If you want to know how to get into alignment, the short answer is a four step path: learn Python plus the basics of how large language models are trained, take a free structured course such as BlueDot Impact's Technical AI Safety course, build one small public research project, then apply to a funded fellowship (MATS, the Anthropic Fellows Program or ARENA) or straight to safety teams. You do not need a PhD to start. Anthropic says outright that its fellows program does not require one, and 80,000 Hours notes that strong software engineers can apply for empirical research roles right away.
The longer answer depends on which kind of alignment work suits you. This guide maps the main tracks, the programs that are actually open in late 2026, how the money works and the mistakes that slow most people down.
What "getting into alignment" actually means
AI alignment is the part of AI safety that tries to make models pursue the goals their developers and users intended. If you want the wider picture first, our guide on what AI safety is explains how alignment fits next to security, misuse and policy work.
According to the 80,000 Hours career review of AI safety technical research, most technical work sits on a spectrum:
- Empirical research: running experiments on current models. Interpretability, model organisms of misalignment, AI control, scalable oversight and robustness all live here. It involves a lot of programming and machine learning engineering.
- Theoretical research: mathematical and conceptual work on what safe systems would need. There are far fewer of these jobs, and a strong maths or theoretical computer science background is common.
Inside both, you can aim to be a research contributor (doing the engineering and experiments) or a research lead (choosing questions and running projects). Most leads have PhDs; many contributors do not.
There are also nontechnical routes. MATS now runs tracks in policy and governance, strategy and forecasting, systems security, biosecurity and founding new organizations, alongside its empirical and theory tracks, as listed on the MATS site.
The main ways into alignment compared
| Route | Format | Cost to you | Best for | Status (late October 2026) |
|---|---|---|---|---|
| BlueDot Technical AI Safety course | Online, small group discussions, about 6 days intensive or about 6 weeks part time | Free | Your first structured overview | Cohorts running; check the course page for dates |
| ARENA | In person bootcamp in London, 4 to 5 weeks | Travel, housing and meals covered | Hands on safety engineering skills | ARENA 9.0 runs Oct 5 to Nov 6, 2026; applications closed |
| MATS | Full time research fellowship, 12 weeks in Berkeley or London | Funded (stipend, housing, compute) | Producing real research with a mentor | Winter 2027 runs Jan 19 to Apr 10, 2027; applications closed |
| Anthropic Fellows Program | 4 months of empirical research with Anthropic mentors | Funded (stipend and compute) | Strong coders ready for empirical work | Rolling applications |
| BlueDot Career Transition Grants | Grant to work full time on your switch | Funded | People who need time and money to retrain | Rolling applications |
| Direct jobs | Safety teams at labs, nonprofits and government | Salaried | Experienced engineers and researchers | Ongoing |
Step 1: Build the technical basics
80,000 Hours lists three foundations for technical alignment: programming (Python above all), the maths behind deep learning (linear algebra, calculus and probability) and practical machine learning, including a framework like PyTorch. It also stresses that programming and basic practical ML matter more than mastering the maths up front.
You do not need to finish a degree in all of this before moving on. BlueDot's Technical AI Safety course page says you only need to understand how LLMs are trained and fine tuned, that progress is driven by data, algorithms and compute, and that networks are optimised with gradient descent. BlueDot offers a free, self paced AI Foundations course of about two hours to cover that ground.
A useful self test from the same 80,000 Hours review: Google DeepMind's safety team once wrote that if you can reproduce a typical ML paper in a few hundred hours, it is probably interested in interviewing you for a research engineer role.
Step 2: Take a structured AI safety course
Reading at random is the slowest way in. A course gives you a map of the open problems and a cohort of peers.
The BlueDot Impact Technical AI Safety course is the most common starting point. According to BlueDot:
- It is completely free, with no hidden costs.
- You choose an intensive version (about 6 days at about 5 hours a day) or a part time version (about 6 weeks at about 5 hours a week).
- Each session pairs 2 to 3 hours of reading and writing with a 2 hour Zoom discussion among about 8 peers, led by an AI safety expert.
- It covers alignment and RLHF, mechanistic interpretability, evaluations and red teaming, AI control and scalable oversight.
- Graduates can apply to a Project Sprint to produce a first piece of safety work with an expert.
If you prefer to learn alone, the ARENA curriculum is published online for free, and 80,000 Hours also points to Rob Miles' YouTube channel, the AXRP podcast and the Center for AI Safety's Intro to ML Safety course. You can see who runs that last one on our Center for AI Safety page.
Step 3: Do a small, public project
Every program in this guide asks, in some form, what you have actually built. The cheapest proof is a small project you can show:
- Replicate a result from a recent safety paper and put the code on GitHub.
- Write up what you found, including what failed, as a short post. The Alignment Forum is where much of the field discusses research.
- Ask for feedback before you sink more than 100 hours into a replication, as 80,000 Hours suggests.
Good starting topics include interpretability on small models, simple model organism experiments and reproducing a published study of risky model behaviour. Keep the project small and focused on one question.
Step 4: Apply to a fellowship or bootcamp
Fellowships are now the main bridge from "interested" to "employed". They pay you, give you a mentor and leave you with a public output.
MATS
MATS describes itself as an independent research fellowship that pairs researchers with mentors in alignment, transparency and security. The main program lasts 12 weeks in Berkeley, California or London, and fellows can apply for a funded 6 to 12 month extension. MATS says fellows get a stipend, housing, meals, a compute budget, a research manager and office space. The MATS FAQ lists the Winter 2027 cohort as January 19 to April 10, 2027, at 40 hours a week, with applications already closed. The stipend figure varies between cohorts and pages, so check the current one before you apply.
MATS reports that since late 2021, 631 researchers have taken part, and that 75% of alumni now work in alignment, transparency and security. Those are the program's own figures.
Anthropic Fellows Program
The Anthropic Fellows Program funds engineers and researchers for 4 months of empirical safety research with Anthropic mentors, with the aim of a public output such as a paper. Research areas include scalable oversight, adversarial robustness and AI control, model organisms, mechanistic interpretability, AI security and model welfare. Applications are rolling.
What stands out is the bar. Anthropic says it cares more about your ability to execute research than your credentials, and that you do not need a PhD, prior ML experience or published papers. It wants strong coders who can make progress on ambiguous problems and who want to move into safety research. Past fellows came from physics, maths, computer science and cybersecurity. You can read more about the lab's assistant on our Claude page.
ARENA
ARENA, the Alignment Research Engineer Accelerator, runs 2 to 3 in person bootcamps a year, each lasting 4 to 5 weeks, at the London Initiative for Safe AI. It covers travel, visa costs, housing and meals, and it asks applicants to code well in Python and know the maths behind modern AI. Its process is an application form, a one hour coding test and a 30 minute interview. ARENA 9.0 runs from October 5 to November 6, 2026, and there is an expression of interest form for future rounds.
Step 5: Fund the switch if you need to
Money is the barrier that stops many career changers. In September 2026 the main grant program for this moved. Coefficient Giving says it handed its Career Development and Transition Funding program to BlueDot Impact on September 4, 2026, and that new applicants should apply through BlueDot.
BlueDot's Career Transition Grants let you work full time on your transition. BlueDot says these grants generally start at $20,000 and can go up to $200,000, with the amount and duration agreed around your plan; smaller requests go to its Rapid Grants. Applications are rolling.
Government money exists too. The UK AI Security Institute's Alignment Project funds outside alignment research, and our guide to the UK AI Security Institute covers its first grants.
Step 6: Land your first alignment job
The organisations hiring for safety work fall into a few groups, according to 80,000 Hours and the organisations themselves:
| Type | Examples | What they do |
|---|---|---|
| Frontier labs | Anthropic, Google DeepMind, OpenAI | Safety, alignment and preparedness teams inside model developers |
| Safety nonprofits | Redwood Research, Apollo Research, METR | AI control, misalignment and deception research, risk assessments of frontier models |
| Theory groups | MIRI, Alignment Research Center | Conceptual and mathematical alignment work |
| Government | UK AI Security Institute | Testing frontier models and funding outside research |
| Academia | UC Berkeley CHAI, MIT, NYU, Cambridge groups | PhDs and academic research |
80,000 Hours notes that work is concentrated in the San Francisco Bay Area and London, though remote roles are becoming more common. It also warns that some jobs on the way in may advance AI capabilities, which some experts see as harmful, so think hard about what each role actually contributes. Our comparison of the Preparedness Framework and Responsible Scaling Policy shows how two labs describe their own safety commitments.
Pros and cons of an alignment career
Pros
- Funded entry routes now exist: paid fellowships, free courses and transition grants.
- Credentials matter less than shown ability at several programs, including Anthropic's.
- 80,000 Hours describes the field as short of skilled researchers and engineers, and industry roles pay competitively.
- The work is intellectually hard and tied to how frontier models, such as those in our ChatGPT vs Claude vs Gemini vs Grok comparison, are built and tested.
Cons
- Entry is competitive, and 80,000 Hours points to a shortage of managers and mentors.
- Most technical roles need strong quantitative and programming skills.
- Jobs cluster in a few cities.
- Researchers disagree sharply on which approaches will work, and some roles risk speeding up capabilities.
FAQ
Can I get into alignment without a PhD?
Yes. Anthropic says its fellows program does not require a PhD, prior ML experience or published papers, and 80,000 Hours says strong software engineers can apply for empirical contributor roles. A PhD mainly helps if you want to lead research or work in academia.
How long does it take to get into AI alignment?
There is no fixed timeline. 80,000 Hours says some people from other quantitative fields have moved into safety in less than a year, while others spend several years building skills first.
Is there a free AI safety course?
Yes. BlueDot Impact's Technical AI Safety course is free, and so are its AI Foundations course and the ARENA curriculum published online.
Do AI safety fellowships pay?
The major ones do. MATS and the Anthropic Fellows Program pay a stipend and add a compute budget, and ARENA covers travel, housing and meals. Amounts change by cohort, so check each program's current page.
Can I work in alignment if I am not technical?
Yes. Policy and governance, strategy and forecasting, field building and operations roles all exist, and MATS runs dedicated tracks for several of them. Our superintelligence vs AGI explainer is a good primer on the ideas that policy work deals with.
Where should I start this week?
Do BlueDot's two hour AI Foundations course, then apply for the next Technical AI Safety cohort. While you wait, pick one safety paper to replicate.
