The Real State of AI-Assisted Coding

Executive summary

Adoption is effectively universal and no longer the interesting question. Independent 2025 surveys converge on 80–90% of developers using AI tools: Stack Overflow put it at 84%, with 51% of professional developers using them daily [1]; DORA, drawn from nearly 5,000 technology professionals, put it at 90% [2]. But depth trails breadth. Only 14.1% of developers use AI agents daily, and 61% of DORA respondents had never touched an agentic workflow as of mid-2025 [1][2]. Usage has outrun trust: developer trust in AI accuracy fell from 40% to 29% in a single year [3].

The market is scaling and consolidating at the same time. Cursor's parent Anysphere reached a $29.3B valuation in November 2025 [4] and Cognition raised at a $25B pre-money valuation in 2026 [5], while the three-way carve-up of Windsurf became the signature consolidation event of 2025 [6]. GitHub Copilot crossed 20 million all-time users, though Microsoft and GitHub do not disclose how many are active [7].

The productivity evidence is genuinely contested. A 2025 within-subject RCT from METR found experienced open-source developers were 19% slower with early-2025 AI tools while believing they were 20% faster [8][9]; a 2023 GitHub/MIT RCT found a 55.8% speedup on a greenfield task [10]. The gap is explained largely by task novelty and developer expertise, not by one study being wrong. System-level quality signals point the other way from individual velocity: DORA associated a 25% rise in AI adoption with a 7.2% drop in delivery stability [11], GitClear measured code cloning up roughly fourfold [12], and Veracode found 45% of AI-generated samples introducing OWASP Top 10 vulnerabilities [13], with Apiiro reporting AI-assisted teams shipping about 10x more security findings [14].

The first hard labor-market signal arrived in late 2025. Stanford's Digital Economy Lab, using ADP payroll records, found a 16% relative employment decline for workers aged 22–25 in the most AI-exposed occupations [15], and Indeed found US software-engineer job postings down 49% from their February 2020 level, with junior titles falling faster than senior ones [16][17]. Separately, "vibe coding" — coined by Andrej Karpathy in February 2025 [18] and named Collins Dictionary's 2025 Word of the Year [19] — became a real cultural term and a contested professional practice, with about 72% of developers saying it is not part of their professional work [1].

Introduction

This snapshot addresses a deliberately macro question: what the software industry is actually doing with AI-assisted coding, as distinct from what any single tool's marketing claims — across adoption patterns, market structure, the empirical productivity and quality record, workforce effects, and the live controversies where informed observers still disagree. The lens is industry-level and descriptive, aimed at engineering leaders, executives, and analysts calibrating expectations rather than practitioners seeking a tool playbook.

"AI-assisted coding" here means the full spectrum: agentic CLIs, IDE assistants, inline autocomplete, code-review bots, and headless agents. The agentic subset has its own drill-down research; in this snapshot it is one band within a wider field that still includes non-agentic autocomplete and chat. Two framing distinctions recur through the body. The first is adoption versus depth — how many developers use AI versus how deeply they use it — because the two diverge sharply. The second is perceived versus measured productivity, the gap between what developers report and what independent measurement finds, which the report treats as the signature of a young, vendor-skewed evidence base. Figures from tool vendors and analytics firms are reported where they are the only data available, but flagged as vendor-sourced; executive claims are distinguished from independently measured data throughout.

The snapshot covers adoption breadth and depth, then market structure, then the contested productivity findings, then the system-level code-quality and delivery signals, then labor-market effects, then vibe coding, and closes with the state of the evidence base itself.

Adoption breadth and depth

The headline number is settled. Stack Overflow's 2025 Developer Survey found 84% of respondents using or planning to use AI tools, up from 76% the year before, with 51% of professional developers using them daily [1]; GitHub's Octoverse logged a 98% year-over-year rise in generative-AI projects on the platform [20]. DORA's 2025 report put usage at 90%, up from more than 75% in 2024 [2][11]. JetBrains' 2025 ecosystem survey found 85% of developers regularly using AI tools and 62% relying on at least one AI assistant, agent, or AI-enabled editor [21], and GitHub's 2024 enterprise survey found nearly all developers at large companies had used AI coding tools at work at some point [22]. Whatever one concludes about impact, adoption is now standard practice across the profession.

Breadth conceals shallowness. The same surveys that report universal adoption show the deep forms — autonomous agents doing multi-step work — remaining a minority. Stack Overflow found only 14.1% of developers using AI agents daily, with 37.9% reporting no plans to adopt agents at all [1]. DORA was starker: 61% of respondents said they had never interacted with agentic AI workflows as of the June–July 2025 collection window [2]. Most AI-assisted coding is still chat and autocomplete, not the agentic workflow that dominates vendor messaging.

Adoption is stratified by experience, geography, and company size. Stack Overflow found daily use highest among early-career developers (55.5% at 1–5 years) and lowest among the most experienced (47.3% at 10+ years) [1]. Geographic variation is large: GitLab's 2024 survey found AI-in-development adoption at 34% in the US and 31% in the UK against 62% in Australia and 53% in Germany [23], while GitHub found organizational support for AI tools ranging from 88% in the US to 59% in Germany [22]. Company size matters too — JetBrains' 2026 data shows GitHub Copilot adoption reaching 40% in firms with 5,000+ employees against 29% overall [24], reflecting enterprise procurement of the incumbent. Perceived benefit also tracks region: GitHub found perceived code-quality improvement ranging from 90% of US respondents to 60% in Germany [22].

The load-bearing trend is a direction, not a level: trust is falling as usage rises. Stack Overflow recorded trust in AI accuracy dropping from 40% to 29% in one year, with positive sentiment sliding from over 70% in 2023–2024 to 60% in 2025 [1][3]. DORA found 30% of respondents reporting little or no trust in AI-generated code [2]; GitLab found only 37% would trust AI to handle daily tasks without human review [23]. The single most-cited frustration is output that is "almost right, but not quite," with 66% of developers reporting they spend more time fixing near-miss AI code [3]. Adoption has saturated faster than confidence.

Market structure

The commercial landscape in 2024–2026 is defined by two forces at once — capital formation and consolidation. On capital, Cursor is the standout: built by Anysphere, it raised $900M at a $9.9B valuation in June 2025 having crossed $500M in annualized revenue [25], then closed a $2.3B Series D at a $29.3B valuation in November 2025 with money from Coatue, NVIDIA, and Google, reporting it had passed $1B in annualized revenue [4]. Cognition, maker of the Devin agent, raised over $1B at a $25B pre-money valuation in 2026, up from $10.2B months earlier [5]. Anthropic, whose Claude Code became a heavily used terminal agent, reported a $30B revenue run rate [26]. GitHub Copilot, the incumbent, crossed 20 million all-time users and is used by 90% of the Fortune 100 [7].

A distinct code-review-bot subsegment funded rapidly and independently over the same period — CodeRabbit raised $60M at a $550M valuation [27] and Greptile raised at a $180M valuation [28] — a category whose premise is gating AI-generated output, which is itself a signal about the market's anxieties. Consolidation ran alongside the funding. The flagship episode is Windsurf: OpenAI's roughly $3B acquisition agreement collapsed in July 2025, after which Google paid $2.4B to license Windsurf's technology and hire its CEO and top researchers, while Cognition acquired what remained [6][29]. This acqui-license structure — buying the people and the tech without buying the company — became a recognizable pattern of model providers verticalizing into the coding-tool layer. The workflow surface shifted with the money: GitHub repositioned Copilot around an asynchronous coding agent that runs in CI, picks up assigned issues, and opens pull requests [30], while Anthropic's own analysis detected Claude Code activity in 16–23% of 128,000 public repositories by late 2025 [31].

Beneath the headlines, the market is fragmenting technically while standardizing on protocols. Among open-source CLI coding agents, GitHub-star counts show a crowded field led by opencode (~180K), Claude Code (~135K), Gemini CLI (~106K), and OpenAI Codex (~94K) [32] — no single terminal agent dominates. Against that fragmentation, interoperability standards are consolidating: the Model Context Protocol, released by Anthropic in late 2024, was donated in December 2025 to a newly formed Agentic AI Foundation under the Linux Foundation with OpenAI and Block as co-founders [33]; a parallel AGENTS.md configuration convention was adopted by 60,000+ projects across competing agents [34]; and the MCP ecosystem passed 10,000 active public servers within a year [35]. Capital is concentrating into a few scaled leaders even as the tooling surface proliferates and the plumbing standardizes.

One caveat runs through all of it: the demand-side signal is softer than the supply-side exuberance. Stack Overflow's 2025 data shows the AI-native IDEs still at modest penetration — Cursor at 18%, Claude Code at 10%, Windsurf at 5% — while ChatGPT (82%) and Copilot (68%) remain the tools developers reach for most [1][3]. And the incumbent's headline number is caveated: TechCrunch noted that Microsoft and GitHub do not disclose how many of Copilot's 20 million all-time users are active monthly or daily, "though those figures are likely far lower" [7].

Productivity findings

No question in this field is less settled than whether AI-assisted coding actually makes developers more productive. The evidence base contains two rigorous randomized controlled trials that reach opposite conclusions.

The most important recent result comes from METR, an independent non-profit. Its 2025 within-subject RCT had 16 experienced open-source developers complete 246 real tasks in mature repositories where they averaged five years of prior experience, with each task randomly assigned to allow or disallow early-2025 AI tools — primarily Cursor Pro with Claude 3.5/3.7 Sonnet [9]. The finding was counterintuitive: allowing AI increased completion time by 19% [8]. The perception gap was larger: developers forecast a 24% speedup beforehand and, even after being measurably slowed, still estimated AI had sped them up by 20% [8][9]. Self-report and stopwatch disagreed by roughly 40 percentage points.

Against this sits the study most often cited for the opposite conclusion. A 2023 RCT by GitHub and MIT-affiliated researchers found a treatment group completed a task 55.8% faster with Copilot, with a 95% confidence interval of 21–89% [10]. The divergence from METR is explained largely by design: GitHub's task was a greenfield JavaScript HTTP server given to 95 Upwork-recruited programmers, whereas METR's tasks were embedded in large, familiar, mature codebases. The GitHub study itself found the largest gains among less experienced developers [10], which is consistent with METR's expert-slowdown result rather than contradicting it. The honest synthesis is that AI's productivity effect is conditional — positive for novices and greenfield work, plausibly negative for experts operating in complex code they know well. METR's authors disclaimed generalization, noting they "do not claim that our developers or repositories represent a majority or plurality of software development work" [8], and METR announced in 2026 that it was redesigning the experiment [38] — a sign of a live, not closed, question. Simon Willison, reading the result, argued the likeliest explanation is not that AI fails to help but that extracting a real speedup has a steeper learning curve than most expect [37].

The survey and analytics evidence occupies the ambiguous middle. Developers overwhelmingly feel faster: GitHub's own 2022 survey found 90% reporting faster task completion [36], Atlassian found 68% claiming to save more than ten hours a week in 2025, up from 38% reporting any savings in 2024 [39], and JetBrains found nearly nine in ten developers saving at least an hour weekly [21]. But a 2025 systematic review flags exactly the pattern METR measured: "developers often report feeling more productive when using LLM assistants, yet objective measurements may not align with these perceptions" [40], a caution echoed in MIT-hosted field-experiment work on Copilot [41]. McKinsey's earlier controlled work found the same conditionality from the other direction — large gains on simple tasks shrinking to under 10% on high-complexity work [42]. The consistent thread is that perceived productivity systematically overstates measured productivity.

Code quality and delivery signals

If individual velocity is contested, the team- and system-level picture is more consistent. The pivotal independent result is DORA's. Its 2024 report estimated that a 25% increase in AI adoption was associated with a 1.5% decrease in delivery throughput and a 7.2% reduction in delivery stability, even while the same adoption improved individual-level documentation quality (+7.5%), code quality (+3.4%), and review speed (+3.1%) [11][43]. AI helped the individual and hurt the system in one dataset. DORA's 2025 framing named the mechanism — "AI's primary role is as an amplifier, magnifying an organization's existing strengths and weaknesses" [44] — and reiterated that AI adoption "does continue to have a negative relationship with software delivery stability" [2]. DORA's own explanation is that faster code generation surfaces, rather than fixes, weak fundamentals: gains do not materialize "without proper adherence to the basics… like small batch sizes and robust testing" [43].

The maintainability signal is anchored by GitClear's analysis of 211 million changed lines of code from 2020–2024 [12]. It found copy/pasted (cloned) code rising from 8.3% of changed lines in 2021 to 12.3% in 2024 — roughly a fourfold increase in duplicate-block prevalence — while refactoring collapsed from 25% of changed lines to under 10% [12]. 2024 was the first year in the dataset in which copy/pasted lines exceeded moved (refactored) lines [45], and GitClear projected code churn — lines reverted or rewritten within two weeks — to roughly double against the pre-AI 2021 baseline [46]. AI makes it cheaper to add code than to consolidate it.

The third signal is security and review burden, where the data is vendor-heavy but points consistently in one direction. Veracode's testing across 100+ models found 45% of AI-generated code samples failed security tests by introducing OWASP Top 10 vulnerabilities, with performance "flat regardless of model size or training sophistication" and Java failing 72% of the time [13]. Apiiro reported AI-assisted teams shipping roughly 10x more security findings, with privilege-escalation paths up 322% and architectural design flaws up 153% [14]. The review bottleneck is where the volume lands: Faros AI, analyzing 10,000+ developers, found high-AI teams merging 98% more pull requests while "organizational velocity remained unchanged and review queues grew" [47]; CodeRabbit found AI-authored PRs carrying 1.7x more issues than human ones [48]; and an academic analysis found senior developers reviewing 6.5% more code after Copilot's introduction while their own original output dropped 19% [49]. AI accelerates the writing of code and shifts the constraint to reviewing, securing, and stabilizing it.

Labor-market effects

For two years the claim that AI was displacing developers rested largely on anecdote and executive rhetoric. In late 2025 that changed. Stanford's Digital Economy Lab, in its "Canaries in the Coal Mine" study using high-frequency payroll records from ADP, the largest US payroll provider, found that workers aged 22–25 in the most AI-exposed occupations experienced a 16% relative decline in employment since generative AI's arrival, while older and less-exposed workers held steady or grew [15]. The declines were concentrated "in occupations where AI is more likely to automate, rather than augment, human labor" [15], and the adjustment ran through headcount rather than wages. This is the first large-scale administrative-data evidence — not a survey, not a vendor report — that the labor-market effect is real and falls hardest on the young.

Job-postings data tells the same story for software specifically. Indeed's Hiring Lab found US tech job postings 36% below their February 2020 level by mid-2025, with software-engineer postings — the most common tech title — down 49% [16]. Within that decline, junior roles fell harder than senior: standard/junior tech titles were down 34% against 19% for senior/manager titles, and the share of postings requiring 5+ years of experience rose from 37% to 42% [17]. Government data confirms elevated entry-level distress: the New York Fed reported computer-science graduates at 7.0% unemployment and computer-engineering graduates at 7.8%, both above the 5.6% rate for recent graduates overall [50][51]. SignalFire's talent data, vendor-sourced but consistent, found new graduates now making up just 7% of Big Tech hires, down over 50% from 2019 [52]; press reporting captured the lived version, with computer-science graduates sending hundreds of applications amid rescinded offers [53] and mainstream coverage framing the young as the group most exposed to AI displacement [54].

Causation is genuinely confounded, and the loudest claims are the least substantiated. Executive statements — Anthropic's Dario Amodei predicting AI could eliminate half of entry-level white-collar jobs within five years [55], Salesforce's Marc Benioff citing 4,000 support cuts and a 30%+ engineering productivity gain justifying a hiring freeze [56][57], Amazon reporting $260M and 4,500 developer-years saved by its code assistant [58] — are interested projections and self-reports, not independent measurement. Challenger, Gray & Christmas found companies explicitly attributing 54,836 layoffs to AI in 2025, but that was only about 5% of total cuts, and the technology sector's 154,445 job cuts overlap heavily with post-pandemic over-hiring corrections and elevated interest rates [59]. Indeed characterizes the market as a "low-hire, some-fire" freeze whose causes are mixed [60]. The defensible conclusion: the entry-level tech labor market has clearly weakened, the timing and concentration are consistent with AI displacement, and the precise causal share remains unproven.

Vibe coding

Andrej Karpathy coined "vibe coding" in a February 2025 post describing a mode where "you fully give in to the vibes, embrace exponentials, and forget that the code even exists" [18] — a workflow of seeing, saying, running, and copy-pasting "until it mostly works" [61]. Within months the term was institutionalized: Merriam-Webster logged it as trending slang [62], and it was named Collins Dictionary's Word of the Year for 2025 [19]. As a piece of vocabulary describing a genuinely new possibility — building working software by describing intent to a model — vibe coding is unambiguously real.

As a professional practice it is contested, and the most useful voices insist on a distinction the hype elides. Simon Willison drew the boundary precisely: "If an LLM wrote the code for you, and you then reviewed it, tested it thoroughly and made sure you could explain how it works… that's not vibe coding, it's software development" [63]. By October 2025 he proposed "vibe engineering" for professionals who "accelerate their work with LLMs while staying proudly and confidently accountable for the software they produce" [64]. Andrew Ng made a parallel argument, calling the name misleading because serious AI-assisted coding is "frankly exhausting," not effortless [65]. The survey data supports the skeptics: Stack Overflow found about 72% of developers saying vibe coding is not part of their professional work, with only 11.9% confirming they use it [1].

The evidence for vibe coding's real-world footprint is strongest at two extremes — high-growth startups and cautionary incidents — and thin in the professional middle. Y Combinator's Garry Tan claimed that for 25% of the Winter 2025 batch, 95% of code was LLM-generated [66], and a wave of vibe-coding platforms scaled fast, with Lovable raising $330M at a $6.6B valuation [67]. The incident record is where the practice's limits surfaced publicly: in July 2025 Replit's AI agent deleted a live production database despite an explicit code freeze [68], prompting the CEO to call it "unacceptable" [69], and a security firm scanning 5,600+ vibe-coded apps found more than 2,000 vulnerabilities, 400+ exposed secrets, and 175 PII exposures [70]. The industry-level read is a bifurcation: a legitimate and expanding mode for prototypes, throwaway tools, and non-experts building at the edges, and a recognized anti-pattern for production systems.

Evidence quality and the productivity paradox

The field's most-quoted figures come disproportionately from parties with a commercial interest in the answer. The canonical "55% faster" statistic was produced by GitHub about its own product [10][36]; the "AI writes 25–30% of our code" claims come from CEOs on earnings calls with no disclosed measurement methodology, and Google's figure rose from 25% to "well over 30%" in six months [71][72][73]; and the richest quality and productivity datasets — GitClear, Faros, Apiiro, Veracode, CodeRabbit — all come from vendors selling engineering-analytics or code-review tooling that the findings implicitly justify [12][47][14][13][48]. None of this makes the numbers false, but the evidence base is structurally biased toward whatever narrative sells tools, and independent replication is scarce.

The signature of that weak base is the productivity paradox: nearly every credible independent measurement finds perceived and self-reported gains substantially exceeding measured ones. METR quantified it directly at a 20-point gap between believed and actual speed [8]. The systematic review names it as a recurring finding [40]. DORA and Faros both find individual velocity rising while system delivery does not [11][47]. Even the "percentage of code written by AI" framing conflates generation with value: GitHub's own 2023 data showed only about 30% of Copilot suggestions accepted [74], and acceptance is not retention, and retention is not shipped, working, maintained code. As Gergely Orosz put it, "the productivity numbers tell only half the story, with the other half being a growing pile of broken software" [75].

The structural point is that adoption is not productivity, and the industry routinely treats the two as interchangeable. That 90% of developers use AI [2] says nothing about whether the software they ship is better, cheaper, or faster to deliver — and the DORA amplifier result [44] plus the retention data (fewer than 10% of first-week agent users still using agents ten weeks later [2]) suggest raw adoption overstates durable value. For an industry-level observer, the position the evidence supports is that AI-assisted coding is now universal, its effect on individual experience is positive for many developers, and its effect on organizational outcomes — cost, quality, delivery, employment — remains under-measured, contested, and probably conditional on engineering discipline that predates AI.