acquired by a prolonged course of specialized intellectual instruction. A research scientist fits it about as cleanly as any role does. Our breakdown of exempt versus non-exempt classification works through the tests.
Classification is easy here, but write it down
A research scientist with a doctorate doing original technical work is close to the clearest case there is for the FLSA learned professional exemption, which covers advanced knowledge in a field of science or learning customarily acquired by a prolonged course of specialized intellectual instruction. The federal salary floor for white-collar exemptions is $684 per week, or $35,568 per year, and no serious research offer comes near it. Two caveats keep this from being automatic. Your state may set a higher floor than the federal one, and a research engineer classified under the computer employee exemption sits under a separate rate condition. Record the exemption you relied on and why in the employee file at hire, because reconstructing that reasoning two years later during a wage audit is much harder than writing one sentence now. This is general information, not legal advice.
Invention assignment is the whole point of the hire
You are paying a scientist to create new intellectual property, so the agreement that assigns it to the company is not paperwork, it is the asset. Get an invention assignment and confidentiality agreement signed before the first day, not during onboarding week, and make sure it is in the offer package so nobody is surprised by it after resigning from another job. Several states limit how far an assignment can reach into work done entirely on the employee’s own time with their own equipment, and some require written notice of that limit, so use an agreement drafted for your state rather than a template pulled off the internet. Add a clear prior inventions schedule where the candidate lists what they are bringing in, and keep the signed copy with the employee record where you can find it during diligence.
Export controls apply before the first login
This is the requirement small AI teams miss most often. Under the deemed export rule, releasing controlled technology or source code to a foreign national inside the United States is treated as an export to that person’s home country, and it can require a license. Advanced computing hardware and certain model weights fall inside those controls. In practice that means the screen belongs in the hiring sequence, before you grant cluster access, not after the scientist has been running jobs for a month. Decide who owns the check, document the outcome for each hire, and re-run it when someone changes projects or gains access to a new system. A small research team with no compliance function still carries the obligation, and the penalties do not scale down with headcount.
Document your AI risk process while it is still cheap
No single federal statute governs AI development today, but sector rules already apply to what your models touch. A system that influences hiring, lending, housing, insurance, or health decisions is regulated through those existing laws regardless of the technology behind it. Voluntary frameworks fill the gap: mapping your evaluation, monitoring, and incident process against a recognized risk management framework is not legally required, and it is the cheapest evidence of diligence a small company can produce for a customer security review, an enterprise procurement questionnaire, or a regulator. Assign that documentation to the safety or applied scientist explicitly in the job description. Work nobody owns does not get done, and reconstructing an evaluation history after a customer asks for it is far more expensive than keeping it as you go.
Export controls deserve one more sentence because the sequencing matters. Under the deemed export rule, releasing controlled technology or source code to a foreign national inside the United States is treated as an export to that person's home country and can require a license. Run the screen before granting cluster access, not after.
For the fourth item, governance, no single federal statute covers AI development, but existing sector law already reaches whatever your models touch. Mapping your evaluation and monitoring process against the NIST AI Risk Management Framework is voluntary and is the cheapest diligence evidence a small company can produce.
Requirements That Actually Filter
The requirements section of a research posting should filter on evidence of original work, not on a tool list. A framework list filters for people who used a framework, which is nearly everyone, while a request for a defensible research result filters for people who produced one.
Mark the degree expectation as required or preferred and mean it. A vague requirement loses both groups: credentialed candidates read it as a role that does not value the credential, and candidates without one read it as a wall.
Requirement
How to write it so it filters correctly
Doctorate
Mark required or preferred; offer equivalent research experience as the honest alternative
Research output
Ask for a publication list, preprints, or a public record of original technical work
Subfield depth
Name the subfield concretely instead of listing every area of machine learning
Engineering skill
State the real bar: research code, or production code, or both, and which one matters more
Compute budget
Not a requirement but a disclosure; state it or expect strong candidates to assume the worst
Publication policy
Free to publish, internal review, or restricted; write the actual answer in the posting
Classification
Exempt learned professional for scientists; check the separate condition for research engineers
IP agreement
State that the offer includes an invention assignment so nobody is surprised after resigning
One more filter worth adding: ask candidates to describe a result they produced that turned out to be wrong, and what they did about it. Research is mostly negative results, and a candidate who cannot discuss one comfortably has usually not run enough experiments to be useful.
What to Pay an AI Research Scientist
There is no Bureau of Labor Statistics occupation titled AI research scientist, so the federal data gives you a floor rather than a target. The nearest classification is computer and information research scientists, with data scientists and software developers as secondary reference points.
Nearest Federal Benchmark: Median $140,300
According to the Bureau of Labor Statistics Occupational Employment and Wage Statistics survey (May 2025), computer and information research scientists had a national median annual wage of $140,300, with the lowest ten percent under $82,200, the twenty-fifth percentile at $103,570, the seventy-fifth at $188,700, and the highest ten percent above $230,630 (U.S. Bureau of Labor Statistics). No federal occupation carries the AI research scientist title, so treat this as the nearest available classification rather than a direct match.
Nearest classification
Median (BLS OEWS, May 2025)
90th percentile
How to use it
Computer and information research scientists
$140,300 per year
$230,630
Closest match for a research scientist role; treat as a base floor
Data scientists
$120,230 per year
$199,130
Reference point for applied and analysis-heavy research roles
Software developers
$135,980 per year
$214,670
Reference point for a research engineer posting
Statisticians
$105,650 per year
Not cited here
Reference point for methods and evaluation work
Read those as base wage medians across a broad employer population, not as offers at an AI lab. National compensation surveys show total compensation for frontier research roles running far above the federal figures once bonus and equity are counted. Benchmark to companies at your stage and funding level, explain what the equity is actually worth rather than quoting a share count, and publish a good-faith range where pay transparency rules apply.
Companies Using FirstHR Onboard 3x Faster
Join hundreds of small businesses who transformed their new hire experience.
Most small companies that write an AI research scientist posting need a different hire. That is the uncomfortable part the template libraries skip, and it is the single highest-value thing to get right before you publish anything.
You may not need a research scientist at all
The honest question first: does your problem require a new method, or does it require competent application of methods that already exist? Almost every small company that posts a research scientist role needs the second one. A research scientist is worth the cost when the capability you need does not exist yet and your business depends on creating it. If a strong engineer using published methods and available models can get you there, you want a machine learning engineer or an applied engineer, and you will fill that role faster and cheaper. Posting a scientist role you do not need has a specific failure mode: you hire someone who wants to do research, hand them integration work, and lose them within a year. Write down the problem the hire will own before you write the job description, and if the answer is application rather than invention, use the applied template or a different role entirely.
You are competing against employers who pay in a different currency
Well-funded labs pay research scientists total compensation that a small company cannot match, and pretending otherwise wastes everyone’s time. What you can offer is different and genuinely valuable to some candidates: ownership of a whole research direction rather than a slice of one, results that reach real users in months rather than years, direct access to the founder and the decision, a permissive publication policy, and a hiring decision made in a week. Those are the things worth putting in the posting. Be equally direct about the constraints, especially the compute budget, because that is the first question a serious candidate asks and a vague answer reads as a small budget you are hiding. A candidate who accepts knowing the real numbers stays; one who finds out in month two does not.
The paperwork for a research hire is unusually heavy and lands at once
A research hire arrives with more documents than almost any other role: signed offer, invention assignment and confidentiality agreement, prior inventions schedule, publication policy acknowledgment, export control screening record, data access and security agreements, work authorization, plus the ordinary handbook and payroll forms. Missing one of those is not a filing problem, it is a claim on your intellectual property or an access control gap. FirstHR was built for exactly this kind of burst. The onboarding wizard runs the same sequence for every hire, e-signature covers the invention assignment and policy acknowledgments, document management stores the signed agreements and screening records against each employee profile, and training modules handle security and data handling orientation before day one. Applicant tracking is coming soon to FirstHR. Note that FirstHR is an onboarding and HR platform, not a payroll provider.
The Mismatch That Costs You a Year
The predictable failure is hiring a researcher into an integration job. Someone joins expecting to own a research direction, spends nine months wiring up an existing model behind an API, and leaves. You lose the hire, the recruiting cost, and the year. Prevent it by describing the actual work in the posting rather than the aspirational version, and by naming the split between research and product work as a rough percentage. Candidates who want the mix you are offering will self-select in, and the ones who do not will save you both the trouble.
From Offer to First Experiment
The job description is step one. Once a candidate accepts, the same document becomes the basis for the offer, the classification record, the invention assignment, and an onboarding sequence that has to clear access controls before any real work starts.
Research onboarding has a hard dependency most roles do not: compute and data access cannot be granted until screening and agreements are complete, so a slow paperwork week is a slow first month of research. Run it as a repeatable onboarding checklist with owners and dates rather than a folder of documents someone chases.
Key Takeaways
An AI research scientist owns a question and creates new capability, while an engineer executes an approach that is already chosen, and posting the wrong one of the two is the most expensive mistake in this hire.
Compute budget and publication policy are the first two things a serious candidate screens on, so state both in the posting instead of leaving them for the second interview.
The nearest federal benchmark, computer and information research scientists, had a median wage of $140,300 with a ninetieth percentile of $230,630 (BLS OEWS, May 2025), and no federal occupation carries the AI research scientist title.
Research scientists are exempt learned professionals in nearly all cases, research engineers may sit under the computer employee exemption, and paid research interns are usually non-exempt with hours tracked.
The invention assignment and confidentiality agreement is the asset you are buying, so it belongs in the offer package and signed before day one, not during onboarding week.
Export control screening belongs in the hiring sequence before cluster access is granted, because releasing controlled technology to a foreign national inside the United States can count as an export.
A research hire arrives with more signed documents than almost any other role. FirstHR runs the same onboarding sequence every time, with e-signature for the invention assignment and policy acknowledgments, document storage for screening records and agreements, and security and data handling training assigned before the first day. Applicant tracking is coming soon to FirstHR.
Frequently Asked Questions
What does an AI research scientist do?
An AI research scientist creates new capability rather than applying existing capability. The work is formulating a research problem, designing and running controlled experiments, developing or adapting methods, evaluating results honestly against real baselines, and turning what survives into a publication, a prototype, or a product feature. The distinguishing test is ownership of the question: a scientist decides what is worth trying, while an engineer executes an approach that is already chosen. Most hold a doctorate, because the job requires reading and extending a technical literature rather than consuming finished tools. In a small company the role usually blends into applied work, which is fine as long as the split is stated in the job description instead of discovered after the hire starts.
What should an AI research scientist job description include?
Eight things: the specific research problem stated concretely, the subfield and how it connects to your product, the compute budget in real terms, the publication policy, the degree expectation marked required or preferred, the split between open research and product work, the FLSA classification, and pay with the equity component explained. The two that small companies skip are the two candidates screen on first: compute and publication policy. A vague answer on either reads as a constraint you are hiding, and strong candidates move on rather than asking. Add the invention assignment requirement, the export control screening notice where it applies, an equal opportunity statement, a named person to apply to, and a real decision timeline. The six templates on this page follow that structure.
Is an AI research scientist exempt from overtime?
Yes, in nearly every real case. A research scientist doing original technical work that requires advanced knowledge in a field of science or learning, customarily acquired through a prolonged course of specialized intellectual instruction, fits the FLSA learned professional exemption about as cleanly as any role does. The federal salary floor for white-collar exemptions is $684 per week, or $35,568 a year, which no serious research offer approaches. Three cautions. Your state may set a higher salary floor than the federal one. A research engineer classified under the computer employee exemption sits under a separate rate condition, so check it rather than assuming. And a short-term paid research intern rarely satisfies the exemption on their own, so classify interns as non-exempt and track their hours. This is general information, not legal advice.
How much does an AI research scientist make?
There is no Bureau of Labor Statistics occupation titled AI research scientist, so the closest federal benchmark is computer and information research scientists. According to the Bureau of Labor Statistics Occupational Employment and Wage Statistics survey (May 2025), that occupation had a national median annual wage of $140,300, with the lowest ten percent under $82,200 and the highest ten percent above $230,630. Data scientists had a median of $120,230 and software developers $135,980 in the same survey. Treat all three as floors rather than targets for a frontier AI role. National compensation surveys show total compensation at well-funded AI labs running far above the federal figures once bonus and equity are counted, and the federal survey measures base wages across a much broader population of employers. Benchmark to companies of your stage and publish a good-faith range where pay transparency rules apply.
What is the difference between an AI research scientist and a machine learning engineer?
The difference is whether the method already exists. A research scientist owns an open question and creates new capability, measured by novel results that hold up under scrutiny. A machine learning engineer takes methods that already work and makes them run reliably in production, measured by systems that serve real traffic at acceptable cost and latency. That distinction drives everything else in the posting: the scientist role expects a doctorate, a publication record, a compute budget, and a stated publication policy, while the engineer role expects production systems experience, monitoring, and cost discipline. Most small companies that post a research scientist role actually need the engineer. Decide by writing down the problem the hire will own, and if it is solvable with published methods and available models, post the engineering role instead.
Do I need a PhD requirement in an AI research scientist job description?
For a true research role, a doctorate or equivalent research experience is a reasonable bar, because the job is reading and extending a technical literature. Write it as required or preferred and mean it, because a vague requirement filters out both groups at once. The honest alternative is equivalent research experience demonstrated by publications, preprints, or a substantial public record of original technical work, which lets you consider candidates who left a program or built their record in industry. What matters more than the credential is the evidence: ask for a publication list, a code repository, or a writeup of a result the candidate produced and can defend in detail. For applied research roles measured by shipped capability rather than papers, a doctorate is often the wrong filter, and the applied template on this page is written accordingly.
How does a small company hire an AI research scientist without an HR department?
Run a short, fixed sequence and decide fast, because speed is one of the few advantages a small employer has. Write down the problem the hire will own, then confirm it genuinely requires new method development rather than application. Post a specific description with the compute budget, publication policy, pay range, and classification stated up front. Screen on a real research result the candidate can defend in detail rather than on interview puzzles. Include one working session with the team they would join. Then move the offer within days, paired with the invention assignment and confidentiality agreement so nothing surfaces late. Onboarding is a checklist, not a pile: signed agreements, export control screening, data access approvals, security training. FirstHR runs that sequence with e-signature, document management, and training modules. Applicant tracking is coming soon to FirstHR.
