The Assumption That AI Replaces Simple Jobs First Is Wrong
For years, the dominant narrative around automation and AI tools has followed a reassuring logic for knowledge workers: machines will come for the simple, repetitive jobs first. Manual labour before professional work. Routine before skilled. That logic gave many software developers, data analysts and IT professionals a sense of security. They were wrong to feel it. Research from the World Economic Forum and independent academic work now points in a strikingly different direction — AI replacing coding jobs is a more immediate risk than AI replacing truck drivers, and the reason has nothing to do with which job is intellectually harder.
The determining factor, it turns out, is not complexity. It is data availability. The jobs most exposed to large language models (LLMs) are not necessarily the most routine. They are the most legible — the ones that have left behind a dense, searchable, structured trail of recorded work that AI systems can learn from. And few professions have documented themselves as thoroughly as software development.
What the World Economic Forum Actually Found About AI and Jobs
The World Economic Forum's white paper Jobs of Tomorrow: Large Language Models and Jobs, produced in partnership with Accenture, examined the likely impact of large language models on the labour market. The report's central insight is not that AI will neatly replace one occupation and leave another alone. Instead, exposure to AI depends on the fit between a job's specific tasks and the material that AI systems can observe, imitate and reproduce.
That framing shifts the entire debate. Instead of asking "Is this job hard enough to be safe?", the right question becomes: "Are this job's outputs visible enough for a model to learn from them?" Under that lens, software engineering is far more exposed than most people expected, while truck driving — despite being seen as a prime automation target — faces a different and in many ways harder technical barrier.
A parallel research paper from OpenAI and the University of Pennsylvania, GPTs are GPTs, reached similar conclusions. Its authors measured "exposure" as a proxy for potential economic impact — not as a guarantee of layoffs. They found that programming and writing skills were positively associated with LLM exposure, while manual task intensity and robotics exposure showed negative correlations. In short: the more your job involves producing structured language and code, the more of it a model can plausibly assist or replicate.
Why Code Became the Perfect Training Dataset for AI Tools
To understand why AI replacing coding jobs is a more immediate concern than AI replacing logistics workers, you need to understand what large language models actually learn from. These systems are trained on enormous quantities of text and structured data. Code is both. It is text, written in formal languages with strict syntax, and it is functional — it can be run, tested, linted and verified against expected behaviour in ways that a paragraph of prose cannot.
Software development has also, over decades, built a culture of radical transparency. Code is stored in version-controlled repositories. It comes with comments, documentation, unit tests, issue trackers, pull requests and complete histories of how one version became another. Much of this material is publicly accessible, especially through open-source projects. According to the OpenAI paper Evaluating Large Language Models Trained on Code, the Codex model — which later powered GitHub Copilot — was fine-tuned on publicly available code from GitHub. That single detail encapsulates the core dynamic: developers built a comprehensive public archive of their own professional practice, and that archive became the raw material for AI tools that can now perform meaningful parts of that practice.
"The jobs most exposed to large language models are not necessarily the easiest — they are the most thoroughly documented. Code is the clearest example of a profession that inadvertently trained its own replacement."
— Analysis drawn from WEF Jobs of Tomorrow reportFor an AI system, code has another critical advantage over most human work. Feedback loops are easy to construct. A model can be evaluated not just on whether its output looks like code, but on whether the program actually runs correctly. That testability accelerates training and refinement in ways that make coding a uniquely tractable problem for AI systems to improve at rapidly.
Why Autonomous Trucking Is a Harder Problem Than It Looks
Truck driving appears, on the surface, like a prime automation candidate. It is physically demanding, involves repetitive routes and is already partially assisted by GPS and fleet management software. Yet it has proven far more resistant to full automation than many predicted. The reason is not that AI is incapable of understanding roads. It is that the data required to train a safe autonomous driving system is vastly more expensive, dangerous and incomplete than the data required to train a code-completion model.
A truck driver navigating a real delivery route makes thousands of micro-judgements per hour: noticing a van drifting at the edge of peripheral vision, applying subtle brake pressure, assessing whether a warehouse gate is wide enough, deciding not to trust the satnav when local knowledge says otherwise. None of that generates a clean public archive of labelled decisions. It is embodied, situational and largely invisible to any recording system. Autonomous vehicle datasets exist, but they are expensive to collect, difficult to label at scale and dangerous to test inadequately. A coding tool that suggests a wrong function can be corrected and rerun within seconds. A heavy vehicle that makes a wrong judgement at speed cannot.
Furthermore, deploying autonomous trucking at scale is not merely an AI problem. It requires perception hardware, real-time mapping, vehicle control systems, regulatory approvals, liability frameworks and the tolerance of logistics operators, insurers and road safety authorities. The barrier is systemic, not just algorithmic.
AI Exposure Does Not Mean Immediate Job Elimination
Precision matters here. When the WEF and academic researchers describe software development as "highly exposed" to large language models, they are not predicting mass layoffs of developers by a fixed deadline. Exposure is a measure of task overlap — the degree to which portions of a job can be assisted, accelerated or replicated by current AI tools. It is not a replacement schedule.
| Job Role | LLM Exposure Level | Primary Barrier to Full Automation | Key Human Advantage |
|---|---|---|---|
| Software Developer | High | System design, judgement, responsibility | Problem framing, architecture, review |
| Technical Writer | High | Accuracy, audience understanding | Context, editorial judgement |
| Truck Driver | Low (LLMs); Medium (robotics) | Physical world, regulation, safety | Embodied judgement, adaptability |
| Data Analyst | High | Interpretation, business context | Stakeholder communication |
| IT Security Analyst | Medium | Novel threat identification | Adversarial thinking, incident response |
The more plausible near-term change for developers is a structural shift in which tasks employers value. If models can generate boilerplate code, translate between programming languages, explain common errors, draft unit tests and refactor small functions — all tasks that tools like GitHub Copilot already perform with meaningful reliability — then the scarce human skill moves upward: toward system architecture, security design, product understanding, ethical review and accountability for consequences.
There is a specific concern for junior developers worth flagging for IT managers and policy professionals. Entry-level software engineering work typically involves exactly the tasks that AI tools now handle most competently: small bug fixes, writing tests, implementing clear specifications, reading error logs. If organisations automate too aggressively at this layer, they risk hollowing out the training pipeline that produces experienced senior engineers. According to analysis by McKinsey & Company on generative AI's workforce implications, the transition period between AI augmentation and genuine structural displacement is where workforce planning decisions become most consequential.
The Open Source Irony: How Developers Built the Dataset That Trained Their Competition
There is a deeper and politically uncomfortable dimension to this story that matters particularly for the open-source community and digital sovereignty advocates. The coding profession's culture of radical transparency — version control, public repositories, Stack Overflow answers, detailed documentation, open-source licensing — was built to accelerate collaboration and knowledge sharing among humans. That culture produced some of the most valuable intellectual infrastructure in computing history. It also, inadvertently, produced the most comprehensive training dataset any AI developer could ask for.
This dynamic has already generated significant legal and ethical debate. The question of whether AI companies had the right to train on publicly licensed code without attribution or compensation was central to lawsuits filed against GitHub and Microsoft over Copilot. For European digital sovereignty advocates, the issue takes on additional dimensions: European open-source contributions fed training pipelines predominantly owned by US technology companies, raising questions about where value is created, who captures it, and what data governance frameworks under GDPR should apply to training data derived from public repositories.
The broader principle holds beyond code. Any profession that systematically turns its practice into searchable, structured, public text — legal filings, medical research papers, architectural drawings converted to digital formats — creates a more navigable surface for AI tools to learn from. This is not an argument against openness or collaboration. It is a reason why data governance, AI regulation and questions of digital sovereignty are inseparable from discussions about labour market impact.
Which Jobs Are Quietly Becoming Readable Next?
The coding-versus-trucking comparison is instructive precisely because it overturns comfortable assumptions. But it also raises a forward-looking question that deserves attention from IT decision-makers, policy professionals and anyone thinking about workforce strategy in an AI-regulation context.
The physical world is becoming more instrumented. Cameras are cheaper. Sensors are proliferating across vehicles, warehouses and logistics corridors. Every forklift movement, delivery route and depot manoeuvre is increasingly being logged, timestamped and uploaded. Autonomous driving datasets, while still expensive and incomplete relative to public code repositories, are growing in volume and quality. The gap between "jobs that have become data" and "jobs that remain embedded in physical reality" may be narrowing faster than current AI exposure analyses suggest.
