FutureJobPath logo
The career map for the AI era
This page explains how the Durability Score is built — the components, the evidence behind each one, and the named sources. For who this work fits and what a career path through it looks like, see the Deep Read. For your personalized match, take the free quiz.
Where the 52 comes from.

Three components - Automation Resistance, Structural Moat, and Demand - add up to 52.

Data note

Federal labor data does not count AI/ML engineers separately; the wage, workforce, openings, and AI-exposure numbers use Data Scientists as the public comparison. That gives useful modeling scale, but it is broader than roles that train, ship, and maintain machine-learning systems.

FJP Durability Score
52/100
Automation Resistance
16/40

Automation pressure is high because the job uses the same tools that now draft code, tests, baselines, and research summaries, while senior model judgment still matters. The exposed layer includes code, tests, experiments, and summaries; the durable layer is choosing evidence standards and owning failures after deployment.

Sub-components
Substitution Resistance
6/30

The data-scientist occupation shows high observed AI exposure, and that fits this role because much starter work is code, notebooks, experiments, and documentation. The score keeps room for human judgment where the engineer owns data strategy, failure thresholds, deployment risk, and model behavior after launch.

Sources feeding this sub-component
Anthropic labor-market impacts → Observed exposure for the Data Scientists occupation category is 46.05%.
Tufts American AI Jobs Risk Index → Median modeled job-loss pressure for the occupation category is 37.18%.
Augmentation Leverage
10/10

AI gives strong leverage here: it can draft training code, tests, documentation, analysis, and experiment ideas. The lift is valuable for someone who can review and improve the result. It is less protective for beginners whose work is mainly boilerplate around a model.

Sources feeding this sub-component
IMF Staff Discussion Notes on AI and labor markets → Links AI-related skills with wage premiums in exposed labor markets.
Structural Moat
14/35

The moat comes from hard-to-fake skill depth: statistics, software engineering, model evaluation, production systems, and evidence that models worked outside a class project. The barrier is strongest for workers who combine statistics, software engineering, data quality, evaluation, and production systems in a visible project record.

Sub-components
Physical & Environmental
1/10

The work is almost entirely screen-based. Lifting, hazardous settings, and physical presence do not protect the occupation. Any durability has to come from technical judgment, production accountability, and employer demand, not from the body being hard to replace.

Regulatory Moat
1/12

No state license controls this work. Regulated deployments can require documentation, validation, and review, but those rules create project demand rather than a legal gate to the job. The individual moat is skill and trust, not permission.

Robotics Resistance
8/8

Robotics does not drive the main substitution risk. These engineers may work on robotics models, but their own work is code, data, evaluation, and deployment. The replacement path is AI software doing more engineering support.

Sources feeding this sub-component
Credential Depth
4/5

Preparation depth is meaningful. A strong candidate usually has computer science, statistics, math, engineering, or data-science depth plus projects that show evaluation and production judgment. Graduate school matters for some research roles, but shipped proof matters across industry.

Sources feeding this sub-component
O*NET Online occupation summary → Lists this occupation in Job Zone 4, a higher-preparation category.
Demand
22/25

Demand is real but concentrated around employers investing in models, deployment, evaluation, and infrastructure, with data scientists supplying the closest federal scale. Employer interest is real around models, deployment, evaluation, and infrastructure, but the public numbers come from data science rather than this exact specialty.

Sub-components
Volume
9/10

Federal labor data does not isolate AI and machine-learning engineers. The closest data-scientist occupation has about 245,900 jobs and around 23,400 annual openings, which gives scale for nearby analytical and modeling work.

Sources feeding this sub-component
Bureau of Labor Statistics Employment Projections → Data Scientists: 245.9K jobs, 33.5% growth, and 23.4K annual openings.
Source Quality
8/8

The source base is strong for data-scientist labor data and broad AI demand signals, but weaker for exact AI-engineer headcount. Role-specific sources point toward active hiring and high pay in pockets, not a uniform market for all beginners.

Resilience
5/7

Resilience depends on staying near production value: model evaluation, data quality, monitoring, and business tradeoffs. If employers can use tools to generate acceptable baselines with fewer junior engineers, entry hiring weakens even while senior model owners stay valuable.

What would move the score
Scenario 1
Model setup becomes mostly automatic

The case weakens if teams can produce reliable baselines, tests, and deployment scaffolding with little engineering review. The trigger is employers cutting junior modeling roles while keeping only senior reviewers and production owners. That pattern would make portfolios and internships more important because classroom baselines would no longer show enough judgment.

Direction
Down
Components affected
Automation Resistance, Demand
Scenario 2
Production model ownership broadens

The case strengthens if more employers hire engineers to own evaluation, monitoring, data quality, and failure response for deployed models. That would show demand beyond research labs and make the role less dependent on a small set of elite employers.

Direction
Up
Components affected
Demand
Scenario 3
Research and product lanes diverge

A mixed outcome needs review if research-heavy roles stay strong while general product-model work compresses. The important distinction would be whether the job owns novel modeling decisions or mostly adapts available models inside products. A reader should watch whether entry postings ask for deployment and evaluation, or only routine model adaptation.

Direction
Mixed; review required
Components affected
Automation Resistance, Demand
Personalized job matches →
Want to find the careers that fit your specific profile? Take the free FJP quiz — 3 personalized matches.
Last reviewed June 2026 · Next September 2026