Blog · Research · 12 October 2026 · 13 min read

Four Numbers About Software Developers in One Story. Four Different Bases.

AI-generated. This article was written by an AI system and published without substantive human review. Facts are drawn from primary sources and linked inline — verify anything you intend to act on.EU AI Act Art. 50(4)

On 11 October 2026 CNBC published a piece asking why software developer job postings are rising if AI is automating coding. It is a reasonable question and the piece is carefully sourced. It puts four numbers side by side: postings up almost 15%, postings about 27.5% below their pre-pandemic level, the Bureau of Labor Statistics projecting 10% growth, and employment for developers aged 22 to 25 down nearly 20%.

Every one of those four reproduces exactly against the document it came from. We downloaded the underlying data files and checked. The problem is not accuracy. It is that the four numbers are measured against four different things, two of them are months-old snapshots of a series that is revised every week, and one of them is not a fall at all.

What do the postings numbers actually describe?

The 15% and the 27.5% both come from anIndeed Hiring Lab post published on 8 July 2026. That is the first thing to know: the story is from October, the figures are from July, and they describe the chart as it stood around mid-June.

Indeed publishes the series behind that post as apublic CSV, so the claim is checkable rather than arguable. We downloaded it on 11 October 2026; its last data commit was 6 October and its last observation 2 October. The software-development series, indexed so that 1 February 2020 equals 100, reads as follows.

Take 72.50 over 63.23 and you get +14.7%, which is the "almost 15%". Take 72.50 against the baseline of 100 and you get -27.5%, which is the second figure. Both reproduce to the decimal. They describe mid-June.

Run the same two calculations on the latest observation in the file and the picture moves. 78.33 over 63.23 is +23.9%. 78.33 against 100 is -21.7%. The recovery is larger than the story says and the remaining gap is smaller, because nearly four months of data arrived in between.

Two things stay true on both vintages and are worth more than either headline figure. The February 2022 peak was 233.85, so the October level is about a third of it. And the comparison series for all US postings moved very little: 108.97 in February 2025 against 100.54 in mid-June, which is the 7% fall the story mentions, recovering to 103.84 by 2 October. Software development fell much further than the market and is climbing back from much lower down. Indeed's own post makes the same point, calling the rebound's starting point low.

Indeed is careful about what the timing means, and the caution deserves repeating: its post says "correlation does not imply causation" while also calling the coincidence of dates one "that cannot be ignored". Two further figures from that post are routinely mis-added. Of the net increase in software-development postings between May 2025 and May 2026, 71% came from senior roles and 37% from jobs mentioning AI in the title. Those are shares of the net increase, not of the level, and they overlap, so they cannot be summed.

Which 10% growth, for which occupation?

The projection figure has the cleanest base problem of the four, and the BLS documents it openly.

The2025 to 2035 employment projections, released on 27 August 2026, put software developers, occupation code 15-1252, at 1,717,800 jobs in 2025 growing to 1,892,600 in 2035. That is +174,700 jobs, or +10.2%, with 95,300 annual average openings and a 2025 median wage of $135,980.

TheBLS profile page for the same occupationprints 1,905,400 jobs in 2025 and growth of 10%. The difference is not an error. The profile page covers "Software Developers, Quality Assurance Analysts, and Testers", merging in occupation code 15-1253. Add that occupation's 187,600 jobs to 1,717,800 and you get 1,905,400 exactly.

So both pages say about 10%, and they are describing populations that differ by 188,000 people. Quote the count from one and the rate from the other and you have built a number that exists nowhere.

The more interesting line in that workbook is one the story does not use. The BLS assigns every occupation a relative AI exposure category, and software developers are "Very high". The same sheet projects them growing 10.2%. Computer programmers, occupation code 15-1251, are also "Very high" and are projected to shrink 7.3%, from 110,800 jobs to 102,700. Identical exposure rating, opposite direction.

The BLS method pagepre-empts the obvious misreading, under a heading called Interpretation limits:

And the limit that matters most for a story about agentic coding tools is on the same page: the theoretical sources behind the measure "conceptualize AI capabilities as those available no later than mid-2023". The measure predates the tools. Summing the workbook ourselves across all 831 occupations, the "Very high" group covers 206 occupations and 53.2 million of 170.3 million jobs, and that group is projected to grow 2.0% while all occupations together grow 3.5%. Slower, not shrinking.

Is young developer employment down 20%, and from when?

This is the figure most likely to be quoted wrongly, including by the report that publishes it.

The Stanford AI Index 2026states on page 173, in its economy chapter highlights, that "Employment for software developers ages 22 to 25 has fallen nearly 20% from 2024." On page 221 the same report says that "By September 2025, employment for software developers ages 22-25 had fallen close to 20% from its 2022 peak."

Those are different claims. One measures from 2024, the other from a 2022 peak. Both land on about 20%, and both appear in the same document.

A third figure from page 222 is a different kind of thing again. Among workers aged 22 to 25, employment in the most AI-exposed occupations has fallen "roughly 16% relative to the least-exposed, after controlling for firm-type effects". That is a gap between two groups, not a fall in a level. A relative gap can widen while both groups grow.

All of it rests on one dataset. TheCanaries paper, in its August 2026 version, describes a balanced ADP panel of 3.5 to 5 million employees per month from January 2021 to June 2026, and reports young workers in AI-exposed occupations standing 19% below where they would be had they kept pace with less-exposed peers. Its authors call these "early, descriptive indicators" rather than causal estimates, and note the effects are "more pronounced in the ADP analysis sample than in national survey benchmarks".

The reason no one can simply check this against the official household survey is in a footnote of that paper. The monthly Current Population Sample contains between 23 and 48 software developers aged 22 to 25, and the authors note those estimates are "highly volatile, with common fluctuations of 20% or greater in estimated employment month-to-month". A 20% month-to-month wobble is the same size as the finding.

Four numbers, four bases

What about jobs that actually exist?

None of the four is a current count of employed developers, so it is worth adding one.

BLS payroll series CES6054150001 covers all employees in computer systems design and related services. That industry is not the same as the occupation, since developers work in banks and retailers too and the industry employs plenty of non-developers, but it is the closest monthly read available. It stood at 2,355,400 in September 2026, flagged preliminary, against 2,416,700 in February 2025 and 2,446,900 in January 2024. That is about 2.5% below the February 2025 level and about 3.7% below January 2024.

So postings are climbing from a deep trough while industry payrolls drift slowly down. Both are true at once, and neither tells you why.

What to take from this

First, ask what a figure is a share of before asking whether it is true. In this set that single question separates a postings index from a projection from a panel headcount, and it is why the four cannot be read as a trend.

Second, ask when the data ends, not when the article was published. Two of the four figures here describe mid-June and one describes September 2025. The weekly-revised series had moved nine points by the time the story ran.

Third, treat an exposure score as a ranking, not a prediction. The agency that publishes the score says so twice in its own spreadsheet, and puts a growing occupation and a shrinking one in the same category.

Fourth, if you write software for a living, the honest summary is unsatisfying and still better than any single number: demand for the work is recovering from a very deep fall and is nowhere near its 2022 level, the official ten-year projection is mild growth, the industry's payrolls are slightly down, and the sharpest signal of all concerns people aged 22 to 25 and comes from a single payroll panel whose authors decline to call it causal.

Methodology limits of this piece

Every figure above was read from its primary source on 11 and 12 October 2026. Five limits. The Indeed index is seasonally adjusted and re-estimated weekly, so every value we recomputed is tied to the file downloaded on 11 October 2026 with a last observation of 2 October, and the next refresh will change it; we published the vintage so the arithmetic can be redone. Indeed's sectors are assigned from normalised job titles, and a posting is an advertisement, not a hire. The CES September figure is preliminary and will be revised. The Stanford figures come from a single ADP panel whose authors describe their own results as descriptive rather than causal. And this piece makes no claim about what caused any of these movements, because none of the sources it cites makes one either.