Will AI replace developers? Look at who is getting hired

Short version. US software development job postings have grown by almost 15 percent since Claude Code launched in late February 2025, while postings across the whole market fell 7 percent over the same stretch. Employment for 22 to 25 year olds in the occupations most exposed to AI sits 19 percent below where it would be if it had tracked their less exposed peers.

Both of those come from payroll and postings data, not from a prediction, and holding them at once is the answer. Demand for people who can build software went up. The door you walk through to become one of those people got narrower. If you already do this job, the market wants more of you. If you are trying to get the first one, the thing you are attempting got harder for reasons that have nothing to do with you.

What follows is every number I could verify in August 2026. That includes the most quoted study on the subject, which its own authors have since updated, and one week of my own work running a studio that ships client software.

Where the hiring went

Indeed Hiring Lab indexed software development postings to 100 on 24 February 2025, the day Claude Code launched, and tracked them through June 2026. Postings grew almost 15 percent. Postings across the whole US market fell 7 percent over the same period (Indeed Hiring Lab, AI and Job Postings: From Destruction to Creation?, 8 July 2026).

That is a recovery from a deep hole rather than a boom. Software development postings remain about 27.5 percent below their February 2020 level, while postings across the market as a whole are roughly level with February 2020. The occupation fell much further than everything else and has climbed back part of the way.

The composition of the climb is the part that answers the question. Of the increase in software development postings between May 2025 and May 2026, 71 percent came from senior roles, and 37 percent came from postings with AI in the job title. The two groups overlap.

The same data set supports opposite headlines depending on the window. Measured from the May 2022 hiring peak to May 2026, the more exposed an occupation was to AI, the more its postings fell, and the relationship was statistically significant. Measured from May 2025 to May 2026, the sign flips: the more exposed the occupation, the more it rebounded. One exposure measure, two windows, two conclusions.

The study everyone quotes has been updated

In July 2025, METR published a controlled trial that became the standard citation for people arguing AI does not help. Sixteen experienced open source developers worked 246 issues drawn from repositories they had contributed to for years, averaging more than 22,000 stars and a million lines of code. Issues were randomly assigned to AI allowed or AI disallowed. Developers took 19 percent longer when AI was allowed. They had predicted a 24 percent speedup beforehand and still believed they had gained 20 percent afterwards.

That page now opens with a banner reading "These results are out of date." METR ran a continuation on late 2025 tools and published it on 24 February 2026: 57 developers, 143 repositories, more than 800 tasks, a median of 10 years of experience each.

The sign flipped. The ten developers who returned from the first study showed an 18 percent speedup, with a confidence interval running from a 38 percent speedup to a 9 percent slowdown. The 47 new recruits showed 4 percent, on an interval from a 15 percent speedup to a 9 percent slowdown. Neither interval excludes zero.

A confidence interval that crosses zero is not a finding. It is a request for more work.

The authors' own reading of their data is the most useful thing in either paper. They call the new numbers "an unreliable signal of the current productivity effect of AI tools" and the central estimate "likely a bad proxy." They list why. Developers who refuse to work without AI now decline to take part, so the sample is missing the most enthusiastic users. Pay dropped from 150 dollars an hour to 50, which made that worse. Between 30 and 50 percent of surveyed participants admitted holding back tasks they did not want to attempt unaided, so the hardest work never entered the experiment.

One line in the limitations says more than the headline number. Completion rates were lower in the arm where AI was banned, and one developer finished none of their assigned tasks without it.

So nobody has a trustworthy figure for how much faster AI makes a working developer. The best designed attempt produced opposite answers a year apart, and its authors now say the design itself is losing validity as the tools improve. Anyone quoting a clean percentage in either direction is quoting something that was withdrawn or something that was never measured.

The squeeze is at the entry point

The Stanford Digital Economy Lab paper Canaries in the Coal Mine?, revised 12 August 2026, uses ADP payroll records covering millions of US workers through June 2026. Employment for 22 to 25 year olds in the most AI-exposed occupations is 19 percent below where it would be had it kept pace with less exposed peers. Experienced workers in the same occupations show no comparable gap.

The mechanism matters more than the size. The gap runs through reduced hiring of young workers rather than through separations, and the adjustment shows up in employment rather than in base pay. Nobody is being walked out. The postings are not opening.

The caveats belong next to the number. The authors present the work as early descriptive indicators rather than causal estimates. The divergence holds when technology firms and computer occupations are excluded, and when they control for interest rate exposure and remote work. It attenuates when they control for education, some divergent trends predate generative AI, and the pattern is more pronounced in the ADP sample than in national survey benchmarks.

Set that beside the seniority split in the postings data and two independent sources agree. Payroll shows no gap for experienced workers and a 19 percent gap for the youngest. Postings show a recovery of which 71 percent is senior. The profession is hiring. It is hiring people who have already done the job.

A profession that stops hiring beginners runs out of seniors on a delay of about ten years.

What the people using it every day report

Stack Overflow's 2025 developer survey found 84 percent of respondents using or planning to use AI tools, up from 76 percent in 2024, with about 51 percent of professional developers using them daily. Adoption is close to settled.

Trust moved the other way. Among 33,244 responses, 46 percent distrust the accuracy of AI output and 33 percent trust it. Three percent highly trust it. Positive sentiment fell from over 70 percent in 2023 and 2024 to 60 percent in 2025. Developers with ten or more years of experience are the least likely to highly trust the output and the most likely to highly distrust it.

The specific complaints describe the new shape of the work. Sixty-six percent name "AI solutions that are almost right, but not quite" as a frustration. Forty-five percent say debugging AI-generated code is more time-consuming. Twenty percent say they have become less confident in their own problem solving.

Two numbers from the same survey put a floor under the replacement question. Seventy-five percent said they would still ask another person when they do not trust an AI answer. Four percent picked "I don't think I'll need help from people anymore."

The tool got adopted faster than it got trusted, and it created a category of work that did not exist at this volume before. Reading code you did not write, at speed, and being accountable for it is a different skill from writing it.

Throughput went up and stability went down

The 2025 DORA report, published 24 September 2025, surveyed nearly 5,000 technology professionals and collected more than 100 hours of interviews. Ninety percent report using AI at work. More than 80 percent believe it increased their productivity. Thirty percent report little or no trust in the code it generates, which sits oddly next to the productivity figure and is the most honest pair of numbers in the report.

The delivery findings point both ways. AI adoption showed a positive relationship with software delivery throughput and with product performance, a reversal from 2024. It continues to show a negative relationship with delivery stability.

DORA's framing of that split is worth quoting exactly: "AI doesn't fix a team; it amplifies what's already there." Teams with automated tests, mature version control and fast feedback absorb the extra volume of change. Teams without them push more changes per week into a system that cannot catch mistakes, and the report describes the outcome as instability rather than speed.

What the tools are being pointed at

Anthropic's Economic Index, published 15 September 2025, classified a million Claude.ai conversations and a million API transcripts from August 2025. Coding is 36 percent of Claude.ai usage and roughly 44 percent of API traffic, the largest category in both.

The mode of use shifted. This was the first edition where automation exceeded augmentation on Claude.ai. On the API the gap is wide: 77 percent of transcripts show automation patterns against 12 percent for augmentation, and at task level 97 percent of API traffic is automation-dominant. People are handing over whole tasks rather than sitting beside the model.

One geographic line is worth pulling out if you are reading this from India, as I am writing it. Coding is over half of all Claude usage in India, against roughly a third globally.

Automation-dominant usage is what produces the stability number DORA measured. Work arrives finished, in volume, from something that does not know whether it is correct, and somebody downstream owns the consequences.

What changed in my week

I run a small studio. I quote the work, write most of the code, and send the invoice. When people ask whether AI is replacing us, I can answer for one week of one business, which is a narrower claim than any of the studies make and a more direct one.

The hours the model took off my week were the hours I was already fastest at. First drafts of components, a migration file, a regular expression I would have looked up, the second and third variant of a layout I was going to throw away. None of those hours ever decided whether a project shipped.

# one project, from the first call to the invoice
#
# work out what they need                     a model cannot
# put a price on it in writing                 no
# agree what is out of scope                   no
# build the thing                              yes, most of it
# read the errors it produces                  partly
# decide the layout is wrong                   no
# the review call where they change the plan   no
# deploy, domain, mail, payments               partly
# the invoice, and chasing it                  no

One row changed. Everything above and below it is a conversation, a decision, or an accountability, and the pricing row is the one that decides whether the studio survives the quarter. I wrote about that separately in a post on why the quote is the hard part of the job.

The second change is that I read more code than I write now. Reviewing a diff I did not produce, quickly, while being the person who is wrong if it is wrong, is the job description shift. It is the same shift the DORA stability figure describes at team scale, and it is why learning to read an error message stopped being a beginner topic.

What I tell students now

If you are 22 and applying, the 19 percent gap is the number that concerns you. It says the first rung is thinner, it says the mechanism is hiring rather than firing, and it says nothing about your ability. Reading it as a verdict on yourself is the most common mistake I see.

The route that assumed a company would take you on and train you is no longer reliable, so the evidence has to come from you. Something deployed, at a URL, used by somebody who is not your friend. I set out the specific version of that path in a post on becoming a full-stack developer after high school, and nothing in the 2026 data changes the shape of it. The data raises the price of skipping it.

Use the model, and be able to defend every line it hands you. That 66 percent frustration figure describes the trap precisely. Code that is almost right passes a glance and fails in production, and a beginner who cannot tell the difference has automated the part of the work that was doing the teaching.

# you are using the model as a pair when:
#
#   1. you can say why the code it wrote is correct
#   2. you can name the failure case it did not handle
#   3. you would have known how to write it, slower
#   4. you read the error yourself before pasting it back
#
# Fewer than four out of four is a gap you are hiding
# from yourself, and the interview will find it.

There is one advantage to starting now that is easy to miss. The same tools that thinned the entry rung also mean a person on their own can ship a working product in their first year that would have taken a small team in 2015. The bar for evidence went up. So did your ability to clear it.

Where the forecasters disagree

Indeed Hiring Lab and Pulsenomics surveyed about 120 US economists and employment researchers in July 2026. Fifty-two percent expect AI to be at least a mild drag on employment over the next year. Thirty-five percent expect a net gain. Thirteen percent expect no effect.

Software development appears on both of their lists. Panelists naming the occupations facing the largest AI-driven employment declines put software development at the top, alongside administrative assistance. Other panelists put software development among the gainers. No single panelist named the same occupation as both, so the split runs between forecasters rather than inside them.

On productivity, 70 percent expect a modest to moderate boost over three years and 4 percent expect one large enough to reshape the economy. On pay, 57 percent expect at least slight downward pressure on the wages of degree holders over the coming year, against 34 percent for workers without a degree. Fifty-one percent expect the gap between high and low skill pay to widen, and not one panelist expected it to narrow much.

The disagreement is the finding. People whose job is forecasting employment cannot agree on the sign for this one occupation. That is a defensible place to be in August 2026, and it is worth more than any confident headline built on the same data.

So, is it replacing us

It is replacing a set of tasks. The first draft, the boilerplate, the lookup, the migration, the test scaffold. That covers a large share of the typing and a small share of the job.

Nobody has automated working out what to build, agreeing what it costs, being reachable when it breaks at two in the morning, or the call in week three where the client changes their mind. Those are the parts clients pay for, and they are the parts that were expensive before any of this shipped.

The risk worth taking seriously is structural rather than personal. A profession can keep every job it has today and still be in trouble if it stops taking people in at the bottom, and the payroll data says that is happening now. That is a hiring decision, made by people, and it can be reversed by the same people. It is not a capability of the models.

If you already have the job, the market wants more of you and your work has moved toward review. If you are trying to get the first one, arrive with proof, because fewer companies are going to hand it to you.