If you generate code for a living, you’ve probably asked yourself: How long until an AI takes my job? The new generation of large language models can produce text and code that rivals human performance. A paper from OpenAI warns that “80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of LLMs.” Tech workers, specifically developers, are among the most exposed. 

As these technologies continue to improve, understanding their impact on the labor market will be essential. Forecasting the trajectory of developer roles provides early insight into the future of work. We can use historical analogies and economic insights as a rough guide to answer the question on everyone’s mind: Will LLMs lead to mass displacement across developer roles?

Forecasting requires formulating a question with a clear and specific time horizon and criteria that can be easily validated. I’m interested in the impact of the current generation of LLMs, so I chose to focus on the near term: 2025. The Bureau of Labor Statistics’s (BLS) Occupational Employment and Wage Statistics (OEWS) program provides reliable job data and includes a category for “Computer and Mathematical Occupations,” under which developers fall. 

With a data source and a time horizon, I can create my question:

What will be the percentage change in Computer and Mathematical Occupations employment between 2023 and 2025, as reported by the BLS?

Before starting a forecast, it can be helpful to review the available data. 

BLS data on overall employment shows a modest upward trend between 1999 and 2021, with interruptions from the fallout of the dot-com bubble, 2008 financial crisis, and the  COVID-19 pandemic.

Tech employment has grown more robustly — on average 2.6% per year. The year-to-year changes are relatively smooth compared with overall employment, suggesting more dramatic fluctuations in tech employment are less likely.

With a baseline in mind, historical events similar to the one we’re modeling can help put the situation in context. I looked at three cases: The first, smartphone adoption, sheds light on new technology adoption in the modern era. The second and third are examples where the diffusion of new technologies upended an existing labor market: the impact of automation and outsourcing on manufacturing jobs, and bank-teller employment in the wake of ATMs.

Smartphone Adoption

Smartphone adoption provides a relatively recent example of a new disruptive technology, and it seems plausible that LLMs may follow a similar pattern. Smartphone adoption also has uncomplicated data and provides an intuitive understanding of how technological adoption usually works. Adoption trends usually follow an S-curve pattern: Uptake starts out slowly, experiences rapid growth as the technology becomes mainstream, and finally slows down as the technology approaches saturation. In 2005, almost no one had a smartphone, but in 2007, the iPhone kicked off a wave of rapid growth in the industry. By 2015, most people were carrying smartphones.

Manufacturing: Outsourcing and Automation Impacts

Manufacturing jobs in the United States expanded rapidly during World War II and reached a plateau in the late 1960s. They began falling in 1989. This was due to a combination of factors, notably higher outsourcing and greater automation (though trade policies and other factors undoubtedly played a role). Worker displacement began slowly, then rapidly accelerated throughout the late 1990s and early 2000s, finally bottoming out around 2010. Economic fluctuations complicate this picture, but the general pattern is clear: Job losses didn’t begin accelerating until almost a decade after the dawn of the outsourcing business strategy.

Bank Telling: ATM and Digital Banking Impacts

ATMs entered widespread use in the late 1970s. Although there aren’t easily accessible data sources for teller employment prior to the late 1990s, records from the BLS indicate teller employment peaked in 2007 and has experienced steady erosion ever since, despite the robust growth of the financial sector. While the work that tellers do has changed to keep up with trends, technological progress has meant that fewer are needed. The BLS expects this trend will continue.

Even if LLMs have the potential to upend the economy, it probably won’t happen tomorrow.  So how quickly do we expect the process to play out — in other words, where are we on the S-curve? To answer that question, I’ll break it down into four parts:

1) The counterfactual scenario: What job growth would we expect to see in a world without LLMs?

2) Existing growth: How much job growth can we expect in existing industries as a result of LLMs?

3) Emerging growth: How much job growth can we expect in new industries that arise as a result of LLMs?

4) Job loss: How much job displacement can we expect as a result of LLMs? 

The basic model looks like this: 

counterfactual + emerginggrowth + existinggrowth -  job_loss ¹

Along the way, I’ll also include estimates for some key variables: productivity boosts, adoption rates, and integration in business practices. 

counterfactual + existing_growth + emerging_growth -  job_loss

Before we can estimate the impact of LLMs on tech jobs, we need to project the counterfactual: How many tech jobs would be added to the economy if everything continued as normal?

4% — Extrapolating from the geometric mean of tech-job growth over the past two decades, we should expect to see growth of about 2.6% per year, leading to about 5% more roles in 2025 than today. I revised this down slightly — to 4% — due to higher interest rates, recent tech layoffs, and continuing economic uncertainty. ²

Now we’ll turn to estimating these components.

counterfactual + existing_growth + emerginggrowth -  jobloss

To get to an estimate of job growth within existing industries, I need to estimate a few other parameters. 

First, there’s developer adoption. What percentage of developers will be using LLMs for the majority of their professional programming work by 2025?

I estimate 55%. To arrive at this estimate, I used a few pieces of information: the 1.2 million people who signed up for Github Copilot during its technical preview, the percentage of those I estimate who use Copilot for business purposes (50%), and the proportion who are based in the U.S. (about 25%). ³ This would indicate that between Copilot’s release in October 2021 and the end of its technical preview in June 2022, about 3.5% of developers had adopted the technology. 

If I map this to an idealized S-curve, it would indicate we’re somewhere between 5% and 15% adoption now. (If you live near a major tech hub, that number may seem small. But tech hubs are early adopters, and many developers work in heavily regulated industries that prohibit LLM use for legal and security reasons.) If I take the midpoint of that estimate and assume we’re at 10% adoption today, then I expect, following the adoption S-curve, we’ll be in rapid growth stages by 2025: 55% is my midpoint in an estimated expected range of 35% to 75% adoption.

The second parameter is integration. After adoption takes place, businesses still need to make sense of what LLMs mean for their planning processes and decisions. I estimate this parameter by assuming that average employee tenure is a fair estimation of how long it will take business practices to change. The average tenure of computer and mathematical occupations, according to the BLS, is four years. Between now and 2025, half that tenure will elapse (50%), but I’ll adjust my estimate to 35% given that organizational changes tend to occur at a slow pace. 

The final parameter is productivity. How much more productive will LLMs make tech workers? 

I estimate 25%. Recent research from Github Next suggests LLMs allow developers to code 55% faster. Based on Stack Overflow developer surveys, as well as personal experience, I estimate the typical coder spends a little less than half their time coding (about 45%). The end result would be an approximately 25% improvement in productivity (55% * 45%). This implies that, under ideal conditions, four programmers will be able to do the work of what used to require five. If salaries remain the same, companies could expect to see a 20% decrease in costs on a per-unit basis of code.

With these in mind, we can answer our next questions.

How much will LLMs increase demand for existing tech roles by 2025?

1% — If firms can get more developer productivity for the same cost, projects that previously hadn’t made financial sense may become attractive, leading to increased job demand. 

To illustrate this, imagine company XYZ runs an e-commerce site and employs three programmers at $100,000 per year each. The company has considered building a personalized recommendation engine that they expect would increase profits by $150,000 per year but would require two additional programmers to build and maintain. At the current salary and productivity levels, this project does not make economic sense: It would cost an extra $200,000 per year, resulting in an annual loss of $50,000. But with a 25% developer productivity boost, the company would only need to hire one additional programmer (four can now do the work of five),  and the project becomes economically viable. These sorts of decisions are rarely so clear-cut, and there’s no single source of data that we could use to calculate how increased productivity might lead to more job creation, but we can do a back-of-the-envelope estimation using labor elasticity of demand.  

Economists use elasticity of demand to estimate how much demand changes in response to price changes. If the elasticity of demand for a product is -0.5, every 2% increase in the price would lead to a 1% drop in demand. Estimates for the elasticity of demand of labor range from -0.15 to -0.7, with higher-wage professions generally falling in the lower end of the range (after all, someone is still paying them despite the wage premium they’re charging). In the absence of good empirical data, let’s estimate the elasticity of demand for developers at -0.25. This means that a 20% cost reduction per unit of work should lead to a 5% demand increase in the long run (-0.25 elasticity * -20% cost).

This change will take time to play out. To capture this lag, I scale down the 5% by my values for adoption (55%) and integration (35%), which leaves us with a 1% increase in existing job roles.

counterfactual + existing_growth + emerging_growth -  job_loss

This represents the new roles created directly due to the possibilities opened by LLMs. 

How much growth in tech roles will be created by LLM-enabled industries by 2025?

3.5% — LLMs will lead to the creation of new startups and new industries. With their creation, we might see the emergence of entirely new kinds of tech jobs, such as specialized trainers who fine-tune LLMs for specific applications, or prompt engineers who specialize in designing tools to generate prompts that get improved responses. As new frontiers are unlocked, we might expect to see fierce competition over talent as companies vie to establish themselves. Machine learning is a helpful analogy here: ML-related roles now comprise about 10% of developer roles. Over time, LLM-enabled companies may end up supporting a similar 10% of the tech job market. From now until 2025, however, it is likely that only a fraction of the roles will be created. Because this category uses LLMs intrinsically, adoption will always be 100% within this category, so I only scale down according to the integration factor I calculated previously. That leaves us with 3.5% (35% integration x 10% long-run estimate).  

counterfactual + existing_growth + emerging_growth -  job_loss

Now we’ll consider the job loss from tech roles displaced by LLMs.

How much will LLMs decrease demand for non-LLM-related tech roles by 2025?

3% — Most disruptive technologies also introduce economic dislocation. While some firms will experience increased demand, those who don’t could choose to cut costs by reducing head count.   

As an example, suppose company ABC employs three programmers at an annual cost of $100,000 each. They mainly service one client and they have enough slack that it would really only take two full-time programmers and one part-time programmer to maintain the business. The problem is, they’ve never found someone who is willing to work part time reliably, so they keep three full-time developers on payroll. If their developers are each 25% more productive with an LLM coding assistant, the company might now choose to employ just two and save $100,000 per year.  

Here, the bank-teller analogy is helpful. As ATMs became more efficient at handling routine transactions, banks reduced the number of tellers and shifted the focus of the remaining tellers to more complex tasks. Similarly, as LLMs make developers more productive, firms might reduce the number of developers they employ or reassign them. 

In order to estimate the long-term net change, I’ll presume that LLMs will have less of a substitution effect for developers than ATMs did for bank tellers (at least for current-generation AI). If a technology that can substitute for 60% of what labor can do leads to a job reduction of 45% (as in the case of ATMs), ⁴ that gives us a starting point whereby each percent of substituted value constitutes a 0.75% drop in employment. 

If LLMs can substitute for 20% of developer labor, that would suggest a 15% reduction in developer jobs in the long run. Scaling down that 15% for adoption and integration leaves us with a 3% decline in jobs by 2025.

We can now make our forecast for 2025:

counterfactual (4%) + emerginggrowth (3.5%) + existinggrowth (1%) -  job_loss (3%) = 5.5% job growth for 2025

The impact of LLMs on the developer job market and the broader economy will probably be significant, but in the short term it won’t be transformative. Broader economic forces, such as recessions and interest rates, will continue to be the predominant factors shaping the overall job market, including the demand for developers. Sweeping structural changes in the industry will take time to unfold, and their full impact will not be realized by 2025. Legal and regulatory considerations surrounding the use of LLMs may also play a crucial role in shaping the speed at which these transitions take place.

At least in the short term, job loss will likely be balanced by job growth. The improved efficiency realized by LLMs will lower the barrier to entry for new start-ups, many of which will capitalize on the new opportunities enabled by advanced AI. While higher productivity will lead to some job displacement, it will also drive demand for developers, as more projects become economically viable. The next two years are more likely to see LLMs open new opportunities and allow businesses to expand and innovate. But it’s not clear how long that trend will continue.

Of course, 2025 isn’t that far away. What about in the medium term? While my predictions in this case aren’t based on a formal model, it’s worth considering how these trends may change over the next seven years. 

As LLMs become more and more capable, they will most likely encroach on tasks previously performed by humans. They may be able to automatically anticipate human needs and desires through a more comprehensive understanding of our preferences and behaviors. And they’ll probably be able to conduct A/B tests to validate their decisions and fully integrate with cloud providers to create scalable applications.

In this context, human input will still be essential, but the traditional label of “developer” might become less meaningful as job roles evolve and adapt. By 2030, I suspect that LLMs will have significantly transformed the nature of software development, blurring the lines between human and machine contributions to the process. 

And as technology advances and LLMs become more sophisticated, a larger share of developer roles may become susceptible to automation, potentially leading to a tipping point where the demand for human developers starts to decline. Despite this, there will likely still be opportunities for individuals who can adapt and find new ways to collaborate with AI-driven tools. What is clear is that the skills required for success in the developer job market of 2030 will differ significantly from those needed today.