When the Machine Writes the Code: What 80% Means for Australian Software Engineers and the Universities Training Them

In June 2026, Anthropic published a figure that should be read carefully in every engineering team and every university computer science department in Australia. More than 80% of the code merged into the company's own codebase in the month of May 2026 was written by its AI, Claude. When the company's coding agent launched in early 2025, that figure sat in the low single digits.

The same report noted that the typical Anthropic engineer now ships around eight times as much code per day as they did in 2024. Not because they type faster. Because, increasingly, they do not type at all. They direct the work and review the output. One engineer quoted in the piece said it had been roughly five months since they last wrote a line of code by hand.

It would be easy to dismiss this as a single AI company describing its own unusually AI-saturated workflow. That would be a mistake. The trend Anthropic is documenting from the inside is already visible from the outside, in the labour market data, in university enrolment figures, and in the lived experience of graduates trying to enter the profession. The question this raises is not whether software engineering is changing. It plainly is. The question is whether the institutions that train software engineers, and the businesses that employ them, are changing fast enough to keep up.

This article argues they are not, and sets out what responsible adaptation would actually look like.

The frontier and the labour market are telling the same story

Anthropic's internal numbers are striking, but they are not an outlier. They are the leading edge of a pattern now appearing in serious economic research.

The most rigorous evidence comes from Stanford's Digital Economy Lab. In a widely discussed 2025 working paper, “Canaries in the Coal Mine?”, economists Erik Brynjolfsson, Bharat Chandar and Ruyu Chen analysed high-frequency payroll data from ADP, the largest payroll processor in the United States, covering millions of workers across thousands of firms. They found that since the widespread adoption of generative AI, early-career workers aged 22 to 25 in the most AI-exposed occupations experienced a 13 per cent relative decline in employment, even after controlling for firm-level shocks. For young software developers specifically, employment by mid-2025 had fallen close to 20 per cent from its late-2022 peak.

The detail that matters most for this discussion is who was not affected. Employment for more experienced workers in the very same occupations remained stable or grew. The decline was concentrated almost entirely at the entry level, and concentrated further in roles where AI automates work rather than augmenting it. The researchers were careful to note the study is observational rather than a controlled experiment, but the pattern has since been echoed in other large-scale datasets.

The mechanism is not mysterious. Entry-level engineering work has always consisted disproportionately of the well-specified, self-contained tasks that senior engineers delegate downward. Fix the broken export button. Write the tests. Clean up the legacy module. That category of work is precisely what current AI coding tools do most reliably. The first rung of the career ladder is, in effect, the rung most exposed to automation.

Australian students have already noticed

If the economics are shifting, Australian school leavers appear to have sensed it before the universities have responded.

In early 2026, the Australian Computer Society reported a notable fall in university ICT enrolments, a category that includes computer science, software engineering, cybersecurity and data analytics. This is happening despite the sector's stated ambition to grow Australia's technology workforce to 1.2 million workers by 2030, a target that declining enrolments now place under real strain. Commentary around the figures consistently points to the same anxieties: AI automating coding work, intense global competition for remote roles, and a string of high-profile tech layoffs eroding the perception that a computer science degree is a guaranteed path to a good job.

There is an irony worth naming here. The demand for genuine technology skills in Australia has not fallen. The Occupation Shortage List still flags shortages in several ICT categories, and PwC's AI Jobs Barometer has reported a substantial wage premium for workers with AI skills. The problem is not that technology careers are vanishing. It is that the entry point is narrowing at exactly the moment students are deciding whether to commit three or four years and a significant HECS debt to the field. Students are responding rationally to a signal the institutions have not yet fully absorbed.

The capability that still belongs to humans

The Anthropic report is unusually candid about where human value is migrating, and this is the part educators and employers should study most closely.

The company is explicit that its AI is now strong at producing code and at executing well-specified experiments, but that a meaningful gap persists when it comes to judgment: deciding which problems are worth solving in the first place, knowing which results to trust, and recognising when an approach is a dead end. Anthropic calls this “research taste.” It is the capability the machine is least good at, and, not coincidentally, the capability that traditional technical education is least structured to teach. Judgment is difficult to assess in a written exam. It is cultivated slowly, through exposure to real problems, good mentorship, and the experience of being wrong and understanding why.

This points to an uncomfortable misalignment. A great deal of computer science education still centres on the production of working code, the very skill that has been automated fastest, and assesses students on whether that code compiles and runs. Meanwhile the skills that are becoming the core of the job, evaluating and directing AI output, exercising judgment about correctness and quality, knowing when to override the machine, are often treated as something graduates are expected to pick up later, on their own, in the workplace that no longer offers them an entry-level position in which to learn it.

What responsible adaptation looks like

None of this is an argument that software engineering is a dead field, or that students should avoid it. The field is not disappearing; its shape is changing. The responsible response, for both universities and employers, is to take that change seriously rather than to ban the tools or pretend nothing has happened.

Some Australian institutions have begun. A number of computer science programs are placing greater emphasis on system design and architecture, the kind of higher-order, AI-resistant skills that are harder to automate, and weaving real-world projects and industry partnerships more deeply into their degrees. These are sound instincts. The broader shift they point toward involves several moves.

Teaching judgment as a discipline, not a by-product. If the durable human contribution is choosing the right problem and evaluating the result, those skills deserve explicit instruction and assessment, not incidental exposure. Code review, critical evaluation of AI output, and the ability to spot a plausible-but-wrong answer should be core competencies, not afterthoughts.

Teaching students to supervise AI, because that is the job they are entering. A graduate's first role increasingly involves directing and verifying AI-generated work rather than writing software from scratch. Knowing how to do that well, and how to do it responsibly, with appropriate scrutiny of where the AI might be confidently wrong, is now a foundational professional skill.

Building governance and oversight into technical training. As AI takes over more of the production of code, the human responsibility shifts to verification, accountability and the ability to stand behind what a system produces. That is a governance capability as much as a technical one, and it belongs in the curriculum. It is the same shift toward direction and oversight over hands-on delivery that responsible AI governance has been pointing to all along.

For employers, the parallel obligation is to resist the temptation to simply stop hiring juniors because AI can do entry-level tasks. An organisation that eliminates its entry-level pipeline today is, in effect, choosing not to have senior engineers in a decade. The talent that exercises good judgment is grown, not bought ready-made, and it is grown by giving people early-career work and the mentorship to learn from it.

The wider pattern

Software is the first knowledge profession to feel this shift at full force, but it will not be the last. It is simply the most exposed, because it is the field building the tools, and because code is unusually easy for an AI to generate and verify against an objective test. Accounting, legal services, design, financial analysis and large parts of medicine sit on the same trajectory, a little further behind. The questions Australian universities and employers are being forced to confront about software engineering are a preview of questions every knowledge-work sector will face, much as the debate about preparing the next generation is no longer confined to one classroom.

What separates the organisations and institutions that will navigate this well from those that will not is unlikely to be the speed at which they adopt AI. On that measure, as the Microsoft AI Diffusion data shows, Australia is already near the front of the global pack. The differentiator will be the quality and responsibility of that adoption: an honest reckoning with what is now automatable, a clear-eyed view of what genuinely still requires human judgment, and the governance, oversight and accountability structures to manage the gap between the two.

That work, deciding how AI should be used, where humans must remain in the loop, and how to verify and stand behind what these systems produce, is not a distraction from the real job. Increasingly, it is the real job.

For Australian businesses navigating exactly this transition, an independent standard for responsible AI use is no longer a nice-to-have. It is becoming the foundation on which the credibility of an AI-augmented workforce will rest.

Responsible AI Australia helps Australian organisations adopt artificial intelligence responsibly, with the governance, oversight and accountability that a rapidly changing workforce demands. To learn more about certification across our three tiers, get in touch.

Frequently asked questions