Where the work moved, and what is genuinely unknown

Sort by what the work was actually selling

The useful question is not “which job titles are safe”. It is: what was the customer paying for?

If the answer was largely transcription (converting a known, well-specified requirement into code of a common shape) the price fell hard. If the answer was deciding what should exist, and confirming that what exists is right, it did not.

That cut does not follow seniority or language. It runs through the middle of most job descriptions, which is why the effect feels uneven and why people in the same title report opposite experiences.

Where the price fell

Work sold mostly as transcription: brochure sites from a template, small CRUD applications to a written spec, straightforward ports between languages, first-pass boilerplate, simple scripts for non-technical clients, routine glue between two documented APIs.

Freelance marketplaces are where this shows first and fastest, because prices there move weekly rather than annually. Independent developers doing small well-specified jobs have reported rate pressure since 2023. Note what is happening mechanically: the client can now produce a mediocre version themselves for nothing, so your price competes with free rather than with the next freelancer.

The response that works is not being faster at transcription. It is selling the part that is still scarce: I will find out what you actually need, and I will make sure it is right, and I will still be here when it breaks.

Where the work expanded

Everything downstream of writing: review capacity, testing and test infrastructure, release engineering, observability, incident response, security review, data engineering and the endless work of making data trustworthy, and anyone who can hold a domain model in their head and say what the system must never do.

The pattern is the one from module one, restated at the level of a career: value pooled at the constraint, and the constraint moved downstream. If you want to know where to go, find your organisation’s longest queue and go there.

A second, quieter expansion: integration and operations for AI features themselves: evaluation harnesses, cost control, fallback behaviour, prompt and context management, safety review. This is ordinary engineering with an unusual failure distribution, and there is more of it than there are people who have done it.

What is unknown, and why

The honest answer about total employment is that nobody knows, and the data cannot currently settle it.

Entry-level technical hiring in the United States and Europe fell substantially from 2023 through 2025. That fall has at least three candidate causes running simultaneously: interest rates rose sharply, ending a decade of cheap money that funded speculative hiring; the industry had over-hired dramatically in 2021 and 2022 and was correcting; and assistants arrived. These are not separable with the data available. Any confident attribution to one cause is a position, not a finding, including the confident ones in either direction.

The two stories people tell are both under-evidenced:

• ”Demand rises because software gets cheaper.” This has happened before with other cost collapses. It is plausible. It is not a law, and it took years each time.
• ”The junior role disappears.” Also plausible. Also an extrapolation from two years of a confounded series, and it requires assuming the rung is never rebuilt.

Hold both loosely. The behaviour that pays under either is the same, which is convenient.

What to actually do about it

Move toward verification and decision. Not as a title change, as where you spend your attention. Be the person who can say whether it is right.

Get a domain. Payments, clinical records, logistics, tax, energy trading. The scarce combination is not two languages, it is one language plus knowing what a reconciliation break means. Domain knowledge lives in conversations and regulations, much of it unwritten, and that is exactly the material no corpus contains.

Stay close to the consequence. Roles insulated from whether the software worked are the exposed ones. Roles where you are woken when it breaks are not.

Do not specialise in the thing that is currently easy. If your work has been getting noticeably easier without you getting better, that is a signal, not a reward.

The access question

One good change: the capital requirement for learning this work fell close to zero. A usable assistant, a free editor, free hosting, free CI minutes and a free database are all available to somebody on a second-hand laptop with an intermittent connection, and open-weight models can be run locally on modest hardware when the connection fails.

Be honest about what still gates, though. Reliable electricity and bandwidth are not universal. Most documentation and most hiring is in English. Time zones and payment rails decide who can take which contract. The barrier moved; it did not vanish, and claiming it did is a story told by people who never faced it.