Should I still learn to code?

Yes, and for a different reason than five years ago. The market signals conflict, so anyone giving you a confident answer in either direction is not reading the data. What follows is the conflict, and then the reasoning that survives it.

The signals point opposite ways

Against. Stanford's payroll analysis found employment among 22 to 25 year olds in highly AI-exposed occupations sitting about 19 per cent below where it would be had it tracked similarly aged workers in less-exposed occupations. Software development is among the most exposed. French data showed employment of under-30s in IT services down 7.4 per cent year on year. The divergence runs through reduced hiring rather than redundancies, which is exactly how an entry-level squeeze looks before it looks like anything.

For. In early 2026, Citadel Securities pointed to Indeed data showing demand for software engineers up 11 per cent year on year, rebutting a viral scenario piece. And the METR trial found experienced developers were 19 per cent slower with AI tools while believing they were faster, which is not the profile of a discipline about to be automated away.

Both are real. The most likely reconciliation is that overall demand is holding while the entry route narrows, which is a different problem from disappearing work and requires a different response.

Why the old reason no longer works

"Learn to code" as career advice rested on scarcity: the ability to translate intent into working syntax was rare and paid accordingly. That specific scarcity is gone. Generating plausible code is now cheap, and any advice premised on typing being the bottleneck is out of date.

What has not become cheap is knowing whether the code is right, understanding a system well enough to change it safely, and deciding what should be built. Those were always the senior parts of the job. The change is that they are now closer to the whole job.

The problem this creates

Those capabilities were previously acquired by doing the cheap parts badly for a few years. You learned to judge code by writing a great deal of it and being corrected. If the cheap parts are automated, the acquisition route is automated with them, and nobody has yet demonstrated a replacement. That is the missing rungs in its clearest single instance. It is the actual risk to a coding career rather than machines writing all the software.

Which means the useful question is not should I learn to code but how do I acquire judgement about systems when the tasks that used to build it are being done for me?

How much of this is causal

A great deal. The Stanford finding is observational and cannot establish causation, and youth hiring is sensitive to interest rates, cohort size and hiring freezes. The METR sample is 16 developers using early-2025 tooling on codebases they knew well; current models may perform very differently. Nobody has run the experiment that matters, which is whether developers who learned with heavy AI assistance become as capable as those who did not.

Anyone telling you confidently that coding is finished, or that nothing has changed, is going beyond the evidence in both directions.

What the reasoning supports

• Learn to code, but treat fluency as the floor rather than the goal. The value is in system understanding, debugging, architecture and knowing what should be built. Those take years and are not accelerated by generation.
• Do the difficult parts unaided, deliberately. Not out of purism. Because the repetitions build the judgement, and you cannot verify what you never learned to do. See desirable difficulty.
• Read far more code than you generate. The scarce skill is now evaluating a system you did not write, which is what reviewing machine output requires.
• Optimise for the second job, not the first. The entry route is narrowing, so proximity to people who will correct you matters more than title or salary. Choose the team that will teach you.
• Do not treat AI fluency as the differentiator. It is not scarce, not durable and not the constraint. See why "learn to prompt" is weak career advice.

Related SuperSkills research

On the entry-level evidence, will AI replace entry-level jobs and the missing rungs. On what appreciates, staying valuable in the age of AI. On judging output, how do I know when AI is wrong. On the discourse, the best writing on AI.

Key sources

• Brynjolfsson, E., Chandar, B. and Chen, R. (2026). Canaries in the Coal Mine? Stanford Digital Economy Lab.
• METR (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.
• INSEE (2026). Youth employment in French IT services.
• Autor, D. (2024). Applying AI to Rebuild Middle Class Jobs. NBER Working Paper 32140.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. The conflicting signals are presented rather than resolved, because they are genuinely unresolved. On a 90-day review cycle, since this is one of the fastest-moving questions on the site.