Roles where AI output has physical consequences, and where regulators rather than firms are setting the pace of adoption.
Software Engineer
One study found AI made developers 56% faster. Another found it made them slower.
Research by Yunzhen Li · Shashank Sharma
Typical employers Large technology firms · financial services · startups · consultancies · in-house engineering teams across nearly every industry
What the job is
Software engineers design, write, test, and maintain the code behind applications, websites, and back-end systems; a backend or full-stack developer builds the server-side logic, databases, and APIs behind a product.
Two controlled studies, opposite directionsFigures are percentages. The studies measure different people on different work — a standardised task versus large, mature codebases. Data: Peng et al. (2023); Becker et al., METR (2025). Chart drawn for this page; the underlying figures are the contributor’s verified sources.
How AI is used today
AI coding assistants are standard tooling rather than novelty — GitHub Copilot, Cursor, ChatGPT, and Claude are used for code completion and generation, debugging, writing unit tests, refactoring, and generating documentation, with developers increasingly working in a pair-programming mode alongside the assistant. Adoption is close to universal: Stack Overflow's 2025 Developer Survey reports "84% of respondents are using or planning to use AI tools," up from 76% the year before, with 51% of professional developers using them daily. But the same survey found 46% distrust the accuracy of these tools against 33% who trust them — a trust gap that makes careful review a normal part of AI-assisted work rather than an extra step.
Surprisingly, we find that allowing AI actually increases completion time by 19% — AI tooling slowed developers down.
Becker et al., METR (2025)
What’s changing
The balance of the job shifts from writing every line by hand toward specifying, reviewing, and integrating code the AI drafts, making judgment and review more central than raw typing speed. How large the productivity effect actually is remains genuinely uncertain, and that uncertainty is the most interesting finding in this entry. In a randomized experiment, developers given Copilot completed a standardized task "55.8% faster than the control group" — but that study used one specific JavaScript HTTP-server assignment, and its result does not transfer automatically to every codebase. A study of experienced open-source developers working in large, mature codebases found the opposite: "we find that allowing AI actually increases completion time by 19% — AI tooling slowed developers down." Both results can be true, because value depends on the task, the engineer's familiarity with the codebase, and the time required to check the output. At the occupational level, Eloundou et al. estimate around 80% of U.S. workers will have at least 10% of their tasks affected by language models, with higher-skill knowledge work among the more exposed. Anyone offering a precise five-year prediction is claiming more certainty than the evidence supports.
What a new entrant needs
Fundamentals matter more, not less: you cannot evaluate, debug, or secure code you do not understand, and the assistant will confidently produce code that is subtly wrong. The central new skill is critical review of AI output for correctness, efficiency, and especially security — and the risk is concrete. A 47-participant study found that developers with access to an AI assistant "wrote significantly less secure code than those without access to an assistant," while being more likely to believe their code was secure. Real fluency includes knowing when not to use AI at all, which the framework names explicitly at the Fluency level. The framework's computer science guidance is specific: "Students should learn to recognize when AI-generated code may be incorrect, inefficient, or insecure." A strong junior engineer can explain what the AI produced, what evidence was used to check it, and why the final code was accepted, changed, or rejected.
Where the evidence disagrees
The METR finding is the one that should change how a new engineer thinks about entering the field: experienced developers got slower with AI while being convinced they were faster. Read optimistically, it suggests the engineers who do well will not be the ones who lean on AI hardest, but the ones with strong enough fundamentals to know when the tool is wrong.
5 sources, located and verified by the contributors
The pace here is set by regulators, not firms — and they have said no for now.
Research by David Lukowiecki
Typical employers Delta · United · British Airways · FedEx · UPS · Amazon Air
What the job is
Commercial pilots fly passengers or cargo, starting as first officers and upgrading to captain.
2Rested, competent pilots that U.S. law requires in the cockpit, as reaffirmed by Congress in 2026
Air Line Pilots Association (2026)
How AI is used today
The job is much more than hand flying: planning the route with dispatch, handling weather and fuel, briefing the crew, monitoring highly automated aircraft in cruise, talking to air traffic control, and dealing with abnormal situations. Most of what passengers think of as automation is rule-based software — autopilot, autothrottle, flight management computer — not machine learning. ML is appearing at the perimeter of the role. OpenAirlines' SkyBreathe On Board, delivered through the electronic flight bag many airlines now issue, gives in-flight advice on shortcuts and other recommendations based on the aircraft's actual performance. Training is changing too, with simulator makers allowing tablet-based procedural practice that reduces full-simulator time. Manufacturers are experimenting further: Airbus flew a DragonFly program on an A350-1000 that can divert to an alternate during an emergency in cruise, and land and assist on taxi — though none of this is on a production aircraft.
Pilots have lived with automation longer than almost any profession, which is why the evidence here is about skill decay rather than job loss. What degrades first is not hand-flying — it is knowing where you are and working out what just failed.
What’s changing
The question that dominates this role is whether aircraft will carry fewer than two pilots. The near-term answer is no, and unusually for this collection, it is a regulatory answer rather than a technological one. The European Aviation Safety Agency said in 2025 it cannot prove single-pilot operation is safe, and in 2026 Congress told the FAA that aircraft must have two rested, competent pilots. That makes it safe to say this will not happen within five years. The FAA's own roadmap is to implement AI in less safety-critical aspects of the aircraft first and work up, so a new pilot should expect AI help with decisions on fuel and weather rather than replacement. Confidence: medium, since manufacturers continue working on autonomy and many parties want two pilots retained.
What a new entrant needs
The specific skill is staying engaged when automation is flying the aircraft. A NASA study of airline pilots in a 747 simulator found that manual flying ability remained somewhat proficient under high automation use — but the abilities that require thinking while flying, such as situational awareness and diagnosing faults, depended on how much the pilot supervised the automation. That maps onto Critical Thinking and Judgement: the pilot must understand what the automation is doing. As these tools reach the electronic flight bag, Understanding AI and Data becomes relevant too — when a computer gives you information about fuel or route, you want to know where that data came from. Because regulators are adding AI slowly, a new pilot can assume continued training in basic airmanship: flying the aircraft, working with the crew, and handling irregular situations.
5 sources, located and verified by the contributor
The job is orchestrating fleets and judging whether to trust what the system decided.
Research by Emanuele Giuliani
Typical employers Exotec · Locus Robotics · inVia Robotics · Amazon robotics · large retailers, 3PLs, and manufacturers
What the job is
A robotics deployment manager sits on the operations side of the robotics boom rather than the engineering side: running the project that gets a robot fleet live in a real warehouse or factory. That means coordinating vendor teams, running site acceptance tests, managing the list of things that go wrong when a robot built on a clean demo floor meets a warehouse with uneven concrete and forklift traffic, and reporting up to both the client's operations executives and the vendor's leadership.
Net job change by 2030: the headline, and the robotics line inside itMillions of jobs, globally. Both bars are net figures from the same report. Data: World Economic Forum, Future of Jobs Report 2025. Chart drawn for this page; the underlying figures are the contributor’s verified sources.
How AI is used today
The AI content of the job is not robots doing clever things alone — it is software coordinating many robots at once. The 2025 shift in the industry was not a better robot but fleet orchestration: platforms that connect autonomous mobile robots, automated guided vehicles, conveyors, and human workers so they are not all fighting over the same aisle. If a warehouse runs robots from three vendors, no single vendor's controller sees the whole picture, and the orchestration layer is what prevents chaos. Digital twins are the other piece, modelling the warehouse before anything on the real floor is touched. Amazon's Vulcan shows how far capability has come: according to a 2026 BCG report it "can already reliably handle around 75% of the more than 1 million unique items in its fulfillment catalog — a level of generalization previously unimaginable for robotic picking systems."
Highly visible demonstrations can obscure which capabilities are mature and which remain experimental.
Boston Consulting Group (2026), warning that firms who cannot tell the difference either commit capital too early or miss the near-term gains
What’s changing
Two kinds of claims circulate here and they do not deserve equal trust. The number of robots installed each year is a countable fact: the International Federation of Robotics reports "542,000 robots installed in 2024 — more than double the number 10 years ago," expecting more than 700,000 by 2028. Predictions about what that means for jobs five years out are a different matter. A 2026 BCG report warns that "highly visible demonstrations can obscure which capabilities are mature and which remain experimental," and that companies unable to tell the difference "risk either overcommitting capital too early or waiting too long to capture near-term gains." A cool robot video does not mean the technology is ready to run your warehouse.
What a new entrant needs
Starting this job in 2020 mostly meant learning to install and configure robots that repeated the same routine. Starting now means dealing with an AI system making live decisions about where robots go and which tasks come first — and the actual job is deciding whether to trust what it just decided. That is Critical Thinking and Judgement. Understanding AI and Data matters too: a digital twin built for one warehouse will not work unchanged at a site with a different layout, and knowing where it is probably wrong is now part of the job rather than something to hand back to whoever built it.
Where the evidence disagrees
The most instructive contradiction in this entry sits inside a single report. The World Economic Forum's Future of Jobs Report 2025 projects 170 million jobs created against 92 million lost by 2030 — a net gain of 78 million. The same report says that "robotics and autonomous systems are expected to be the largest net job displacer, with a net decline of 5 million jobs" (p. 25). These are not contradictory, they measure different things: the 78 million aggregates every trend WEF tracks, while the negative 5 million isolates robotics and automation. Anyone who only ever sees the 78 million headline would have no idea that robotics specifically comes out negative in the same data. Separately, a 2026 review notes that the return on all this has not actually been established: "the high costs of implementing high-precision localization systems, AGV fleet management, and AI training models are substantial, yet their impact on operational efficiency and cost reduction remains underquantified."
4 sources, located and verified by the contributor
AI reads the inspection footage; the engineer still decides what the damage means.
Research by Ryan Zeng
Typical employers Pratt & Whitney · GE Aerospace
What the job is
Jet engine engineers design and refine components of gas turbine engines — compressors, combustors, turbines, nozzles — that must withstand extreme temperatures, pressures, and forces while meeting weight and safety requirements.
30%+Reduction in overall engine inspection time using AI-assisted borescope analysis, with report generation automated
RTX (2026)
How AI is used today
The clearest current AI use is inspection. Pratt & Whitney uses an AI-assisted borescope analysis tool that reviews video captured during internal engine inspections, automatically tracking surface wear patterns and flagging anomalies; it has been used on the V2500 and piloted on the GTF and F135 engine families, with RTX reporting inspection time cut by more than 30% and report generation automated where it previously took hours of manual formatting. A related system, Ascentia, analyses full flight sensor data and maintenance records to predict wear before it becomes detectable, reportedly reaching about 90% accuracy in early fault detection for aircraft cooling and ventilation units.
The AI here reads the inspection footage; the engineer still decides what the damage means and whether the part can be repaired or must be replaced. It can report salt when the cause is something else.
What’s changing
This role most likely will not change dramatically within five years. Employment of aerospace engineers is projected to grow 8% from 2025 to 2035, with about 3,800 openings each year. AI is already handling the time-consuming inspection and defect-detection work, which lets engineers spend more time designing rather than inspecting. Confidence should be moderate, since this reads current trends forward rather than resting on direct evidence about the occupation.
What a new entrant needs
A junior engineer needs to be able to use the borescoping tool without depending on the AI — understanding what caused damage to a turbine, and judging whether a part can be repaired or must be replaced. With AI-assisted inspection and monitoring in place, that means applying critical thinking and judgment rather than accepting the tool's call about a damaged component. The failure mode is concrete: it could report that salt needs removing when the cause is something else. These systems are still being trained on incoming samples.
3 sources, located and verified by the contributor