Should I work as an AI safety engineer or an ML research engineer?

AI safety engineering and ML research engineering both sit inside frontier AI labs and both pay at the top of the engineering bands. The difference is what each role is optimizing for. Safety engineers reduce harm: catastrophic misuse, alignment failures, deceptive behavior, dual-use risks. Research engineers expand capability: new architectures, better training methods, more efficient inference.

Day-to-day work overlaps because both rely on the same skills (PyTorch, training infrastructure, evals). But the objective function is different. A safety engineer who finds a new exploit on a frontier model writes a paper about it and contributes to the next model's training. A research engineer who finds the same exploit either patches it or moves on, depending on the lab's safety posture.

Concentration of roles differs. AI safety engineering is concentrated at a small number of labs (Anthropic, OpenAI, Google DeepMind, Meta FAIR's responsible AI team) plus a few independent research orgs (Redwood Research, METR, Apollo Research). ML research engineering is much more widely distributed and exists at most large tech companies plus every frontier lab.

Compensation parity is real for senior roles. Levels.fyi data and lab disclosure indicate senior AI safety engineering pay tracks senior ML research engineering pay within 10 percent at the same lab. Both roles often include significant equity at the frontier labs.

Career risk diverges. ML research roles are more portable because the market is bigger. AI safety roles tend to be concentrated; if a lab restructures, options outside that lab are narrower. This matters less for early-career candidates and more for late-career bets.

The cybersecurity convergence is strongest in AI safety. Most AI safety teams have an in-house adversarial-evaluation function that overlaps heavily with traditional red teaming and threat modeling. Cybersecurity practitioners moving into AI safety bring directly applicable skills.