Will AI replace software developers?
Code generation worked. That is the part almost nobody disputes. What the 2026 data shows is that generating code was never the constraint — and once output multiplied, the constraint moved to verifying it. In Q1 2026, reviewing AI-generated code overtook writing code as the single largest consumer of developer time. The job did not disappear. It inverted.
Thank you for reading this post, don't forget to subscribe!
Key takeaways
- →Review overtook writing. 11.4 hours a week reviewing AI-generated code against 9.8 hours writing new code. MEDIUM
- →The review queue is buckling. Median pull-request review time is up 441%, PRs are 51% larger, bugs per developer up 54% — and 31% more PRs merge with no review at all. MEDIUM
- →Throughput and instability rose together. DORA finds time saved generating code is re-spent auditing it; every 25pp of AI code adoption costs 7.2% of change-failure-rate stability. MEDIUM
- →Nobody is measuring the cost. 94% of teams omit validation time, tech debt and burnout from their metrics entirely. MEDIUM
- →The entry level took the hit, not the profession. Entry-level postings down 67% since 2022; 54% of companies stopped hiring juniors because of AI. MEDIUM
The inversion: where developer time went
| Measure | Finding | Source | Confidence |
|---|---|---|---|
| Weekly hours reviewing AI code vs writing new code | 11.4h vs 9.8h | Developer survey, Q1 2026 — a reversal from Q4 2024, when writing held a four-hour lead | MEDIUM |
| Same, among self-described heavy agentic-tool users | 14–16h reviewing | Writing hours flat or modestly down | MEDIUM |
| Developers spending more time in review since their team adopted AI — independent corroboration | 81% | Harness, State of Engineering Excellence 2026; 700 practitioners across US, UK, India, France, Germany. 28% report review time up by 30% or more. | MEDIUM |
| Incidents per pull request | +242.7% | Faros AI. Faros’s own caveat, which we repeat because their headline drops it: “This is a ratio, not a probability: a single PR can be linked to multiple incidents, and not every incident traces directly to the most recent merge.” | MEDIUM |
| Monthly incidents | +57.9% | Faros AI — the absolute count, without the per-PR framing | MEDIUM |
| Median time a pull request spends in review | +441.5% | Faros AI — the figure usually quoted | MEDIUM |
| Average time in review | +199.6% | Faros AI — a different metric; routinely confused with the one above | MEDIUM |
| Median time to first review | +156.6% | Faros AI — how long a PR waits before anyone looks at it | MEDIUM |
| Pull request size | +51% | Faros AI | MEDIUM |
| Bugs per developer | +54% | Faros AI — up from +9% in their 2025 report. This is the headline figure, but it does not control for volume. | MEDIUM |
| Bugs per pull request | +28.7% | Faros AI — in the report PDF only, not the public summary. This is the harder-to-dismiss number: it controls for output volume, and it is roughly half the headline. | MEDIUM |
| Pull requests merged with no review at all | +31% | Faros AI | MEDIUM |
| Developers saying AI output is “almost correct” but flawed | 66% | Developer survey | MEDIUM |
“Almost correct” is the expensive failure mode. Code that is obviously wrong costs seconds. Code that is plausible, compiles, passes a shallow read and is wrong in one place costs a reviewer full attention on every line — which is why review time rose four and a half times rather than proportionally with volume.
What happens to the codebase
| Measure | Change | Source | Confidence |
|---|---|---|---|
| Copy-pasted code as a share of all new code | 9.4% → 15.7% (2022–early 2026) | GitClear | MEDIUM |
| “Moved” code — GitClear’s proxy for refactoring — as a share of changed lines | 21% (2022) → 3.8% | GitClear | MEDIUM |
| Two-week code churn — share of authored lines rewritten or removed within two weeks | +15% | GitClear, 623M changes | MEDIUM |
| Code churn on a different definition — ratio of lines deleted to lines added per quarter | +861% ~10× the prior rate | Faros AI — not comparable with the row above; same word, incompatible metrics | MEDIUM |
| Change-failure-rate stability, per 25pp of AI code adoption | −7.2% | DORA 2025 | MEDIUM |
Duplication up four-fifths, refactoring down by more than 80%. Those two move together for an obvious reason: generating a fresh near-copy is now cheaper than understanding and reworking what exists. That is a rational choice per pull request and a costly one per codebase, and it shows up later as churn — code written and rewritten within two weeks, work that produced no durable value.
The Meta proof point
Meta’s Project OT is the only case where a company at frontier scale attempted deep replacement and then published enough numbers to audit it. Under a plan that explored cutting some teams by up to 60%, CTO Andrew Bosworth’s June 2026 figures reported: code changes to internal platforms +220%, changes reaching users as features +36%, major technical and security incidents +40%, and time spent firefighting them +70%. The second planned wave of cuts was cancelled. HIGH
Output tripled; shipped value grew by about a third; the failure rate and the cleanup grew faster than the value did. Full detail in The AI replacement reversal.
The part that does not recover quickly
| Measure | Figure | Period | Confidence |
|---|---|---|---|
| Fall in entry-level developer job postings | −67% | 2022–2026 | MEDIUM |
| Entry-level postings rose while actual hiring fell — roles filled with senior staff instead | +47% / −73% | 2023–2024 | MEDIUM |
| Companies that stopped hiring juniors because of AI | 54% | 2026 | MEDIUM |
| Senior developers reporting burnout | 68% | 2024 | MEDIUM |
| Average developer burnout rating, ~75% sustained 6+ months | 7.4 / 10 | 2026 | MEDIUM |
| Teams omitting tech debt, validation time and burnout from metrics | 94% | 2026 | MEDIUM |
The trap is structural. Verification is now the binding constraint and only experienced engineers can do it. Juniors historically became experienced by doing the routine, low-stakes work — which is precisely the work AI absorbed. The training ground and the automation target are the same ground. A 67% collapse in entry-level hiring across 2024–2026 produces a mid-level shortfall around 2030–2032 and a senior shortfall by the mid-2030s, at exactly the moment AI-generated code makes experienced reviewers more necessary, not less. AWS CEO Matt Garman called replacing junior developers with AI “one of the dumbest things I’ve ever heard.”
So: will AI replace your developer job?
If your value is producing volume of routine code, that is the part being absorbed — boilerplate, CRUD, glue, standard integrations. This is real and it is not reversing.
If your value is judgement about what should be built, whether it is correct, and who answers when it breaks, the evidence says demand for you rose. Every measure above describes an industry short of review capacity, not short of code.
The honest warning is about entry. The profession is not shrinking; the on-ramp is. If you are early-career, the strategy the data supports is to get to demonstrable review-and-judgement capability faster than the old apprenticeship allowed — reading and critiquing code, debugging systems you did not write, and owning production consequences — because that is the work with a queue in front of it.
Frequently asked
Is AI replacing programmers in 2026?
It is replacing routine code production, not the profession. The measured effect is a shift in where time goes: reviewing AI-generated code overtook writing it in Q1 2026, at 11.4 hours per week against 9.8. Median pull-request review time rose 441% and bugs per developer rose 54%, so demand for experienced reviewers increased rather than fell.
Does AI-generated code cause more bugs?
The delivery data says yes at the system level. Bugs per developer are up 54%, DORA finds every 25 percentage points of AI code adoption costs about 7.2% of change-failure-rate stability, and GitClear’s analysis of 623 million code changes shows copy-pasted code rising from 9.4% to 15.7% while “moved” code — its proxy for refactoring — fell from 21% to 3.8%. At Meta, incidents rose 40% and firefighting time 70%.
Are junior developer jobs disappearing?
Entry-level developer postings fell 67% between 2022 and 2026 and 54% of companies stopped hiring juniors because of AI. Because nobody becomes a senior engineer without first being a junior one, this produces a mid-level shortage around 2030–2032 and a senior shortage by the mid-2030s. AWS CEO Matt Garman has called the practice one of the dumbest ideas he has heard.
Which development work is safest?
Work where judgement and accountability are the product: architecture and design decisions, code review, debugging unfamiliar systems, security, incident response, and anything where someone must own the consequence of a wrong answer. The industry’s measured shortage is in review capacity, not in code output.
Methodology & sources
Meta’s internal figures originate in a June 2026 post by CTO Andrew Bosworth as reported by Reuters (26 Aug 2026) and are rated HIGH; we have not independently verified the underlying internal data. Delivery metrics come from Faros AI’s AI Engineering Report 2026: The Acceleration Whiplash, based on two years of telemetry from 22,000 developers across 4,000+ teams. Code-quality metrics come from GitClear’s The Maintainability Gap (June 2026), based on 623 million analyzed changes, 2023–2026 — note that the unit is changes, not lines; GitClear’s earlier papers counted changed lines and are not comparable. Both are vendor research and rated MEDIUM. Four caveats a reader should carry:
- Faros is not a before-and-after study. Every “+X%” is a within-organisation comparison of each team’s two quarters of lowest AI adoption against its two quarters of highest. There is no control group and no clean pre-AI baseline, so these are associations under rising adoption, not measured effects of it.
- GitClear’s “moved” metric is in tension with GitClear’s own product. The research paper uses moved lines as “a proxy for refactoring activity,” while their Diff Delta scoring assigns moved code a value of zero, describing it as “high line counts, little real work.” Anyone citing “refactoring down 70%” is citing a proxy the same company treats as valueless work elsewhere. We report the figure and the tension together.
- Baselines and windows are inconsistent in the source. GitClear’s summary says the −35% / −70% / −74% cluster is “vs 2022 levels”, but the per-signal paragraphs give 2023 baselines; we use 2023, which is the one their numbers support. Separately, copy/paste at 15.7% is “the first half of 2026” while moved code at 3.8% is “year-to-date” — not the same window.
- Faros and DORA disagree. Faros explicitly contests DORA’s finding that strong engineering foundations buffer AI’s downsides. Their observational design cannot really settle that, so where this page cites both, treat them as two readings rather than one consensus.
GitClear’s full whitepaper sits behind an email capture form, which we did not submit; everything above comes from their public research, signal-graph and methodology pages. Hiring figures describe job postings, which lead actual hiring and can overstate turning points in both directions. Quarterly cadence. Next review: December 2026.
Correction, 11 September 2026. On first publication this page described Faros’s +54% figure as “bugs per pull request”; the metric is bugs per developer (up from 9% in the 2025 report). It also gave GitClear’s sample as 211 million lines — the correct figure is 623 million analyzed changes — and labelled the 21% → 3.8% decline “refactored code”, where GitClear’s metric is “moved” code, used as a proxy for refactoring. Third, it attributed a “up to 9× higher code churn” figure to GitClear. That was mis-sourced: GitClear’s churn figure in this paper is +15% two-week churn, and the ~9× number is Faros’s 861% churn increase — measured on an entirely different definition (lines deleted to lines added per quarter, versus GitClear’s share of lines rewritten within two weeks). Both now appear, labelled, with a note that they are not comparable. The percentages in the first two corrections were themselves accurate; the labels and attribution were not. Corrected the same day.
Part of the Will AI Replace My Job? pillar. See also customer support and The AI replacement reversal.