To become an AI researcher, most people take one of two routes: a PhD followed by a research scientist role, or strong software and machine learning skills followed by a research engineer role. Fellowships and residencies now offer a third door. Either way, the currency is a visible research record: papers, replications or open-source work.
Definition: An AI researcher studies and improves how AI systems work, through new training methods, architectures, evaluations, interpretability or safety techniques, and shares results through papers, code or internal reports.
It's the narrowest door into AI. It's also the one people romanticize most. Both things can be true, so let's look at what the job really asks for.
What an AI researcher actually does
The day is less glamorous than the headlines. Reading papers. Designing an experiment. Waiting for it to run. Finding a bug in the data loader. Running it again. Writing up what you found, including when nothing worked.
Research happens in three main places:
- Industry labs, where research feeds products or frontier models and teams share huge compute budgets.
- Universities, where you pick your own questions but compete for funding and compute.
- Nonprofits and government, smaller in number, often focused on safety, policy or public-interest science.
Industry has become the main employer. The Stanford AI Index 2026 education chapter reports that 62.75 percent of new AI PhDs in the US and Canada went to industry in 2024, against 31.59 percent to academia, with government below 5 percent.
Research scientist vs research engineer
These two titles cover most AI research jobs, and the difference matters for how you prepare.
| Research scientist | Research engineer | |
|---|---|---|
| Main job | Sets research direction, designs experiments | Makes experiments run, often at large scale |
| Typical background | PhD, publications | Strong software and ML engineering |
| Hiring signal | Papers, research taste | Code quality, systems skill, replications |
| Common risk | Few openings | Work can drift toward pure infrastructure |
Google DeepMind's careers page puts it plainly: its research scientists normally hold a PhD, while research engineers are software engineers with deep machine learning knowledge. That split is fairly typical of large labs, though not universal.
At some labs the line barely exists. Anthropic's careers page notes that its engineers do lots of research and its researchers do lots of engineering, and it encourages engineers to apply as engineers rather than as researchers.
Our read: if you don't have a PhD, the research engineer title is usually the more realistic way into a research team. It's not a consolation prize. Anthropic, for one, notes that engineers appear as authors on all its papers, often as first author.
Do you need a PhD to become an AI researcher?
Short answer: for research scientist roles at major labs, it's common. For research overall, it's not strictly required. The evidence points both ways, so here's both.
The case that a PhD matters:
- The BLS lists a master's degree as the typical entry education for computer and information research scientists and notes that some employers prefer a PhD.
- Google DeepMind states that its research scientists normally hold one.
- A PhD gives you years to build research taste, a publication record and a network of collaborators, which is hard to replicate on nights and weekends.
The case that it doesn't:
- Anthropic's careers page says about half its technical staff have PhDs, and about half had no prior ML experience before joining.
- Anthropic's Fellows Program says no PhD, prior ML experience or published papers are required.
- OpenAI's Residency page explicitly welcomes self-taught and non-traditional candidates with a strong track record of building.
Our take: a PhD is the most common path into research scientist roles, and a strong one if you love the work. It's a poor choice if you're doing it only for the credential. Several years is a long time to spend on something you're merely tolerating.
Worked example: the PhD's opportunity cost
All numbers here are illustrative. Say a PhD stipend is $45,000 a year for five years. Compare that with the BLS median for software developers, $135,980 in May 2025, as a stand-in for what a strong engineer might earn instead.
The gap is about $90,980 a year, or about $454,900 over five years before taxes. That ignores raises, tuition waivers, equity and the higher research pay a PhD can unlock later.
The point isn't the exact figure. It's that the PhD is a real investment, and it's worth knowing what you're buying: time to do research, mentorship and a credential that some research scientist roles still screen for.
How to build an AI research record without a PhD
If you're taking the engineer route, the goal is the same: evidence you can do research. Some ways we'd consider:
- Replicate a paper. Pick a recent result, reproduce it, and write up what matched and what didn't. Replications are useful to the field and show you can run careful experiments. OpenAI cofounder John Schulman's guide to ML research, first written for OpenAI Fellows, recommends the same habit: reimplement ideas from papers and compare your results with the published ones.
- Extend the replication. Change one variable the authors didn't test. That's your first original result.
- Publish openly. A clean write-up with code on arXiv or a blog, and later a workshop paper at a major conference. Workshops tend to be smaller and less selective than main tracks, in our view, which makes them a sensible first venue.
- Contribute to open research code. Fixes and features in widely used training or evaluation libraries put your name next to real research infrastructure.
- Join a research community. Open research groups and online reading groups let you collaborate with working researchers.
- Apply to a fellowship. See the next section.
For students, a research assistant role in a university lab is often a strong first step. Our guide on getting an AI internship covers that route.
AI research fellowships and residencies to know
These change often, so check each program's current page before you plan around it. The details below come from each program's own page:
- Anthropic Fellows Program. A four-month empirical research program in AI safety areas such as interpretability, adversarial robustness and scalable oversight, with about $15,000 a month of compute per fellow. It states that no PhD, prior ML experience or publications are required. Anthropic reports that over 65 percent of fellows published work from the fellowship and over 35 percent from past cohorts later joined Anthropic.
- OpenAI Residency. A six-month, on-site program in San Francisco paying $18,333 a month, aimed at technical people from other fields who want to move into AI research. Applications open on a rolling basis, so check whether it's currently accepting candidates.
- MATS. An independent 12-week research program in Berkeley and London for AI alignment, governance and security, with a $1,250 weekly stipend and a $2,000 weekly compute budget. MATS, founded in late 2021, reports that 75 percent of its alumni now work in AI alignment, transparency and security.
- Google DeepMind Student Researcher Program. Paid, in-person placements of 12 to 24 weeks for students enrolled in bachelor's, master's or PhD programs.
Our view: these programs are competitive, and they reward what you've already built. Treat them as an accelerant, not a starting line.
AI researcher pay and job outlook
The BLS doesn't track "AI researcher" as its own occupation. The closest is computer and information research scientists: median pay of $140,300 in May 2025, about 38,600 jobs in 2025, and projected growth of 22 percent from 2025 to 2035, with roughly 2,900 openings a year.
Where you work changes pay a lot. Per the same BLS page, the median for this occupation was $211,270 at software publishers and $85,460 at colleges and universities in May 2025. In our view, that gap helps explain why most new PhDs head to industry.
The pipeline is growing too. CRA's Taulbee Survey 2025 reports a record 1,351 computer science doctorates across the departments it tracks year over year, and says AI and machine learning specialties claim the largest share of new doctorates. Among US-awarded computing PhDs with known employment, 61.7 percent went to industry and 32.2 percent to academia.
Our read: a small, very well-paid job market with a growing number of qualified candidates. Strong odds for the exceptional, harder for the merely good. That's a reason to build a visible record early.
A step-by-step path to AI research
One way to sequence it, depending on where you start:
As an undergraduate. Take linear algebra, probability, statistics and machine learning. Join a faculty lab as a research assistant. Aim for one co-authored paper or a serious replication before graduating.
Deciding on a PhD. Apply if you want to spend several years choosing your own questions, and pick advisors over school rankings. If you're unsure, a research engineer job first is a good way to test the fit.
As a software engineer. Move toward ML-heavy work at your current company, replicate two papers, publish one extension, then apply for research engineer roles and fellowships.
Applying. Expect coding interviews, ML fundamentals and a deep discussion of your past research. On the 1752vc careers board's AI track, searching for research engineer or research scientist and filtering by level helps you see which labs and AI startups are hiring right now.
"You need to be a genius to do AI research"
Some people believe AI research is reserved for prodigies with perfect math scores. There's a fair version: the competition is intense, and the field rewards unusual depth.
But most research progress comes from careful experiments, good engineering and persistence. Many results come from people who ran the boring ablation nobody else bothered with. We think steady, rigorous work counts for more than raw brilliance in most research teams, though brilliance obviously doesn't hurt. Two well-known researchers lean the same way: Andrej Karpathy, a founding member of OpenAI, wrote in 2019 that patience and attention to detail were the traits he saw tied most closely to success in deep learning, and Schulman argues that picking the right problems matters even more than raw technical skill.
Common mistakes
- Treating the PhD as the goal. It's a tool for doing research, not a finish line.
- Reading without running. Experiments teach what papers can't.
- Ignoring engineering. Weak code slows every experiment you'll ever run.
- Applying only for research scientist roles. Research engineer jobs are often the realistic entry.
- Working alone. Collaborators improve your ideas and your odds.
Where we land
We see AI research as a real but narrow path. If you love open questions and can live with months of failed experiments, it can be one of the most rewarding jobs in tech. If what you want is to build AI products, our guide on how to become an AI engineer is probably a better fit, and AI career paths compares the options.
PhDs coming from academia should also see our guide on moving from academia into tech. It's our answer, not the answer, and the field shifts fast. The AI jobs track shows each employer's own posting date, which helps you see what research teams want today.
The bottom line
The degree opens some doors.
The work you can show opens the rest.
Key takeaways
- AI research jobs split mainly into research scientists, who set direction and often hold PhDs, and research engineers, who make experiments work at scale.
- A PhD is common for research scientist roles but not universal; Anthropic reports about half its technical staff have one.
- Anthropic's Fellows Program and OpenAI's Residency say they take candidates without a PhD, MATS lists no PhD requirement, and Google DeepMind's student researcher program serves enrolled students.
- The BLS puts median pay for computer and information research scientists at $140,300 in May 2025, with 22 percent projected growth from 2025 to 2035.
- Replicating papers, publishing extensions and contributing to open research code are practical ways to build a record without a PhD.
Frequently asked questions
Not always, but it helps for research scientist roles. Google DeepMind says its research scientists normally hold a PhD, and the BLS lists a master's as the typical entry education for computer and information research scientists. Research engineer roles and fellowships such as Anthropic's Fellows Program accept candidates without a PhD who can show strong research or engineering work.
A research scientist usually sets the research direction, designs experiments and leads papers, and often holds a PhD. A research engineer usually builds the code and infrastructure that make experiments run, often at large scale, and is hired mainly on software and ML skills. At some labs the roles overlap heavily, and both appear as paper authors.
The BLS doesn't track AI researchers separately. Its closest category, computer and information research scientists, had a median wage of $140,300 in May 2025, with medians of $211,270 at software publishers and $85,460 at colleges and universities. No consistent public dataset covers total pay at private AI labs, so treat any single figure with care.
It's possible, especially through research engineer roles and fellowships. Anthropic's Fellows Program states that published papers aren't required, and OpenAI's Residency welcomes self-taught candidates with a strong building record. Even so, some public evidence of research ability, such as a careful paper replication with code or an open-source contribution, tends to improve your chances significantly.
Examples include Anthropic's Fellows Program, a four-month AI safety research program; OpenAI's six-month paid Residency for people moving into AI research from other fields; and MATS, a 12-week AI alignment and governance research program that lists no PhD requirement. Each is competitive and changes over time, so check the program's own page for current eligibility and dates.
Sources
- US Bureau of Labor Statistics: Computer and Information Research Scientists, Occupational Outlook Handbook
- US Bureau of Labor Statistics: Software Developers, Quality Assurance Analysts, and Testers, Occupational Outlook Handbook
- Stanford HAI: AI Index Report 2026, Education
- Computing Research Association: Taulbee Survey 2025, Doctoral Degree Production and Employment
- Google DeepMind: Careers
- Google DeepMind: Student Researcher Program
- Anthropic: Careers
- Anthropic Alignment Science: Anthropic Fellows Program for AI safety research
- OpenAI: Residency
- MATS: ML Alignment and Theory Scholars Program
- John Schulman: An Opinionated Guide to ML Research
- Andrej Karpathy: A Recipe for Training Neural Networks
Disclaimer: This guide is for general education only and is not legal, tax or investment advice. Laws, market data and program terms change, so it may not reflect the latest developments or fit your situation. Treat it as a starting point, not a source of truth, and talk to a qualified lawyer, accountant or financial adviser before you make decisions.
