For two decades, earning a computer science degree was like winning a lottery ticket. Wages were flat for many, but tech professionals did well. Opportunities dried up in other sectors, but STEM careers — those in science, technology, engineering and mathematics — were a ticket to the good life. Foremost among these were software engineers. Demand for them seemed insatiable. But it was never a level playing field.
In 2022, a typical software developer earned 49 per cent more than the median wage in Canada. But a wider gap lay within the profession itself. According to salary data posted on levels.fyi, a salary aggregation website, the median compensation for a senior software engineer was 82 per cent higher than for an entry-level software engineer in Canada. While the typical senior software engineer earned just over $155,000 a year in Canada, compensation for the top 10 per cent started at $262,000. The use of AI pair programming, in which a user collaborates with artificial intelligence to develop software, may shake that up, narrowing the wage gap between non-programmers and average developers, while reinforcing the higher wages of the most skilled engineers.
The size of developer teams over the last 10 years has grown, partially due to a deluge of easy money. As AI helps engineers become more productive, and the flood of easy money recedes, tech bros may be in for a reckoning.
In a recent study by Microsoft Research and MIT Sloan School of Management, GitHub Copilot, an AI pair programming tool, improved the productivity of software developers by 55 per cent. The research involved two groups of developers. One group had access to the internet, including Stack Overflow, a question-and-answer website for programmers. The other had access to Github Copilot, an AI pair programmer that can write code based on natural language prompts from the developer.
Researchers found that developers with access to Github Copilot completed programming tasks faster, with less experienced developers benefitting more. If the dramatic productivity improvement persists in future studies, it points to a flood of programming talent by supporting “job transitions into software development” and a probable narrowing of the wage difference between software developers and other professionals.
“I believe technologies like Copilot and ChatGPT will democratize the pay structure for software engineers,” says Siva Sivakolunthu, an engineering manager based in Toronto. “More people will be able to code, so the pay gap between engineers and other professions will close.”
When ChatGPT first came out, Sivakolunthu says, he worried about software development devolving into low-wage factory work. But after he used Copilot, he says, he realized that it accelerated development and reduced the tedious bits of writing code. “If it took an hour previously, it now takes 10 minutes,” Sivakolunthu says, then he adds: “But you still need to know how to code.”
I believe technologies like Copilot and ChatGPT will democratize the pay structure for software engineers
Siva Sivakolunthu, engineering manager
Sivakolunthu envisions a broader talent pool entering software development, capitalizing on skills that are well-suited to AI pair programming. He points to the example of a former schoolteacher who is now a software engineer on his team. “He is great at asking questions and structured thinking,” says Sivakolunthu. His colleague’s teaching experience enables him to break problems down and prompt AI with clear instructions, says Sivakolunthu. The resulting productivity boost is so significant that he is now one of the most productive engineers on the team.
In another study, this one by Andres Comparo and Michele Vacarro at the MIT Center for Collective Intelligence in Cambridge, Mass., AI-powered developer tools enabled non-programmers to create websites almost as quickly as programmers. First, the researchers studied the productivity improvements by giving programmers access to Generative AI to write code. Then non-programmers were asked to perform the same tasks using the same AI tools. This is where things got really interesting: The performances of the two groups were so similar that the study indicated that they were not “statistically different.” The authors noted that this may lead to “deskilling,” the process by which skilled labour is eliminated by the introduction of technologies operated by semi- or unskilled workers. The programmers’ jobs could then be performed by people with less skill – and for lower wages.
We’ve seen this kind of “deskilling” before. At the turn of the 18th century, hand weavers and knitters were skilled craftsmen. It took years of learning to be productive, for weaving required mastery of technique and physical dexterity. These skilled workers were able to earn a decent wage, until they faced a competitor — the steam loom. Richard Guest, inventor of the power loom, wrote in A History of Cotton Manufacture: “A very good Hand Weaver, a man twenty-five or thirty years of age, will weave two pieces… (whereas), a Steam Loom Weaver, fifteen years of age, will in the same time weave seven similar pieces.”
The productivity improvements were too attractive to ignore, as factory owners could now employ younger, less-skilled workers at lower wages while dramatically increasing production. Capital poured in, and factories sprang up, first in Manchester and then across England. Industrialization benefitted the public as manufactured goods became cheaper. But there were also economic losers. Previously well-paid hand-knitters, who had spent years honing their skill, were no longer in demand. Gradually, they were replaced by more efficient machines.
“Either you’ll be a beneficiary — or a victim (of generative AI),” says Adnan Lawai, CEO of Folio3, a California-based software development company. At Lawai’s company, a small group of engineers evaluated Copilot, and initial results indicate a 30 to 40 per cent improvement. “That’s huge,” says Lawai. “If another company is 30 per cent more productive, then we’ll lose. And vice versa.”
Lawai thinks of AI as the evolution of programming. A few decades back, an engineer had to write much of the code from scratch on simple text editors, juggling between tools for version control, compiling and documentation. It was time-consuming and error-prone. Then Integrated Development Environments (IDEs) came along, dramatically improving the developer experience by bundling all of these tools. At around the same time, Application Programming Interfaces (APIs) became pervasive; now a developer could simply call an API to send a text message or process a credit-card payment without having to spend months writing code. These innovations changed the shape of global commerce. “Generative AI is a revolution of that nature,” says Lawai.
In computer science, programming languages are classified as low-level if they’re close to the hardware; it’s like speaking the computer’s native tongue. They offer precise control over hardware but demand meticulous manual coding. Hardly anyone codes in these languages anymore. The languages commonly used today are considered higher level in that they’re closer to how humans communicate. As code became easier to read and write, more people started to write code – from finance geeks writing quantitative models to public health specialists analyzing disease outbreaks.
Either you’ll be a beneficiary — or a victim (of generative AI)
Adnan Lawai, CEO of Folio3
With generative AI, the barrier to entry will get dramatically lower – and that creates new possibilities. “Business analysts think clearly and write detailed requirements,” says Sivakolunthu. “In the future, AI may convert user stories to code and BAs may evolve to become high-level programmers.” Similarly, product managers would prompt AI to create functional prototypes. It’s ultimately a question of economics: What happens when the speed of software development goes up and the cost comes down?
“It’s like a horse race,” says Anton Korinek, a professor in the Department of Economics at the University of Virginia and the Economics of AI Lead at the Centre for the Governance of AI, “and it’s hard to tell.” AI-based automation tools will increase the productivity of developers, probably reducing the demand for junior programmers. But as software becomes cheaper to create, there will be more demand for it. So, we have two scenarios. In one scenario, says Korinek, more software engineers would be needed as the increase in productivity is more than offset by the demand for more software. In the other scenario, productivity would increase so fast and by so much that it would reduce the demand for engineers. “Looking at the pace of progress,” Korinek says, “I’ll speculate that productivity will win out.”
Translation: That’s scenario 2, in which the demand for software engineers will go down because the AI is just so good.
Programmers are not the hand weavers of the modern era. Software engineering is a creative profession, and the best engineers are multitalented. They understand the business domain and collaborate across cross-functional teams to solve complex problems. These tasks can’t be automated. Additionally, the research on developer productivity improvement from AI is nascent. The programming tasks used to quantify productivity improvements in recent studies do not represent the real-world complexity of large codebases and the spaghetti of technical dependencies.
“There is more to senior devs than dev skills,” says Lawai. He says he expects senior engineers to become more valuable at the expense of entry-level programmers as AI tools can do the simple stuff that typically gets farmed out to junior developers. This points to a further concentration of compensation and wealth. Sivakolunthu envisions that “the top one per cent of engineers will work on AI and emerging tech that powers everything else.” They will have up to 30 times the impact and will make the most money. “It’s a bit like the billionaires today, where a small number control most of the wealth. In the future, that may also be the case for elite engineers.”
Even if the productivity gains are overstated (they probably are), they will still make programming available to a much wider array of professionals, especially for less sophisticated programming projects. Software developers, used to dictating terms in a tight labour market, will feel the squeeze. Just as more people, helped by an AI, will be able to write production-quality code, companies will need fewer developers.
“Complex engineering disciplines like mechanical or civil engineering, they don’t get paid as well,” says Lawai. “Maybe that’s what happens to software engineers.”
Adnan Haider is the senior vice president of analytics at a financial technology firm in Toronto and a fellow in the Fellowship in Journalism and Health Impact at the University of Toronto.
Recommended from Editorial