It’s become commonplace to hear predictions that, because of AI, software development is headed into decline as a profession. Earlier this year, the CEO of Google said that AI writes 25% of the code at Google, and the CEO of Anthropic predicted that, within 12 months, AI will write “essentially all of the code”. In July, the Federal Reserve Bank of St. Louis posted data showing that software development job postings on Indeed declined more quickly post-pandemic than many other areas. In addition, a new paper from Stanford's Digital Economy Lab, coincidentally posted at the same time as this piece, reports declines in employment for young workers in a number of AI-exposed job categories including software development.
These seem like real warning signs for the profession. Since at least 2000, the number of software developers has always risen at a much faster rate than the average increase in employment. Thus, a long-term decline in software engineering jobs would be unprecedented.
What precisely is the prediction, though? In these situations, I find it helpful to proceed like a superforecaster: (1) formulate a very precise prediction, (2) learn as much as possible about the relevant factors, paying special attention to areas of persistent uncertainty, and (3) make a precise (but revisable) Brier score prediction in the range 1 to 100 (0 ≈ No and 100 ≈ Yes).
Here is a candidate superforecasting-style prediction based on the doomsday predictions for the field:
The percent change in employment in the US in the category “Software Developers, Quality Assurance Analysts, and Testers” for the period 2023–2033 will be lower than average growth rate for all occupations in the US in that period.
This is adapted from the U.S. Bureau of Labor Statistics page for this job category. That page is currently predicting a 17% change in this period vs. an average growth rate of 4%.
I have posed this prediction to a number of very smart and knowledgeable people recently, in response to their musings about “the fall of software development”. Almost invariably, they begin to complain to me that it’s not clear how we will define this job category in the future. I reply that the prediction statement is very clear – they have to figure out how to deal with the uncertainty around what society will do with the job category.
I feel that this response gets at the heart of the issue, though: what will software development be like in the future? It does seem clear that many of today’s coding skills will become devalued or obsolete. The BLS pages for the categories Computer Programmers and Database Administrators and Architects do predict declines. However, this is a very familiar sort of technological progress (perhaps sped up by AI). The key question is what will replace these skills.
I would venture that the majority of jobs that involve working with GenAI will be placed in the category “Software Developers, Quality Assurance Analysts, and Testers”. This will include people tasked with writing prompts, shaping and orchestrating agent behaviors, and running GenAI evals.
If my guess about the job category sounds correct to you, then I predict you will not want to endorse the core prediction above. It seems so clear that GenAI is going to massively expand into essentially every industry. In that context, the prediction turns out to be true only if this category of jobs is significantly more susceptible to automation than other categories. This seems very unlikely to me because of how much of the work relates to shaping AI to align with very particular and often ephemeral organizational needs – brand identity, custom regulations, and so forth.
I would add that it seems entirely appropriate to me to expand our notion of software development to include this work. Shaping AI behavior via prompts, data annotation, and tool orchestration is very challenging, highly technical work. If you watch experienced people do the work, you see many familiar design patterns and principles from computer programming and computer systems design. The fact that their “code” is natural language quickly starts to seem incidental.
Overall, then, for the above prediction, my superforecasting Brier score is currently 10 – essentially a No. Indeed, as long as this category includes the people who design and shape GenAI products, I would predict that its growth over the next 10 years will be above 40%.
