Menu
AI Safety / Alignment Engineer
Three components - Automation Resistance, Structural Moat, and Demand - add up to 50.
Federal labor data does not count AI safety or alignment engineers separately; the wage, workforce, openings, and AI-exposure numbers use Data Scientists as the public comparison. That gives a large modeling base, while actual alignment hiring is much narrower and more concentrated.
Automation resistance is moderate because research judgment remains human, while experiment scaffolding is easy to accelerate. Routine experiments, summaries, and test generation are tool-friendly; resistance comes from research taste and evidence standards for difficult model behavior.
Substitution resistance is limited for routine experiments and summaries, but stronger for research direction and evidence judgment.
Augmentation leverage is high because AI can help generate tests, write code, summarize papers, and organize results.
The moat is research credibility, math, programming, and trust in a small field rather than a protected credential. The barrier is built through math, programming, machine-learning depth, fellowships, papers, evaluations, and trusted feedback in a small research community.
Physical and environmental protection is minimal because the work is mostly computational and desk-based.
Regulatory protection is low; safety concern can raise attention, but no license reserves the job.
Robotics are not the main threat because the work is analytical and research-focused.
Credential depth is moderate through advanced study, fellowships, papers, evaluations, and strong technical portfolios.
Demand is narrow and lab-centered even though the broader data-science public row is much larger. The available labor scale comes from data science, while real alignment hiring clusters around labs, fellowships, evaluation teams, and safety nonprofits.
Volume is high only in the broader data-scientist row; actual alignment roles are much narrower.
Source quality is decent but imperfect because data scientists are a broad anchor for a specialized research path.
Resilience is fair because frontier-model risk keeps attention on the work, but funding and roles can remain concentrated.
The case weakens if labs need fewer junior researchers because AI can run routine evaluations, draft writeups, and reproduce common experiments. The title would remain important, but entry access would narrow. The training risk is losing the small tasks that help new researchers learn how evidence is built.
The case strengthens if labs, buyers, and governments require credible behavioral evidence before major model releases. That would support more roles for people who can design tests and interpret uncertain results. That would support more roles for people who can design tests, audit results, and explain uncertainty to decision-makers.
A mixed outcome needs review if a few research roles stay elite while more jobs move toward evaluation checklists and compliance support. The important question would be whether the worker sets evidence standards or only runs them. Readers should watch whether new jobs ask for original research judgment or mainly checklist execution.