Will AI Take Programming Jobs?

You are about to spend months learning to write code by hand, at the exact moment a machine can produce it in seconds. It would be strange not to ask whether that is a waste of your time.

So here is the honest answer, and it is not "don't worry about it".

Writing code stopped being the bottleneck

When something becomes nearly free, it stops being the constraint. Code generation is nearly free now. What did not get cheaper is deciding whether the code is right.

When production becomes free, the bottleneck moves to verification. You cannot verify what you cannot read.

Think about what a working developer now does in a day: read a diff they did not write, spot the thing that will break at 3am, notice the missing error case, decide the data model is wrong before it costs six weeks. Every one of those is a reading job, not a typing job.

So why write it by hand here?

Because hand-coding is no longer the deliverable. It is the training input that produces a reader.

Nobody is going to pay you to type for (let i = 0; ...). They will pay you because you can look at 200 lines a model produced and say "this is fine, except that will blow up on an empty array". You do not get that eye from watching. You get it from having written the thing wrong yourself a few times.

That is the whole design of this course: you build by hand first, so that when you do lean on AI, you are its editor rather than its passenger.

This is not a vibe. It has been measured.

The Anthropic randomised trial (January 2026)

Developers were asked to learn an unfamiliar Python library. One group used AI, one did not. The result:

Read that middle number again. They paid two letter grades of understanding and did not even get the speed.

But the study says something more interesting

The AI group was not one group. It split:

• Participants who fully delegated the coding got some productivity improvement, and lost the learning.
• Participants who asked conceptual questions and used AI to check their understanding scored well.

The difference was whether you were thinking alongside the tool, or letting it think for you.

That distinction is the single most useful thing in this chapter, and it is entirely under your control. Same tool, same session, opposite outcome, decided by what you type into it.

What this looks like in practice

This is exactly why Sigmo AI never hands you an answer. It asks you one question at a time until it clicks. That is not us being difficult, it is the column on the right.

The previous lesson was about you. This one is about the market.

Who is saying this

Andreessen Horowitz (a16z) is one of the largest venture capital firms in the world, managing tens of billions of dollars. They were early investors in Facebook, Airbnb, Stripe, Coinbase, GitHub, Slack, Instagram and Lyft, and they are among the heaviest investors in AI companies today.

That last part matters both ways. They have every reason to be optimistic about AI, so read them with that in mind. They also have billions of dollars riding on reading the labour market correctly, which is a strong reason to get it right.

Their piece is called "The AI Job Apocalypse Is a Complete Fantasy" (David George, May 2026). Here is the argument, broken down.

1. The "lump of labour" fallacy

The doom case quietly assumes there is a fixed amount of work to be done, so any work a machine takes is work a human loses. Human wants are not fixed. Keynes predicted automation would give us a 15-hour week; instead we invented more things to want, and more work to do.

2. Every previous time, the same shape

The spreadsheet row is the one to hold on to. A bookkeeper did arithmetic. A financial analyst decides what the numbers mean. Excel did not delete the person, it deleted the typing and promoted the judgement. That is the same trade being offered to you right now.

3. What the current AI data actually shows

Not predictions. Measurements, as of 2026:

• NBER (working paper 34984): AI adoption "has not yet led to meaningful changes in total employment". Routine clerical work is exposed to substitution; analytical and managerial work is complemented.
• Atlanta Fed: more than 90% of firms report no employment impact from AI over the last three years.
• US Census Bureau: only about 5% of AI-using firms report any headcount impact at all, and it splits both ways, roughly 2-4% reporting increases and 2-2.4% reporting decreases.
• Yale Budget Lab (April 2026): the picture "largely reflects stability, not major disruption".
• Overall: no statistically significant relationship between AI adoption and unemployment.

4. And specifically for software developers

This is the part that should matter most to you. In a16z's data, software engineering is the flagship example of augmentation, not substitution:

• Git pushes are climbing steeply.
• Software developer job creation has been accelerating since early 2025.
• SWE job postings are rising rapidly.
• Wages in computer systems design are growing above trend.
• Product manager openings are at their highest since 2022, at the same time.

That last one is the quiet proof. If AI were simply eating roles, you would not see the people who specify software and the people who build it both in demand at once. Both rising means the same team is now shipping more, and wants more hands.

On earnings calls, executives talk about AI augmenting their workforce about 8 times more often than replacing it (Goldman Sachs).

5. The honest caveats

The article is not all sunshine, and neither are we:

• Customer support hiring is falling. Some roles genuinely are being substituted. This is real, and it is happening now.
• There is some evidence that entry-level, AI-exposed roles are harder to find (Stanford, Dallas Fed, Census), even while AI-augmented entry-level roles increase. The junior door is narrower and shaped differently.
• Transitions are painful for the people inside them, even when the aggregate numbers are fine. "The economy adjusts" is cold comfort if it is your job.

Which is the actual lesson: the risk is not "no jobs". The risk is being the kind of developer AI substitutes rather than the kind it multiplies.

Where that leaves you

A developer who can only produce code competes with something that produces code for free. A developer who can read it, verify it, debug it and decide what should be built is the one holding the part that got scarcer.

That is what the next few weeks are for.