GenAI and the Software Engineering Economy: Three Variables, One Industry, Zero Consensus

Three Variables, One Industry, Zero Consensus

The global software engineering industry can be described with three variables:

N — the number of software engineers
L — the volume of code produced (a proxy for features, products, systems)
C — the market cost (what the world spends on software)

A caveat before we start: L is a deliberately crude proxy. Software economists will correctly point out that code volume is a liability, not an asset — more code can mean more bugs and more maintenance, not more value. The real output is reliable functionality delivered. Established frameworks like SPACE and DORA measure that better. But N/L/C has a virtue that rigorous models don't: it maps directly to the three numbers that executives actually track — headcount, throughput, and budget. This post uses L as shorthand for "system capability produced," not "lines written." Where that shorthand breaks down — and it does — the data will show it.

For 25 years, the story was simple: demand for software (L) grew, so the workforce (N) grew, so the market (C) grew. Revenue CAGR of ~12% outpaced workforce CAGR of ~6%. The industry generated more value per engineer every year. Enterprise software spending climbed from $74B in 2001 to $1.25 trillion projected for 2025. The developer population went from roughly 10-12 million at the turn of the millennium to 36.5 million professionals today.

Then Generative AI arrived and scrambled the equation.

Now every analyst, CEO, and VP of Engineering is trying to figure out what happens to N, L, and C when a machine can write code. The predictions range from mass displacement to exponential expansion. Most of them are wrong — not because the data is bad, but because they're solving for one variable when the system has three.

This is an attempt to solve for all three. With data.

The Boilerplate Tax: Why N Was Always Inflated

Before we can model what AI does to the equation, we need to understand what engineers actually spend their time on. The answer is unflattering.

Developers spend only 11-16% of their time writing code. The rest goes to maintenance (33-42%), communication (~20%), and fighting environment issues. This comes from Microsoft's internal "Time Warp" study and corroborating IDC research.

Of the code that does get written, 60-70% is orchestration and boilerplate — controllers that delegate, services that wrap domain objects, configuration files, CRUD endpoints. The business logic that defines a product's unique value? Roughly 10-20% of the codebase.

Component	% of Codebase	AI Suitability
Boilerplate/scaffolding	30-40%	High
CRUD & data access	20-30%	High
Glue code/routing	10-15%	High
Business logic	10-20%	Low-Medium
Tests/validation	15-25%	Medium

This means a significant portion of N has always existed to produce commodity code. Not because engineers are lazy — because systems require plumbing, and plumbing requires humans. Every previous wave of abstraction (compilers, frameworks, cloud) attacked this problem:

Transition	Boilerplate Reduction	What Happened to N
Assembly → C	~2.5x fewer lines per function point	N grew. New markets opened.
C → Java/Python	~5-15x fewer lines per function point	N grew. Web/mobile exploded.
Manual config → Spring Boot	~80% less data access boilerplate	N grew. Microservices everywhere.
Obj-C → Swift (Lyft)	75K → 22K lines (~70% reduction)	N grew. More iOS apps built.
Manual setup → create-react-app	~70% less initial boilerplate	N grew. React dominated frontend.

Every time the industry reduced the cost of producing code, total demand for software expanded faster than efficiency gains could absorb. This is textbook Jevons Paradox — and it's happened five times in software engineering history.

The question is whether GenAI triggers a sixth cycle or breaks the pattern.

The Productivity Paradox: What the Studies Actually Show

The headline numbers from AI productivity studies are seductive. They're also contradictory.

The Acceleration Studies

Study	Finding	Context
GitHub Copilot (2022)	55% faster task completion	Greenfield HTTP server, 95 devs
MIT/Princeton/Microsoft (2024)	26% more tasks per week	4,867 devs at Fortune 100 companies (RCT)
UChicago/Cursor (Late 2025)	39% more PRs merged	Tens of thousands of devs, 1,000+ orgs

The Friction Studies

Study	Finding	Context
METR (July 2025)	19% slower	16 experienced devs on familiar codebases (RCT)
McKinsey (2023)	<10% gain on complex tasks	40+ developers, internal study

Read those together. AI makes greenfield boilerplate 55% faster. It makes complex maintenance work on large codebases 19% slower for experienced developers.

The METR study drops a bomb: despite being 19% slower, participants believed they were 20% faster. That's a 39-point gap between perception and reality. Engineers feel productive while delivering less.

The Plot Twist

But the METR story has an epilogue. In February 2026, the researchers tried to replicate their own study with a larger cohort. It failed — not because of methodology, but because 30-50% of developers refused to participate in the no-AI control group. One participant compared coding without AI to "walking across a city when you're used to taking an Uber."

The partial data from those who did participate told a different story:

Cohort	Result
Returning developers (from original study)	+18% speedup (reversing the -19%)
Newly recruited developers	+4% speedup

The reversal tracks with model improvements. The original METR study tested Claude 3.5 Sonnet (~49% on SWE-bench). By early 2026, Claude 3.7 Sonnet hit 70.3% on the same benchmark — a massive leap in the model's ability to solve real-world engineering problems without human cleanup. Meanwhile, Salesforce deployed Cursor to 20,000 engineers and saw an 85% reduction in legacy code coverage time — a complex maintenance task, not greenfield.

The honest read on productivity: it depends on what you're measuring, when you measured it, and which model was doing the work. The tools are improving faster than the studies can track them.

The Adoption Paradox

84-90% of developers use AI tools
Only 29% trust AI accuracy (down from 40%)
66% spend more time fixing "almost-right" AI code
Code review times up 91% in heavy-adoption teams
Delivery stability down 7.2% in organizations with aggressive AI adoption

Spotify's best engineers reportedly haven't written a single line of code since December 2025, using an internal system powered by Claude Code. Meanwhile, the industry average shows PR volume up 98% but code churn doubling — meaning much of the new code gets rewritten within two weeks.

The productivity story isn't "AI makes developers faster." It's "AI makes certain tasks faster while creating new bottlenecks in verification and review." The constraint has shifted from writing code to trusting code.

The "Meatbag" Error Rates

Here's the argument that gets deployed against AI code generation: "AI hallucinates." The implication is that human code is reliable and AI code is not.

Let's check that assumption against the data. Because it's not like meatbags are right every time — especially the less experienced ones.

The Human Baseline

Context	Defect Rate
Industry average (development phase)	15-50 bugs per KLOC
Microsoft applications (internal testing)	~10-20 bugs per KLOC
Cleanroom development (high rigor)	~3 bugs per KLOC
Top-quartile vs bottom-quartile developers	10-20x productivity and quality gap

Humans inject somewhere between 15 and 50 bugs per thousand lines of code during development. That's the baseline. The industry has spent decades building review processes, testing frameworks, and CI/CD pipelines to catch them before production.

The AI Baseline

Metric	AI vs Human	Source
Total issues per PR	1.7x more (10.83 vs 6.45)	CodeRabbit, Dec 2025
Logic errors	75% more	CodeRabbit
Security vulnerabilities	1.5x-2x more	CodeRabbit, Veracode
Readability issues	3x higher	CodeRabbit
SWE-bench success rate	~49% (failure > 50%)	Anthropic benchmark
Security flaws in AI code	45% of samples contain vulns	Veracode 2025
Java AI code security failure	72%	Veracode 2025

AI-generated code is measurably buggier — at least in these studies. 1.7x more defects per pull request, 75% more logic errors, up to 2x more security vulnerabilities. SonarSource's 2025 analysis is even grimmer: 90% of AI-generated code contains "code smells" — indicators of poor design and high cognitive complexity that don't cause immediate bugs but make future modifications exponentially harder. Refactoring activity has dropped over 60% while code duplication has spiked 8x. The AI is writing code that works today and creates maintenance liabilities tomorrow.

But the CodeRabbit data has problems. The study analyzed open-source PRs without controlling for developer seniority. Market data shows junior developers lean heavily on AI for complex logic they don't understand, while seniors use it for boilerplate. The "1.7x defect rate" likely reflects misuse by inexperienced engineers more than an inherent flaw in AI output.

A GitHub controlled experiment tells a different story. With 202 developers (all 5+ years experience), those using Copilot were 53.2% more likely to pass all unit tests and had 13.6% fewer readability errors — with blind reviewers rating the AI-assisted code higher on reliability and maintainability.

Google's DORA 2025 report resolves the contradiction: AI is an amplifier. High-performing teams with strong testing, CI/CD, and review practices produce better code with AI. Struggling teams produce worse code faster. The defect rate is a function of organizational maturity, not just the tool.

And the models are moving fast. SWE-bench scores — the benchmark for solving real GitHub issues — jumped from ~49% (Claude 3.5 Sonnet) to 70.3% (Claude 3.7 Sonnet) in under a year. GPT-4.1 reduced "unnecessary edits" from 9% to 2%. The defect data from 2025 studies was collected on models that are already obsolete.

Here's the part that never makes the headlines: human developers using AI write less secure code but believe it's more secure. That's from a Stanford study. The "hallucination" isn't just in the model — it's in the human's confidence about the output.

The Economics of Bugs

Detection Stage	Cost per Bug
Requirements/Design	~$100
Coding/Unit Test	~$1,000
QA/System Testing	~$1,500-2,000
Production	$10,000-100,000+

The cost of a bug rises 100x from design to production. AI accelerates injection (more code, faster) without proportionally accelerating detection. If review processes don't scale with generation speed, defects leak downstream into the expensive end of the curve.

Code churn has doubled. Copy/paste up 48%. Refactoring down 40%. The data suggests AI encourages adding code rather than improving existing code — inflating codebases without proportionally increasing value.

There's a twist in the TCO data, though. When AI is aimed at cleaning existing code rather than generating new code, the economics flip. A global bank used multi-agent systems to refactor 15 legacy microservices and achieved a 60% reduction in refactoring time, 30% lower maintenance costs, and $2.3 million in labor savings. The pattern: AI as a maintenance tool pays for itself. AI as a generation engine creates debt.

The honest read: AI code quality is worse than human code quality by measurable margins. But human code quality was never great to begin with. The question isn't "is AI perfect?" — it's "does the speed/volume gain offset the quality/maintenance cost?" And the answer depends entirely on what you're building, who's reviewing it, and whether you're using AI to generate or to maintain.

Four Economic Scenarios

Using the N/L/C framework, here are the four possible futures — and what the data supports for each.

Scenario A: Displacement

Thesis: AI replaces a significant percentage of developer tasks. Companies maintain output, reduce headcount.

The model: N shrinks. L stays flat. C drops (or shifts to AI compute costs).

Evidence for:

Junior hiring collapsed 73% year-over-year (Ravio, 2025-2026)
70% of hiring managers believe AI can do intern-level work
Entry-level roles now require 3-5 years experience ("experience inflation")
CS graduate unemployment at 6.1% — higher than philosophy (3.2%) or art history (3.0%)
ADP Research: the US employed fewer software developers in 2024 than in 2018 — before agentic AI even matured
Average SWE salary down 18.79% from 2023 levels
Mid-level displacement accelerating: employment for 22-25 year-old engineers in AI-exposed roles down ~20% (Stanford)

Evidence against:

Senior roles are expanding, not contracting
AI specialists command a 70% wage premium ($450K+ packages)
Total developer population still projected to grow to 45M by 2030

Probability: High in the short term (2024-2027), particularly for junior and mid-level positions. The ADP data suggests this may already be structural, not cyclical.

Scenario B: Augmentation

Thesis: Same number of developers, dramatically more output per person.

The model: N stays flat. L grows 3-5x. C grows (more features shipped, more value created).

Evidence for:

26% more tasks completed (MIT/Microsoft RCT)
39% more PRs merged (UChicago/Cursor)
87% of developers report AI preserves mental effort on repetitive tasks

Evidence against:

Review times up 91% — the bottleneck shifted, didn't disappear
Delivery stability down 7.2%
Net productivity for experienced developers: possibly negative (METR: -19%)

Probability: Moderate. The "productivity paradox" — faster generation offset by slower verification — throttles the pure augmentation story.

Scenario C: Expansion (Jevons Paradox)

Thesis: Cheaper code production unlocks massive latent demand. The market grows far beyond current size.

The model: N grows. L explodes. C expands dramatically.

Evidence for:

Every previous abstraction layer (compilers, frameworks, cloud) expanded the developer population
58% of SMBs now using GenAI (up from 23% in 2023)
Custom software market growing at 22.6% CAGR (2025-2030)
25% of Y Combinator startups have 95%+ AI-generated codebases
Low-code developer population growing at 40.4% CAGR

Evidence against:

Demand for code is elastic, but demand for software engineering (the full process) is constrained by human-speed activities: meetings, requirements, compliance, architecture decisions
The "maintenance wall" — generating code is cheap, maintaining it is not. The TCO data is sobering: first-year costs for AI-heavy projects run 12% higher than traditional development when you account for review overhead, testing burden, and defect remediation. By year two, maintenance costs can spiral to 4x traditional levels as "subprime code" — syntactically correct but architecturally brittle — accumulates interest. Gartner projects 40% of agentic AI coding projects will be abandoned by 2027.
Jevons has failed before: agriculture went from 40% to 2% of the US workforce despite massive productivity gains. Food demand is inelastic. Software demand might be too.
Telephone switchboard operators: call volume exploded (Jevons in consumption) but operator employment dropped 80%. The labor decoupled from the demand.
Acemoglu's "so-so technologies" framework: if AI displaces labor without lowering costs enough to trigger massive new demand, the Jevons rebound never arrives
Baumol's Cost Disease: as writing code approaches zero cost, verifying code becomes the bottleneck. Human verification doesn't scale — creating a hard constraint on the expansion cycle.

Probability: Historically the default — but the counter-evidence is stronger this time than in any previous abstraction wave. The outcome depends on whether software demand is truly elastic (like energy) or approaching saturation (like food).

Scenario D: Transformation

Thesis: The definition of "software engineer" fundamentally changes. New roles emerge. Total N is stable or growing, but the composition shifts dramatically.

The model: N transforms. L grows. C grows and restructures.

Evidence for:

Gartner: By 2030, 25% AI-only, 75% AI-augmented, 0% human-only IT work
New roles emerging: Vibe Coder ($110K-$240K), AI Orchestrator ($120K-$180K), AI Engineer ($134K-$185K, up to $400K+)
Spotify engineers managing AI agents, writing zero manual code
The value chain shifting from syntax → architecture → verification → orchestration

Evidence against:

The pipeline problem — entry-level hiring collapsed 73%, and if juniors can't enter the pipeline, where do future seniors come from?
Citrini Research models a "2028 Global Intelligence Crisis" — a negative feedback loop where AI displaces high-income workers, reducing aggregate demand, triggering further automation. Their model projects 10.2% unemployment by 2028.
The "architecture premium" may not be permanent: agentic AI is moving into system design, dependency mapping, and modernization — tasks previously reserved for principal engineers

Probability: Very high across all timeframes. This is the scenario most supported by current data — but the Citrini bear case deserves attention. If the displacement-demand spiral activates, Scenario D collapses into Scenario A.

The Verdict

Scenario D is happening now, carrying elements of A (short-term) and C (long-term). But the bear case — A dominating permanently — has more supporting data than the industry wants to admit.

The labor market isn't simply shrinking or growing. It's polarizing. The "middle class" of coding — writing boilerplate Java/Python — is being hollowed out. The cost of code is trending toward zero. The cost of trust and architecture is trending upward — for now.

The ADP data showing fewer developers in 2024 than 2018 should give pause to anyone banking on Jevons Paradox. Agriculture, telephony, and manufacturing all saw massive productivity gains that eventually decimated the workforce in those specific professions, even as the broader economy grew. Software engineering may not be exempt from this pattern.

Scenario	Short-Term Probability	Long-Term Probability	Wage Impact
A: Displacement	High	Low	Collapse (Junior) / Stable (Senior)
B: Augmentation	Moderate	Moderate	Moderate growth
C: Expansion	Low	High	High growth (volume)
D: Transformation	Very High	Very High	K-shaped (polarized)

What This Data Doesn't Tell You

Intellectual honesty section, because anyone selling you certainty about AI's impact on software engineering is either lying or hasn't read the studies.

The productivity studies are small and contextual. The METR study (19% slower) had 16 participants — and its own replication attempt reversed the finding. The GitHub Copilot study (55% faster) measured a greenfield task, not maintenance. The UChicago study (39% more PRs) doesn't measure whether those PRs delivered more value. And Salesforce's 85% improvement is a single team on a single legacy system. Every study is a snapshot.

"Lines of code" is a terrible proxy for value. PR volume up 98% and code churn doubling could mean "teams are shipping more" or "teams are generating disposable code." GitClear counts rapid iteration as "churn" — but in agentic workflows, a developer asking an AI to "scaffold with Stripe instead of PayPal" generates high churn that represents successful iteration, not failure. The metric was designed for human workflows.

The error rate comparisons aren't apples-to-apples. The CodeRabbit study (1.7x defects) didn't control for developer seniority. The GitHub controlled experiment (53.2% more likely to pass tests) did. Google's DORA report says AI amplifies whatever's already there — good teams get better, struggling teams get worse. The defect story is about organizational maturity as much as tool capability.

The models are a moving target. Most negative studies tested Claude 3.5 Sonnet (~49% SWE-bench) or GPT-4o. Current models hit 70%+ on the same benchmark. Extrapolating defect rates from 2025 models to 2026 models is like judging the iPhone by the Blackberry.

The Jevons Paradox isn't guaranteed. Every previous abstraction wave expanded the developer population. But agriculture, telephony, and manufacturing all saw permanent workforce shrinkage despite massive productivity gains. The difference was demand elasticity. Nobody needed 20x more food. The question is whether the world needs 20x more software — or whether we're approaching saturation. ADP data showing fewer developers in 2024 than 2018 suggests the optimistic view needs defending.

The "architecture premium" may have an expiration date. Today, senior engineers who design systems and verify AI output command the highest premiums. But agentic AI is already moving into system design, dependency mapping, and migration — tasks previously reserved for principal engineers. The premium is real today. Whether it's permanent is an open question.

The bear case is credible. Citrini Research models an "Intelligence Displacement Spiral" — AI displaces high-income workers, aggregate demand falls, companies automate further. Their projection: 10.2% unemployment by 2028. Forrester projects 6% structural job loss by 2030 (10.4 million roles). Goldman Sachs says AI added "basically zero" to GDP in 2025 despite massive capital expenditure — suggesting companies are using AI for cost-cutting (substitution), not growth (expansion).

The bull case is also credible. McKinsey projects $2.6-4.4 trillion in annual value creation. Custom software spending is growing at 22.6% CAGR. 58% of SMBs are using GenAI, up from 23% in 2023. The demand expansion may just be starting.

Analyst projections are guesswork with footnotes. The track record of Goldman Sachs, McKinsey, Gartner, and Forrester on technology predictions is worse than their confidence suggests. Gartner called the dot-com bust — but also predicted $100 billion/year in IP losses from 3D printing by 2018 (actual: negligible) and that 20% of businesses would own no IT assets by 2012 (they didn't). Goldman coined "BRICs" and nailed China/India growth — then closed the BRIC fund after assets dropped 88%. McKinsey's 2011 Big Data predictions showed only 30% value capture by 2016; they admitted to underestimating organizational barriers. Forrester's 3.3 million offshoring jobs prediction remains debated two decades later. A 2024 Economist analysis found only about a fifth of technologies follow Gartner's famous Hype Cycle path. The pattern: these firms reliably identify which technologies matter but are poor at predicting when mass adoption arrives and how many jobs are actually affected. Goldman's headline "300 million jobs exposed to AI" is an estimate of exposure, not a forecast of unemployment — a distinction that gets lost in the press release. Use the ranges, not the point estimates.

The Solow Paradox applies here too. Goldman Sachs' chief economist noted AI added "basically zero" to GDP in 2025 despite massive investment. This mirrors the "you can see the computer age everywhere but in the productivity statistics" pattern from the 1980s-90s. McKinsey's own Big Data retrospective admitted that widespread value capture typically takes 5-10 years longer than predicted because organizational inertia — culture, processes, incentive structures — moves slower than technology. The current AI employment projections conflate technical feasibility (what AI can do) with process integration (what companies actually do).

We don't know what "experienced developer + AI agents" looks like at scale. Spotify's engineers writing zero code is one data point at one company. It might be the future. It might be a bubble. The METR replication failure — developers refusing to work without AI — suggests the trajectory, but not the destination.

The Strategic Read

Three things the data tells us with reasonable confidence:

1. The boilerplate economy is over. AI generates 46% of code in active repositories (61% in Java). The 60-70% of codebases that were commodity plumbing are being automated. This isn't coming — it's here. If your competitive advantage is "we write CRUD faster," you don't have a competitive advantage.

2. The verification economy is beginning. The bottleneck has shifted from writing to reviewing. PR review times up 91%. Delivery stability down. But solutions are emerging: AI review agents (CodeRabbit, Qodo, Google Conductor) already reduce manual review effort by 35-50%, and workflow restructuring (stacked diffs via Graphite) increases shipping volume by 26% while keeping PRs reviewable. The organizations that build verification infrastructure — automated quality gates, AI-aware review processes, spec-driven compliance checking — will pull ahead. The ones that treat AI as "faster typing" will drown in technical debt.

3. The architecture premium is real — but calling it permanent is a bet. AI specialists command 70% wage premiums. Senior engineers who can orchestrate AI agents, verify complex outputs, and make architectural trade-offs are the scarcest resource in the market today. But agentic AI is already moving into system design, dependency mapping, and modernization. The premium exists because AI can't yet do reliable end-to-end architectural reasoning. When it can — and the trajectory of SWE-bench scores from 49% to 70%+ in under a year suggests "when," not "if" — the value chain shifts again. Bet on it being real for the next 3-5 years. Beyond that, hedge.

The N/L/C Projection

Variable	Short Term (2024-2027)	Long Term (2028-2035)
N (engineers)	Polarizing. Juniors down, mid-level squeezed, seniors up, net declining (ADP: fewer devs in 2024 than 2018).	Uncertain. Bull case: new roles expand N to 45M by 2030. Bear case: agriculture pattern — massive output, fraction of the workforce.
L (code volume)	Exploding. PR volume +98%, AI generating 46%+ of code. Quality improving with reasoning models but maintenance costs rising.	Massive expansion into long tail (SMB, niche, personal). The question is whether the maintenance wall limits sustainable growth.
C (market cost)	Restructuring. Spend shifting from headcount to AI compute + senior talent. $1.25T and climbing.	Expanding — but the mechanism matters. If C grows via Jevons (more demand), N benefits. If C grows via substitution (same output, fewer humans), N doesn't.

The world may or may not need fewer software engineers. It definitely needs different ones. And a lot more software — the open question is how many humans are required to produce it.

What To Do About It

If you're a CTO or VP of Engineering reading this, the data points to three concrete moves:

Invest in verification, not just generation. Your team will produce 2-3x more code. Build the infrastructure to review it at that scale — automated quality gates, AI-specific code review processes, security scanning tuned for AI-generated patterns. AWS CTO Werner Vogels calls the backlog of un-reviewed AI code "verification debt" — and 38% of developers say reviewing AI code is harder than reviewing human code, leading to "rubber-stamping" that defers cost to production. The organizations that win the next five years are the ones that solve the verification bottleneck before the "subprime code" bill comes due.

Price AI tooling as infrastructure, not as perks. AI development tools are becoming a budget category on par with cloud compute. Manage them accordingly: centralize billing, track usage, plan capacity. The teams that treat AI as "individual expense reports" are the ones who'll be surprised by the aggregate number — and who'll have developers self-censoring usage because the meter is running.

Plan for both outcomes. The data supports Jevons Paradox (more demand) and displacement (fewer humans) simultaneously. The organizations that survive are the ones that don't bet everything on one curve. Build your team structure so it works if demand explodes and if AI handles 80% of what your current headcount does. The worst position is being surprised by either outcome.

The specific numbers in this post have a shelf life. The structural forces don't. Code is getting cheaper. Trust is getting more expensive — for now. The gap between those two curves is where the next decade of software engineering economics lives. But keep watching the ADP data. If the developer population keeps shrinking despite rising software demand, the Jevons bet is losing.

Sources

The Industry Baseline (N, L, C)

SlashData, "Global Developer Population Trends 2025" — 36.5 million professional developers, 47.2 million total — slashdata.co
Gartner, "Worldwide IT Spending Forecast 2025" — $1.25 trillion in software spending, $5.74T total IT — cloverinfotech.com
Gartner, "IT Spending 2025 Press Release" — enterprise software spending projections — gartner.com
University of Michigan / eMarketer, "IT Spending Comparative Forecasts (2003)" — $74B software spending in 2001 — umich.edu
Wikipedia, "Software Engineering Demographics" — historical developer population estimates — wikipedia.org

The Boilerplate Tax

Ecotone.tech, "Write Only Business Logic, Eliminate Boilerplate" — 60-70% of code is orchestration and boilerplate — ecotone.tech
Microsoft "Time Warp" study / IDC Research — developers spend only 11-16% of time writing code — medium.com; corroborated by trunk.io
Stripe / Harris Poll, "Developer Coefficient" — developers spend ~42% of time on maintenance and technical debt — byteiota.com
QSM, "Long-term Trends in Software Development: A 45-Year Perspective" — lines of code per function point across language generations — qsm.com
Capers Jones / IFPUG, "Software Economics and Function Point Metrics" — SLOC per function point benchmarks (Assembly 320, C 128, Python 20-50) — ifpug.org
Medium, "Swift vs Objective-C: A Developer's View" — Lyft rewrite: 75K lines Obj-C to 22K lines Swift (~70% reduction) — medium.com
Medium, "Say Goodbye to Boilerplate: Spring Boot 3.3" — Spring Boot reducing data access boilerplate by ~80% — medium.com
itnext.io, "Building React App: Architecture & Boilerplate" — create-react-app removed ~70% of initial boilerplate — itnext.io

The Productivity Paradox

GitHub Blog, "Research: Quantifying GitHub Copilot's Impact on Developer Productivity" — 55% faster task completion, 95 devs — github.blog
IT Revolution, "New Research Reveals AI Coding Assistants Boost Developer Productivity by 26%" — MIT/Princeton/Microsoft RCT, 4,867 devs, 26% more tasks per week — itrevolution.com
Cursor Blog, "Productivity" — UChicago/Cursor study, 39% more PRs merged, tens of thousands of devs — cursor.com
METR, "Early 2025 AI Experienced Open-Source Developer Study" — 19% slower, 16 experienced devs, RCT — metr.org
McKinsey, "Unleashing Developer Productivity with Generative AI" — <10% gain on complex tasks, 40+ developers — mckinsey.com
Let's Data Science, "Developers Thought AI Made Them Faster, The Data Said Otherwise" — METR perception gap: believed 20% faster while actually 19% slower (39-point delta) — letsdatascience.com
Stack Overflow Blog, "2025 Developer Survey Results" — 84-90% adoption, trust down to 29%, 66% spending more time fixing AI code — stackoverflow.blog
Gitar.ai, "Automate AI Code Reviews" — code review times up 91% in heavy-adoption teams — gitar.ai
dev.to, "Continuous Fluid Flow: How AI is Compressing the Software Delivery Cycle" — delivery stability down 7.2% with aggressive AI adoption — dev.to
Business Insider, "Spotify developers not writing code" — Spotify engineers writing zero code since Dec 2025 using Claude Code-powered system — businessinsider.com
WeAreTenet, "GitHub Copilot Usage Data & Statistics" — 87% of developers report AI preserves mental effort on repetitive tasks — wearetenet.com
GitClear, "AI Code Quality Research 2025" — PR volume up ~98%, code churn doubling — gitclear.com; summarized at jonas.rs

The "Meatbag" Error Rates

Sogeti Labs, "How Many Defects Are Too Many?" — industry average 15-50 bugs per KLOC — sogeti.com
MayerDan, "Bugs per Line of Code Ratio" (citing Code Complete by Steve McConnell) — Microsoft ~10-20 bugs/KLOC, cleanroom ~3 bugs/KLOC — mayerdan.com
Allan Kelly Blog (citing Capers Jones) — top-quartile vs bottom-quartile developers: 10-20x productivity and quality gap — blogspot.com
CodeRabbit, "State of AI vs Human Code Generation Report" — AI PRs have 1.7x more issues (10.83 vs 6.45), 75% more logic errors, 3x readability issues — coderabbit.ai
IT Brief, "Study finds AI-generated code far buggier than human work" — AI code 1.5x-2x more security vulnerabilities — itbrief.com.au
Anthropic, "SWE-bench Sonnet" — ~49% success rate on SWE-bench Verified — anthropic.com
FutureCIO / Veracode, "Study Reveals Flaws and Risks of AI-Generated Code" — 45% of AI code samples contain security vulnerabilities, 72% failure rate in Java — futurecio.tech
IT Pro, "Developers Introduce Security Vulnerabilities with AI Assistants" — Stanford study: developers using AI write less secure code but believe it's more secure — itpro.com
CloudQA, "How Much Do Software Bugs Cost? (2025 Report)" — bug cost escalation from ~$100 (design) to $10,000-$100,000+ (production), IBM/NIST data — cloudqa.io
GitClear, "AI Code Quality Research 2025" — copy/paste up 48%, refactoring down 40% — gitclear.com

Four Economic Scenarios

Quasa.io, "The 73% Collapse: How AI Is Erasing Entry-Level Tech Jobs" — junior hiring collapsed 73% YoY (Ravio data) — quasa.io
ByteIota, "Software Engineer Salary Stagnation: AI Premium vs Junior Market Collapse 2025" — 70% of hiring managers believe AI can do intern-level work, AI specialists command 70% wage premium ($450K+) — byteiota.com
Gartner, "AI Will Touch All IT Work by 2030" — by 2030: 25% AI-only, 75% AI-augmented, 0% human-only — gartner.com
USM Systems, "Small Business AI Adoption Statistics" — 58% of SMBs using GenAI (up from 23% in 2023) — usmsystems.com
Nearshore Americas, "AI Adoption Triggers New Wave of Custom Software Spending" — custom software market growing at 22.6% CAGR (2025-2030) — nearshoreamericas.com
dev.to / Trickle.so, "Y Combinator startups with 95%+ AI-generated codebases" — 25% of YC startups have 95%+ AI-generated codebases — dev.to
VisionPoint Systems / IDC, "Low-Code Developer Population Growth" — low-code developers growing at 40.4% CAGR — visionpoint.systems
Grid Dynamics, "Number of Software Developers in the World" — developer population projected to grow to 45M by 2030 — griddynamics.com
Forbes, "The In-Demand AI Skill That Pays Up to $220,000" — Vibe Coder roles at Walmart ($110K-$220K) — forbes.com
Analytics Insight, "Elon Musk's xAI Hiring Engineers" — Vibe Coder roles up to $240K — analyticsinsight.ae
Applied AI Tools, "Rise of AI Orchestrator Jobs" — AI Orchestrator roles $120K-$180K — appliedai.tools
FlexLab, "AI Engineer Salary 2026" — AI Engineer compensation $134K-$185K, up to $400K+ — flexlab.io

The Strategic Read

Medium / Reliable Data Engineering, "AI Is Writing 46% of All Code" — GitHub Copilot generating 46% of code (61% in Java) — medium.com
McKinsey, "The Economic Potential of Generative AI: The Next Productivity Frontier" — $2.6-4.4 trillion in annual value creation — mckinsey.com
Tom's Hardware, "AI Boosted US Economy by 'Basically Zero' in 2025, Says Goldman Sachs Chief Economist" — Goldman Sachs GDP impact assessment — tomshardware.com

Counter-Evidence & Counterarguments

METR, "Uplift Update" (February 2026) — replication attempt failed due to developer refusal; returning devs showed +18% speedup — metr.org
Salesforce Engineering, "How Cursor AI Cut Legacy Code Coverage Time by 85%" — 20,000 engineers, 85% reduction in legacy code coverage time — salesforce.com
Anthropic, "Claude 3.7 Sonnet" — 70.3% on SWE-bench Verified (scaffolded), up from ~49% for Claude 3.5 — anthropic.com
GitHub Blog, "Does GitHub Copilot Improve Code Quality?" — controlled RCT: 53.2% more likely to pass all unit tests, 13.6% fewer readability errors — github.blog
Google Cloud, "DORA Report 2025" — AI as amplifier: high-performing teams improve, struggling teams degrade — blog.google
ADP Research Institute, "The Rise and Fall of the Software Developer" — US employed fewer software developers in 2024 than 2018 — adpresearch.com
Forrester Research, "AI Job Impact Forecast, US, 2025-2030" — 6% structural job loss (~10.4M roles) by 2030 — forrester.com
Citrini Research, "The 2028 Global Intelligence Crisis" — models 10.2% unemployment by 2028 via Intelligence Displacement Spiral — citriniresearch.com
Business Insider, "Citrini Research AI Job Losses" — market reaction to displacement scenario, software stocks sell-off — businessinsider.com
Acemoglu & Restrepo, "Automation and New Tasks" — displacement vs reinstatement effects, "so-so technologies" framework — mit.edu
Exceeds.ai, "Software Engineer Salary" — average SWE salary down 18.79% from 2023 — exceeds.ai
Stanford / Understanding AI, "New Evidence Strongly Suggests AI is Displacing Workers" — 22-25 year-old employment in AI-exposed roles down ~20% — understandingai.org
Goldman Sachs, "How Will AI Affect the Global Workforce?" — 6-7% displacement risk, developers as "high risk" category — goldmansachs.com
Faros AI, "AI Software Engineering Impact Report" — 21% more tasks, 98% more PRs, but PR review time +91%, bugs +9% — faros.ai

TCO & Maintenance Economics

SonarSource, "The Coding Personalities of Leading LLMs" — 90% of AI output contains code smells, security vulnerabilities at BLOCKER severity in 60-70% of cases — sonarsource.com
CodeBridge, "The Hidden Costs of AI-Generated Software" — first-year AI project costs 12% higher than traditional; maintenance can spiral to 4x by year 2 — codebridge.tech
Gartner, "Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" — escalating costs, weak risk controls, unclear business value — gartner.com
Medium, "AI Agents for Autonomous Code Refactoring" — global bank case study: 15 legacy microservices refactored, 60% time reduction, $2.3M labor savings — medium.com
CodePulse, "AI Coding Tools Impact Guide" — 38% of developers find AI code harder to review; Werner Vogels "verification debt" concept — codepulsehq.com
SD Times, "Beyond Benchmarks: Measuring the True Cost of AI-Generated Code" — TCO analysis and code smell prevalence in AI output — sdtimes.com

Analyst Track Record

Source Code Capital, "Pitfalls in Technological Predictions" — Gartner dot-com burst prediction, analyst Alexander Drobik — sourcecodecap.com
VoxelMatters, "Gartner's Top 3 Predictions on 3D Printing That Did Not Come True" — predicted $100B/year IP losses by 2018; actual losses negligible — voxelmatters.com
ZDNet, "Gartner Issues Its Own 2012 Prediction: End of IT As We Know It" — predicted 20% of businesses own no IT assets; did not materialize — zdnet.com
Quartz, "The BRICs Era Is Over, Even at Goldman Sachs" — Goldman closed BRIC fund after 88% asset decline — qz.com
Computer Weekly, "McKinsey Finds Hard Work to Do in Big Data Revisited Report" — only 30% value captured by 2016; organizational barriers underestimated — computerweekly.com
CNET, "Study Supports Controversial Offshore Numbers" — Forrester 3.3M offshoring jobs prediction: trend correct, magnitude debated — cnet.com
Wikipedia, "Gartner Hype Cycle" — Economist analysis: only about a fifth of technologies follow the prescribed path — wikipedia.org
Dreman & Berry, "Analyst Forecasting Errors and Their Implications for Security Analysis" — systematic optimism and herding biases — bath.ac.uk