Role Definition
| Field | Value |
|---|---|
| Job Title | AI Research Engineer |
| Seniority Level | Mid-Senior |
| Primary Function | Designs and explores novel neural network architectures, implements research papers from scratch or by extending existing frameworks, creates benchmarks and evaluation suites to rigorously test new models, and runs large-scale distributed training experiments. Bridges pure research and production engineering at frontier AI labs (Anthropic, OpenAI, DeepMind, Google, Meta) and enterprise R&D teams. |
| What This Role Is NOT | NOT an Applied AI Engineer deploying existing models to business problems. NOT an ML/AI Engineer focused on production inference pipelines and MLOps. NOT a Research Scientist writing papers without implementation responsibility. NOT an AI Safety Researcher focused on alignment and red-teaming. |
| Typical Experience | 4-10 years. Typically PhD in CS/ML or Master's with strong publication record + 3-6 years industry research experience. Deep expertise in PyTorch, distributed training (DeepSpeed, Megatron-LM), and GPU/TPU optimization. |
Seniority note: Junior (0-2 years) would score lower -- paper implementation and benchmark tasks become more automatable at junior level where the work is more reproduction than novel design. Entry-level positions at frontier labs declined 27.5% in 2025. Senior/Staff Research Engineers (10+ years) score deeper Green with more architectural novelty and strategic research direction.
Protective Principles + AI Growth Correlation
| Principle | Score (0-3) | Rationale |
|---|---|---|
| Embodied Physicality | 0 | Fully digital. Work occurs in code editors, Jupyter notebooks, cluster management consoles, and experiment tracking dashboards. |
| Deep Interpersonal Connection | 1 | Collaborates closely with research scientists, presents findings at internal reviews and conferences, mentors junior researchers. But core value is technical, not relational. |
| Goal-Setting & Moral Judgment | 3 | Decides which architectural directions to explore, designs experiments to test novel hypotheses, determines whether results are genuinely novel or artefacts. Sets research direction in ambiguous territory with no precedent. Every new architecture is a judgment call about what might work. |
| Protective Total | 4/9 | |
| AI Growth Correlation | 2 | Every advance in AI capability creates demand for the next advance. This role IS the engine of AI progress. More AI investment = more research engineers needed to push the frontier. |
Quick screen result: Protective 4 + Correlation 2 = Likely Green Zone (Accelerated). Proceed to confirm.
Task Decomposition (Agentic AI Scoring)
| Task | Time % | Score (1-5) | Weighted | Aug/Disp | Rationale |
|---|---|---|---|---|---|
| Novel architecture design & exploration | 25% | 2 | 0.50 | AUGMENTATION | AI assists with literature search, code scaffolding, and hyperparameter sweeps, but conceiving genuinely novel architectures (e.g., new attention mechanisms, mixture-of-experts variants, efficient multimodal designs) requires creative leaps and deep theoretical understanding no current AI can replicate. The human sets the research direction. |
| Paper implementation & reproduction | 20% | 3 | 0.60 | AUGMENTATION | AI code assistants (Copilot, Claude) accelerate implementation significantly -- translating mathematical notation to PyTorch, generating boilerplate, debugging training loops. But deciphering ambiguous paper descriptions, resolving discrepancies between claims and results, and adapting implementations to new contexts still requires expert judgment. This task is shifting toward AI-led with human validation. |
| Benchmark creation & evaluation design | 15% | 2 | 0.30 | AUGMENTATION | Defining what to measure and why is a judgment call. Creating benchmarks that reveal genuine capability vs gaming metrics requires deep understanding of model failure modes. AI helps generate test cases and automate evaluation pipelines, but benchmark design is fundamentally about deciding what matters. |
| Experimental design & hypothesis testing | 15% | 2 | 0.30 | NOT INVOLVED | Formulating hypotheses about why an architecture might work, designing controlled experiments to isolate variables, and interpreting ambiguous results requires scientific reasoning and domain intuition. This is the irreducible research core. |
| Large-scale distributed training & optimization | 10% | 3 | 0.30 | AUGMENTATION | AI tools increasingly automate hyperparameter tuning, learning rate scheduling, and basic distributed training setup. But debugging training instabilities at scale, optimising CUDA kernels, and troubleshooting multi-node failures across thousands of GPUs still requires deep systems expertise. Automation advancing fast here. |
| Research communication (papers, internal docs, presentations) | 10% | 3 | 0.30 | AUGMENTATION | AI drafts paper sections, generates figures, and structures arguments effectively. The human provides the insight, ensures scientific accuracy, and makes the narrative compelling. Writing quality matters less than research quality -- AI handles the former increasingly well. |
| Mentoring junior researchers & cross-team collaboration | 5% | 1 | 0.05 | NOT INVOLVED | Building research intuition in junior team members, navigating inter-team politics, and aligning research directions across groups requires deep human connection and organisational judgment. |
| Total | 100% | 2.35 |
Task Resistance Score: 6.00 - 2.35 = 3.65/5.0
Displacement/Augmentation split: 0% displacement, 80% augmentation, 20% not involved.
Reinstatement check (Acemoglu): Yes -- AI creates substantial new tasks: evaluating LLM capabilities against novel benchmarks, designing architectures for multi-agent systems, optimising training for new hardware accelerators (TPU v6, Trainium, custom ASICs), developing evaluation frameworks for emergent model behaviours, and creating safety-aware training procedures. The research frontier moves faster than automation can follow.
Evidence Score
| Dimension | Score (-2 to 2) | Evidence |
|---|---|---|
| Job Posting Trends | 2 | AI/ML postings jumped 89% in H1 2025 YoY, with ML Engineer roles up 41.8%. Generative AI skill postings surged from 55 in January 2021 to nearly 10,000 by May 2025 (Lightcast). AI research engineer is the largest ML role category at frontier labs. Every major lab (Anthropic, OpenAI, DeepMind, Google, Meta) hiring aggressively. |
| Company Actions | 2 | OpenAI's 4,000+ employees average $1.5M in stock compensation (WSJ). Anthropic, Google DeepMind, and Meta AI all expanding research teams. Analysis of 1,000 ML job postings (late 2024-early 2025) shows production ML systems and research engineering as the dominant hiring category. No lab is cutting research engineering headcount -- they are all in an "AI arms race" for frontier capability. |
| Wage Trends | 2 | OpenAI pays Research Engineers $210K-$460K base. Anthropic AI Research Engineer averages $148K base with total comp far higher. Research Scientists at top labs command median $1.56M total comp (DataExec). AI-skilled workers command a 56% wage premium, up from 25% prior year. Mid-senior AI engineering salaries jumped 9.2% in 2025 alone. |
| AI Tool Maturity | -1 | AI code assistants (Copilot, Claude, Cursor) significantly accelerate paper implementation and code writing -- the core engineering tasks. AutoML and neural architecture search tools automate portions of architecture exploration. Experiment management increasingly automated. These tools make each researcher more productive, which could reduce headcount even as output grows. The 40% of task time scoring 3 reflects real tooling pressure. |
| Expert Consensus | 2 | Universal agreement that frontier AI research is a decades-long endeavour requiring human-led innovation. WEF ranks AI/ML specialists as the fastest-growing job category globally. BLS projects 36% growth for data scientists (SOC 15-2051, which includes research roles) through 2034. The "AI arms race" between labs ensures sustained investment. No credible source predicts decline in research engineering demand. |
| Total | 7 |
Barrier Assessment
Reframed question: What prevents AI execution even when programmatically possible?
| Barrier | Score (0-2) | Rationale |
|---|---|---|
| Regulatory/Licensing | 0 | No licensing required. No regulatory mandate for human-led AI research (unlike AI deployment under EU AI Act). Research itself is unregulated. |
| Physical Presence | 0 | Fully remote capable. All work is digital -- code, compute clusters, papers. |
| Union/Collective Bargaining | 0 | Tech sector, at-will employment. No collective bargaining protections for research engineers. |
| Liability/Accountability | 1 | Moderate -- if a published architecture has flaws that cause downstream harm, reputational accountability falls on the researchers. Labs increasingly face liability for model capabilities (EU AI Act), which flows back to research decisions. But no personal legal liability for research engineering. |
| Cultural/Ethical | 2 | Strong cultural expectation that frontier research requires human creativity and scientific judgment. The AI research community values human-led discovery, peer review, and intellectual contribution. Publishing AI research authored by AI would face severe scepticism and rejection from the scientific community. Conference submissions require human authorship declarations. |
| Total | 3/10 |
AI Growth Correlation Check
Confirmed at 2. This role has the strongest possible positive correlation with AI growth. AI Research Engineers are the people who build the next generation of AI systems. More investment in AI = more researchers needed to push the frontier. The role is self-reinforcing: every AI capability advance opens new research directions (multimodal, agents, reasoning, safety) that require more researchers to explore. Unlike AI Security (which secures AI) or AI Governance (which governs AI), this role creates AI -- the most direct beneficiary of AI investment growth.
JobZone Composite Score (AIJRI)
| Input | Value |
|---|---|
| Task Resistance Score | 3.65/5.0 |
| Evidence Modifier | 1.0 + (7 x 0.04) = 1.28 |
| Barrier Modifier | 1.0 + (3 x 0.02) = 1.06 |
| Growth Modifier | 1.0 + (2 x 0.05) = 1.10 |
Raw: 3.65 x 1.28 x 1.06 x 1.10 = 5.4476
JobZone Score: (5.4476 - 0.54) / 7.93 x 100 = 61.9/100
Zone: GREEN (Green >= 48)
Sub-Label Determination
| Metric | Value |
|---|---|
| % of task time scoring 3+ | 40% |
| AI Growth Correlation | 2 |
| Sub-label | Green (Accelerated) -- Growth Correlation = 2 AND JobZone Score >= 48 |
Assessor override: None -- formula score accepted. The 61.9 sits logically between Computer & Information Research Scientist (60.2, Green Transforming) and Deep Learning Engineer (64.6, Green Accelerated). Lower than ML/AI Engineer (68.2) because 40% of task time is at score 3 (paper implementation, training optimisation, research writing) -- these are the engineering tasks that AI tools accelerate most. The research core (architecture design, experimental design, benchmark design) at score 2 keeps it firmly Green.
Assessor Commentary
Score vs Reality Check
The Green (Accelerated) label is honest and well-calibrated. The 61.9 score correctly reflects two competing forces: extremely strong demand (evidence 7/10, growth correlation +2) against moderate task-level automation pressure (40% of time at score 3). The role is not as structurally protected as AI Security Engineer (79.3) because it lacks regulatory barriers and has higher engineering automation exposure. But the demand signal is overwhelming -- every major AI lab is hiring, wages are surging 56% above market, and the "AI arms race" shows no signs of slowing. No borderline concerns; the score is 13.9 points above the Green threshold.
What the Numbers Don't Capture
- Function-spending vs people-spending. Labs are investing billions in compute, but each dollar of compute requires fewer researchers to utilise it as tooling improves. Meta trained Llama 3 with a relatively small research team. The market for AI research grows, but headcount may not scale proportionally. Research output per engineer is increasing, which could cap team sizes even as investment grows.
- Supply response is building. PhD programmes in ML have expanded dramatically. Andrew Ng's courses have millions of graduates. The supply of "good enough" research engineers is growing faster than for AI security (which requires a rare skill intersection). This will compress the extreme wage premium over 3-5 years, though demand should keep pace.
- Rate of AI capability improvement. AI tools are improving fastest in exactly the engineering tasks this role performs (code writing, paper implementation, experiment automation). The 40% at score 3 could become 50-55% within 2-3 years. The role remains Green because the creative research core is protected, but the engineering-to-research ratio will shift.
- Concentration risk. The role is heavily concentrated in a handful of labs and Big Tech companies. If AI investment slows (regulation, market correction, capability plateau), demand could contract sharply. The role's strength is also its vulnerability -- it depends entirely on continued AI investment growth.
Who Should Worry (and Who Shouldn't)
If you're designing novel architectures, leading research directions, and your name is on papers -- you're in the strongest possible position. The creative research core is genuinely irreducible. Labs will pay whatever it takes for researchers who can conceive the next breakthrough architecture.
If you're primarily implementing other people's papers and running experiments to specification -- you're in a weaker position than the label suggests. AI code assistants already handle 50-70% of implementation work. The pure "research engineering" layer (implement this paper, run these ablations, generate these plots) is exactly where AI tools are most capable. You need to move toward research direction-setting, not just execution.
The single biggest factor: Whether you set research direction or execute someone else's. The researchers who decide what to explore are deeply protected. The engineers who implement how are facing accelerating tooling pressure. At mid-senior level, most people do both -- which is why this role scores Green. But the ratio matters enormously.
What This Means
The role in 2028: The AI Research Engineer of 2028 will spend less time on implementation (AI handles most code translation from paper to working system) and more time on architecture intuition, experiment design, and evaluating emergent model behaviours. Benchmark creation becomes more important as models become harder to evaluate. The role shifts from "build the architecture" to "design the architecture and validate it works" -- a subtle but important distinction. Team sizes may shrink even as output grows, concentrating value in senior researchers.
Survival strategy:
- Move up the abstraction ladder. Shift from implementing architectures to designing them. The researcher who conceives the next attention mechanism is irreplaceable; the one who codes it up is increasingly assisted by AI.
- Master evaluation and benchmarking. As models become more capable, knowing what to test and how to measure genuine progress becomes the critical bottleneck. This is deeply human work.
- Build systems intuition at scale. Understanding distributed training failure modes, GPU cluster behaviour, and training dynamics at trillion-parameter scale is experience-based knowledge that AI cannot replicate from papers alone.
Timeline: Role strengthens for 5-10+ years, driven by continued AI investment growth. The engineering component transforms significantly (more AI-assisted), but the research component is protected as long as frontier AI capability continues to advance.
