AI Safety Researcher
AI safety researchers study how AI systems fail, deceive, or behave unpredictably at a systemic level — beyond finding individual bugs — and design methods to make models more aligned, interpretable, and controllable. The work ranges from technical research (interpretability, alignment techniques, evaluation benchmarks) to applied safety engineering inside product teams shipping AI features.
It's a young field even globally, and Kenya doesn't yet have dedicated AI-safety labs, but demand is emerging at two ends: multinational AI labs hiring remote research contributors, and local companies needing someone who can translate safety research into practical internal policy and evaluation practice before deploying consequential AI systems.
- AI exposure
- 25 of 100, low exposure
- Hiring trend
- Growing
- Hiring rate
- 40%
- Minimum education
- Bachelor
The role
What the work is, what it pays, and what it costs you.
At a glance
- Remote friendly
- Yes
- Freelance potential
- Medium
- Freelance rate
- Ksh 5,500
- Time to senior
- 6 years
A day in the role
"Most of my week is reading, running small experiments to test a hypothesis about model behaviour, and writing up results clearly enough that an engineer three time zones away can act on them."
What it pays
Kenyan market, per month- Entry
- KES 140,000–220,000
- Mid
- KES 260,000–420,000
- Senior
- KES 450,000–750,000
The trade offs
In its favour
- Intellectually rich, high-impact work at the frontier of a genuinely important problem.
- Strong remote-work access to well-funded international labs and institutes.
Against it
- No local employer base yet — requires building an international remote career from Kenya.
- Slow, competitive hiring processes at the small number of institutions doing this work.
In practice
Complete a structured course like AI Safety Fundamentals, then pick one narrow, testable question (e.g. 'does this open model exhibit sycophancy under X conditions?') and publish a rigorous write-up — this is the standard way people break into the field without a PhD.
Progression typically runs independent researcher → junior researcher at a lab/institute → senior researcher/research lead, with increasing autonomy over research direction.
There is no established local employer base yet; the realistic near-term path is remote contribution to international labs, or a technical-advisor role with an emerging African AI governance body.
A typical week mixes reading recent papers, running small controlled experiments, and writing up findings clearly enough for both researchers and policymakers to act on.
Exposure
How much of this a machine can already do, and how that was worked out.
Where this rating sits
1,516 rated careersRated above 18% of the 1,516 careers in the catalogue, which averages 43. Inside technology the mean is 62, across 125 careers.
What the rating is made of
Share of recorded tasks- Machine does it
- 20%Software can already complete this work end to end.
- Machine assists
- 40%A person still decides, but the drafting is done for them.
- Person does it
- 40%Judgement, relationships and accountability that do not transfer.
Named task by task
Already automated
- Running large-scale evaluation experiments
- Summarising research literature
Still human
- Designing novel alignment/interpretability research methodologies
- Interpreting ambiguous or conflicting evaluation results
- Translating research findings into practical deployment safeguards
- Writing up research for publication or internal policy documents
- Advising product teams on safety trade-offs before launch
Task counts
- Tasks recorded
- 9
- Automatable now
- 1
- Still human
- 6
- Augmenting
- Literature review,Experiment execution at scale
- Creating
- Interpretability tooling,Standardised safety evaluation benchmarks
Sources
Behind the rating- Stanford HAI AI Index 2025
- WEF Future of Jobs Report 2025
Getting in
The routes into the role and what each one asks for.
What to study
8 courses- Certificate in Fashion Design and Textile TechnologyKsh 37,320a year
- Certificate in Desktop PublisherKsh 50,000a year
- Certificate in Mobile Applications and TechnologyKsh 56,420a year
- Certificate in Data Science and Artificial IntelligenceKsh 57,050a year
- Diploma in Photogrammetry and Remote SensingKsh 66,270a year
- Artisan in ICTKsh 67,189a year
- Certificate in Artificial Intelligence & CybersecurityKsh 67,189a year
- Certificate in Big DataKsh 67,189a year
How people get in
Computer Science / Mathematics degree + graduate research in ML safety
4-6 yearsHigh cost
Standard academic route, often via a Master's or PhD focused on alignment/interpretability.
ML engineer transition via independent research + published work
1-2 yearsLow cost
Experienced ML engineers building a public research portfolio (blog posts, open-source evaluation tools) can break in without a formal research degree.
Certifications
AI Safety Fundamentals Course
BlueDot ImpactKsh 03 months
Tools of the trade
TransformerLens
AI ResearchNice to haveFree
PyTorch
AI/MLRequiredFree
Inspect
AI ResearchNice to haveFree
Jupyter Notebook
DevelopmentRequiredFree
Interview preparation
4 questionsHow would you design an experiment to test whether a model is being deceptive rather than simply wrong?
TechnicalSenior
Look for rigorous experimental design thinking: controlling for confounds, defining measurable proxies for 'deception', and acknowledging the difficulty of the question.
Explain a key finding from AI alignment research in terms a non-technical executive could act on.
SituationalMid
Strong candidates translate technical nuance into a clear, non-alarmist, actionable recommendation.
What's the difference between interpretability and evaluation-based safety research?
TechnicalMid
Interpretability tries to understand a model's internal mechanisms directly; evaluation-based approaches test input/output behaviour empirically without opening the black box. Good candidates can explain when each approach is more useful.
Why might a model that performs well on safety benchmarks still behave unsafely in deployment?
BehavioralEntry
Should discuss benchmark overfitting, distribution shift between test and real-world conditions, and the limits of static evaluation.
Common misconceptions
It's mostly philosophical debate about robot uprisings.
The actual work is highly technical — statistics, machine learning research methods, and rigorous empirical evaluation of model behaviour under specific, testable conditions.
You need a PhD from a top-tier Western university to be taken seriously.
Strong independent research output (public evaluations, reproducible findings) increasingly matters as much as formal credentials in this fast-moving field.
What happens next
How the role changes from here, and where it leads.
The near term
Small but professionalising field, increasingly linked to regulatory compliance
- Government AI safety institutes (UK, US, EU) formalising evaluation standards
- Growing bridge roles between research and applied safety engineering
- What to do
- Build a public portfolio of rigorous, reproducible safety evaluations — this matters more for entry than formal credentials in a field this new.
Where pay is heading
2024 to 2030Monthly pay in Kenyan shillings, rounded to the nearest thousand. These are projections, not observations.
Growth outlook
- Net demand change
- 25
- Over
- 2025-2028
- Drivers
- Regulatory frameworks (EU AI Act and successors) requiring safety evaluation,Growing frontier-model capability raising stakes of misalignment
- Headwinds
- Field is small and hiring is slow/selective globally
Supply and demand
- Demand
- 55
- Supply pressure
- 15
- Balance
- High demand
What to learn
- Interpretability research methods
- Statistical evaluation design
- AI policy literacy
Tools worth knowing
Inspect (UK AISI)
Priority: Recommended
Model evaluation framework
TransformerLens
Priority: Recommended
Mechanistic interpretability research
Where people move next
3 recorded movesLine length under each name is the distance of the move: shorter means more of what you already do carries over. Marked lines are steps up rather than sideways.
- Ai Red Teamer
Easy45% skill overlapLateral
Applied, hands-on alternative to research-focused safety work.
- Responsible Ai Governance Lead
Moderate50% skill overlapLateral
Shifts from research to organisational policy and governance implementation.
- Artificial Intelligence Research Scientist
Moderate55% skill overlapPromotion
Broadens from safety-specific research into general AI capabilities research.
Related careers
Kenyan market notes
No dedicated local AI-safety employers yet; the realistic path for Kenyan practitioners is remote research contribution to international labs, or joining an emerging African AI policy/governance body as a technical advisor.
Further reading
This role is rated 25 out of 100 today. Save it and the app keeps that number, then tells you by how much it has moved when the record is next reviewed.