AI Safety Researcher
Make sure powerful AI systems behave safely and as intended.
What does a AI Safety Researcher do?
AI safety researchers study how to make powerful AI systems reliable, interpretable, and aligned with human intentions. The work is a mix of theoretical research, empirical experiments, and writing — and the stakes are considered very high by those in the field. This path takes time to get into but pays off significantly. Good fit if you enjoy deep technical problems and want to specialise in something most people never attempt.
Who it fits
People with strong maths and research ability who want to work on one of the most consequential open problems in technology.
Suits deep, independent learners who enjoy reading academic literature and working on open-ended problems.
Skills you need
Tools you will use
Starter projects
Good projects to practise and add to a portfolio.
- Replicate a key result from an AI safety paper
- Write a summary of a core alignment research agenda
- Run interpretability experiments on a small transformer
- Complete the ARENA curriculum exercises
// ROADMAP
AI Safety Researcher career roadmap
Estimated time: 18–36 months
Badges show what each resource is. Only items marked certification lead to a professional credential — everything else is learning material.
Mathematics & Statistics
12–16 weeksAI safety research requires genuine mathematical depth. Focus on linear algebra, calculus, probability, and information theory. This takes longer than most paths but there are no shortcuts.
Machine Learning (Deep)
12–16 weeksYou need to understand what modern AI systems actually do. Work through the Andrew Ng specialisation, then go deeper with research-level ML. Deep Learning book by Goodfellow et al. is the standard text.
AI Safety Literature
8–12 weeksRead the core AI safety literature. Start with the AI Safety Fundamentals curriculum from BlueDot. Follow the Alignment Forum for current research. This field moves fast.
Interpretability & Alignment Research
12–16 weeksWork through the ARENA curriculum — it is the most structured free resource for learning mechanistic interpretability. Replicate published results to build genuine research skills.
Research & Fellowship Applications
OngoingPublish writeups and research results. Apply for MATS, Redwood Research fellowships, or safety-focused research programmes at Anthropic, DeepMind, and OpenAI. Relationships in the field open doors.
// START LEARNING
AI Safety Researcher learning resources
Videos first — they're the fastest way to get moving — then reading, hands-on practice and courses. Only resources badged professional certification award a formal credential.
Machine Learning with Python and Scikit-Learn
freeCodeCamp · 18 hours
An end-to-end introduction to model training, evaluation and practical machine learning.
Learning resource. This does not award a professional certification.
Machine Learning Crash Course
Google for Developers
A practical introduction to ML concepts with interactive visualisations, exercises and production guidance.
Certificate availability depends on the course provider.
Hugging Face LLM Course
Hugging Face
A free course covering transformers, model fine-tuning, datasets, tokenizers and sharing models.
Certificate availability depends on the course provider.
ARENA AI Safety Curriculum
ARENA
An open technical curriculum covering transformers, interpretability, reinforcement learning and alignment research.
Certificate availability depends on the course provider.
PyTorch Tutorials
PyTorch
Official tutorials from tensor fundamentals through neural networks, computer vision and distributed training.
Learning resource. This does not award a professional certification.
Transformer Circuits
Transformer Circuits
A research-oriented collection introducing circuits-based mechanistic interpretability of transformer models.
Learning resource. This does not award a professional certification.
// PRACTISE
Gain AI Safety Researcher work experience
Build practical evidence before your first role. Provider terms and eligibility can change.
GitHub · Good first issues
open-source · REMOTE
Find newcomer-labelled issues and build evidence of collaboration in public repositories.
Zindi · AI and data challenges
competition · REMOTE
Build solutions to socially and commercially relevant data problems and publish competition work.
// FIND WORK
Find AI Safety Researcher jobs
Start with “AI Safety Researcher”. Platforms without stable public search URLs open with this suggested phrase.
LinkedIn Jobs
Broad professional job search with keyword, location, experience-level and remote filters.
Wellfound
Startup and technology roles with company and compensation context.
Dice
Specialist technology roles across engineering, data, security and infrastructure.
Built In
Technology and startup job discovery, including remote and city-focused listings.
UK visa and international opportunities
Sponsorship depends on the employer, vacancy and current immigration rules. Use the linked official guidance and verify every role before applying.
Going independent
AI safety has unique entrepreneurial paths — policy advising, education, and research programmes are the most viable independent routes.
AI Safety Education and Consulting
Medium effortHelp organisations understand AI safety risks and build appropriate governance processes.
Examples
- AI governance framework design for enterprises
- Board-level AI risk briefings
- AI safety curriculum development for universities
- Policy advising for government technology programmes
Getting started
- 1. Publish accessible writing about AI safety for non-technical audiences
- 2. Engage with policy communities and think tanks working on AI regulation
- 3. Create a newsletter or blog that demonstrates expertise clearly
- 4. Partner with law firms, management consultancies, or think tanks
Independent Safety Research and Writing
High effortConduct independent research funded through grants, fellowships, or Substack-style subscriptions.
Examples
- Open Philanthropy or Long-Term Future Fund grant-funded research
- Substack publication on AI alignment for technical readers
- Independent interpretability research published on ArXiv
Getting started
- 1. Apply to EA Funds and Long-Term Future Fund for research grants
- 2. Publish research writeups on the Alignment Forum to build reputation
- 3. Apply to MATS or Redwood Research residency programmes
- 4. Build a Substack audience while employed before going independent
Communities & tools
// EARNING POTENTIAL
Earning potential
2025Salaries at top AI safety organisations are extremely competitive. Anthropic, DeepMind, and OpenAI offer some of the highest total compensation in the field.
What affects salary
- Publication record in top venues (NeurIPS, ICML) adds substantial leverage
- Interpretability and alignment research are the most valued specialisms
- PhD is often preferred but not always required at top safety labs
- Fellowship experience (MATS, Redwood) is a strong hiring signal
- Philanthropic foundations (Open Philanthropy) also fund this work — different compensation structure
🌍 US AI safety researchers at Anthropic, OpenAI, and DeepMind (US office) earn $180k–$400k+ with equity. UK labs pay well but below US levels.
Based on: Glassdoor · LinkedIn Salary · AI Safety community surveys · Anthropic and DeepMind job postings. All figures approximate.
// RESOURCES
Free resources to get started
Recommended starting points — no payment required.
- Course
- Course
- Community
- Course
- Book
Paid courses worth considering
These are not required — free resources above can get you far.
- Course
- Course
Browse all resources
Resource directory.
Free and paid resources across every tech career path — searchable and filterable.
Not sure this fits?
Take the assessment.
Answer a few short questions and get your top matches — with reasons why they fit you.
Wondering if you are ready for this path? Analyse your CV →
Related careers
Compare paths that share skills, tools or ways of working.