Hiring nowAlignment and scalable oversight

Scalable Oversight Researcher

Also called AI Oversight Scientist · Superalignment Researcher

A scalable oversight researcher designs and tests ways to supervise AI systems on tasks too hard for a person to judge easily. The work shows whether your safety methods still hold as models get more capable.

What they do

The work starts with a simple problem: a person cannot easily check whether a model's answer is right. The researcher breaks hard tasks into pieces a judge can assess, then designs protocols such as debate, recursive decomposition or assisted review. They build controlled experiments, run them in training environments, and test how reliable the judges are, whether human or model. They also attack their own protocols with adversarial evaluations to find where oversight fails. Results go into written findings that the wider alignment team can act on.

What a strong hire delivers

In the first months, a strong hire turns a vague worry about supervising harder tasks into something measurable. Expect a written oversight protocol, a controlled evaluation that can be rerun, and a clear account of where judges fail. They should also leave the team with evidence on whether error detection and correction hold up as tasks get harder, and a sensible view on what to try next.

How you can tell they are good

Ask them to walk through an oversight experiment they designed, including what failed. Strong people explain how they defined task difficulty, how they checked the judge was reliable, and what a null result would have meant. They separate a protocol that looks good from one that catches real errors. Be wary of anyone who talks only in theory or cannot describe a controlled comparison.

Why it is hard to fill

Very few people have done this work, and nearly all of them already sit inside a small number of frontier labs. Titles vary widely, so job boards rarely help. Candidates tend to come from mixed backgrounds: machine learning research, preference learning or technical safety. Few have all the experimental rigour and alignment depth at once. Expect a thin, closely watched market where approach and mission fit matter as much as the offer.

Where to find them

Look first at alignment and technical safety research teams, especially people working on debate, recursive oversight or reward modelling. Adjacent pools include preference learning researchers, model evaluation scientists and ML researchers who have moved toward safety. Many sit in frontier labs, safety-focused research groups and academic labs with strong empirical alignment programmes. Published experimental work is usually the best signal.

Indicative pay

Not enough published evidence yet to give a fair range for this role, which is common for work this new. We benchmark it for you as part of a search.

Interview questions for a Scalable Oversight Researcher

  1. Describe an experiment where you tested supervision on a task that was hard for the judge. How did you define difficulty?

    A strong answer gives a concrete difficulty measure, a baseline and a clear comparison. It covers confounds they controlled for. Worry if difficulty was assumed rather than measured, or if the result was described without any baseline.

  2. How would you check that a judge, human or model, is reliable on tasks it cannot easily solve?

    Look for ground-truth subsets, planted errors, agreement checks and tests across difficulty levels. They should name the limits of each. Worry about answers that treat judge agreement as proof of correctness.

4 more questions, with what a strong answer covers, are in the Role Kit.

Job description you can post

We are building methods to supervise AI systems on tasks that are too hard for any one person to judge easily. We are looking for a researcher to design oversight protocols, test them in controlled experiments and show us where they hold and where they fail.

What you will do

  • Design oversight protocols for tasks beyond easy human judgement
  • Build controlled evaluations that measure error detection and correction as difficulty increases
  • Test the reliability of human and model judges
  • Run adversarial evaluations against our own protocols
  • Break hard tasks into pieces a judge can assess
  • Write up findings clearly, including negative results
  • Work with alignment and training teams to apply what works

What you will bring

  • Published or internal empirical work in alignment, preference learning or model evaluation
  • Strong experimental design, with clear baselines and controls
  • Hands-on experience running experiments in training environments
  • Familiarity with debate, recursive oversight or related methods
  • Ability to reason about judge reliability and failure modes
  • Clear writing that others can act on

Nice to have

  • Experience with adversarial evaluation or red-teaming
  • Work on reward modelling or human feedback methods
  • Experience building experiment tooling

Free to use and adapt. If it saves you time, a link back to this page is appreciated.

Skills that matter

  • Alignment researchAdvancedStudy methods to make behaviour meet intended objectives
  • Preference learningAdvancedLearn behaviour from comparisons and reward signals
  • Model evaluationWorkingMeasure capability and failure modes credibly
  • Experimental designWorkingChoose comparisons that identify credible effects

Other roles in alignment and scalable oversight

EmergingReward Hacking ResearcherA Reward Hacking Researcher finds and studies cases where models exploit the reward or evaluation signal instead of doing the intended task, then…Hiring nowModel Behaviour ResearcherA Model Behaviour Researcher studies how a model's behaviour shifts across contexts and training, and shows what caused the shift. That gives the…Hiring nowModel Behaviour and Character DesignerA Model Behaviour and Character Designer defines and tunes how a model reasons, communicates and handles values trade-offs. They make the product's…

Related roles in other families

Hiring nowPost-Training Research EngineerPost-training and preference optimisationA Post-Training Research Engineer adapts pretrained models into ones that behave as the company needs. They turn preference data and reward signals…Hiring nowReward Modelling ResearcherPost-training and preference optimisationA reward modelling researcher builds models that learn what good output looks like from preference comparisons. They give post-training teams a…Hiring nowReinforcement Learning Research EngineerPost-training and preference optimisationA Reinforcement Learning Research Engineer improves how a model behaves by training it against interactive signals. They matter because reward gains…Hiring nowModel Evaluation EngineerCapability evaluation and benchmarkingA Model Evaluation Engineer builds the tests that show what an AI system can do and where it breaks. Without them, a team cannot tell whether a…

How we wrote this

Role definitions are our own editorial work, checked against occupational classifications, standards, research and employer pages. Titles vary by company, and this is a description of the work, not a job advert.

← Alignment and scalable oversightAll new job families

Start a search

Hiring a Scalable Oversight Researcher?

Tell us the work. We find the people already doing it, whatever their job title says, and you meet them already interviewed for your role.