docs: add Borg et al. 2025 RCT on AI code maintainability (v3.27.7)
- Resource eval: arXiv:2507.00788 "Echoes of AI" (151 devs, 95% pros, 2-phase blind RCT) — 30.7% faster median, ~55.9% habitual users, no significant downstream maintainability impact - guide/learning-with-ai.md: citation + "On maintainability fear" note - guide/ultimate-guide.md: nuance blockquote in §1.7 Trust Calibration - machine-readable/reference.yaml: 4 new RCT/maintainability entries - docs/resource-evaluations/: evaluation file with technical-writer audit Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Florian BRUNIAUX
parent
4c42151151
commit
895ace49f7
6 changed files with 197 additions and 1 deletions
20
CHANGELOG.md
20
CHANGELOG.md
|
|
@ -12,6 +12,26 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
|
|||
- v2.1.47: VS Code plan preview auto-updates, `ctrl+f` kills all background agents, `last_assistant_message` hook field, 70+ bug fixes
|
||||
- v2.1.46: claude.ai MCP connectors support, orphaned process fix on macOS
|
||||
|
||||
## [3.27.7] - 2026-02-19
|
||||
|
||||
### Added
|
||||
|
||||
- **Resource evaluation**: Borg et al. "Echoes of AI" RCT (arXiv:2507.00788)
|
||||
- 2-phase blind controlled experiment, 151 participants (95% professional developers)
|
||||
- AI users 30.7% faster (median), habitual users ~55.9% faster
|
||||
- No significant maintainability impact for downstream developers — first RCT to explicitly target this question
|
||||
- Fact-checked against primary source; v2 (Dec 2025) confirmed via Perplexity
|
||||
- Co-authored by Dave Farley ("Continuous Delivery")
|
||||
- Evaluation file: `docs/resource-evaluations/2026-02-19-echoes-of-ai-maintainability-study.md`
|
||||
|
||||
### Changed
|
||||
|
||||
- `guide/learning-with-ai.md`: Added Borg et al. 2025 RCT citation in Productivity Research bibliography (revised to factual/neutral wording after technical-writer audit)
|
||||
- `guide/learning-with-ai.md`: Added "On maintainability fear" note in "Why Some Teams Get Results" section — the real risks are skill atrophy and over-delegation, not downstream quality degradation
|
||||
- `guide/ultimate-guide.md`: Added downstream maintainability nuance blockquote in §1.7 Trust Calibration — defect rates ≠ maintenance burden (Borg et al. 2025 blind RCT)
|
||||
- `machine-readable/reference.yaml`: Added 4 entries — `productivity_rct_metr`, `productivity_rct_echoes`, `productivity_maintainability_empirical`, `trust_calibration_maintainability_nuance`
|
||||
- Landing `faq/index.astro`: Updated "How much should I trust AI-generated code?" — added maintainability nuance (HTML visible answer + JSON-LD structured data)
|
||||
|
||||
## [3.27.6] - 2026-02-18
|
||||
|
||||
### Added
|
||||
|
|
|
|||
2
VERSION
2
VERSION
|
|
@ -1 +1 @@
|
|||
3.27.6
|
||||
3.27.7
|
||||
|
|
|
|||
|
|
@ -0,0 +1,166 @@
|
|||
# Resource Evaluation: "Echoes of AI: Investigating the Downstream Effects of AI Assistants on Software Maintainability"
|
||||
|
||||
**Date:** 2026-02-19
|
||||
**Evaluator:** Claude Code (eval-resource skill)
|
||||
**Status:** Integrated (section Productivity Research)
|
||||
|
||||
---
|
||||
|
||||
## Resource Details
|
||||
|
||||
| Field | Value |
|
||||
|-------|-------|
|
||||
| **Title** | Echoes of AI: Investigating the Downstream Effects of AI Assistants on Software Maintainability |
|
||||
| **Authors** | Markus Borg, Dave Hewett, Nadim Hagatulah, Noric Couderc, Emma Söderberg, Donald Graham, Uttam Kini, Dave Farley |
|
||||
| **Dave Farley credentials** | Co-author of "Continuous Delivery" (with Jez Humble), Jolt Award winner |
|
||||
| **Publication Date** | July 2025 |
|
||||
| **URL** | https://arxiv.org/abs/2507.00788 |
|
||||
| **Type** | Academic preprint (arXiv, not yet peer-reviewed in conference proceedings as of 2026-02-19) |
|
||||
| **LinkedIn context** | Post by Olivier LOVERDE (Co-founder & CPTO @Innovorder) summarizing the study |
|
||||
| **LinkedIn URL** | https://www.linkedin.com/posts/loverdeolivier_investigating-the-downstream-effects-of-ai-ugcPost-7426914640300802048 |
|
||||
|
||||
---
|
||||
|
||||
## Summary
|
||||
|
||||
Two-phase controlled experiment investigating whether AI-assisted code creation impacts maintainability for downstream developers:
|
||||
|
||||
- **Phase 1**: 151 participants add features to a Java web app (with or without AI: GitHub Copilot / Cursor)
|
||||
- **Phase 2**: A *different* group of developers evolves those solutions **without AI** (blind review — reviewers don't know if code was AI-assisted)
|
||||
|
||||
**Key findings:**
|
||||
1. AI users completed tasks **30.7% faster** (median) than non-AI users
|
||||
2. Habitual AI users showed an estimated **55.9% speedup**
|
||||
3. **No significant differences** in downstream evolution time or code quality — the "AI code is unmaintainable" myth is not supported empirically
|
||||
4. Researchers recommend future investigation into excessive code generation and cognitive debt risks
|
||||
5. Dave Farley's explicit takeaway: developers must guide AI (not autopilot), think about the business problem, and decompose complexity
|
||||
|
||||
---
|
||||
|
||||
## Evaluation Score: **4/5** (Très pertinent — amélioration significative)
|
||||
|
||||
### Scoring Breakdown
|
||||
|
||||
| Criterion | Score | Justification |
|
||||
|-----------|-------|---------------|
|
||||
| **Content novelty** | 4/5 | Directly addresses the #1 FUD against AI-assisted coding ("unmaintainable code") with empirical data |
|
||||
| **Research rigor** | 4/5 | 151 participants, 95% professional developers, 2-phase blind design — solid for this domain. Caveat: arXiv preprint, not yet peer-reviewed in proceedings |
|
||||
| **Guide specificity** | 4/5 | Complements METR 2025 (already in guide) — provides the counter-evidence the guide currently lacks |
|
||||
| **Credibility** | 5/5 | Dave Farley authorship = exceptional signal for the software engineering community |
|
||||
| **Actionability** | 3/5 | Results are confirmatory, not prescriptive — validates approach but doesn't change workflow |
|
||||
|
||||
**Overall: 4/5**
|
||||
|
||||
---
|
||||
|
||||
## Gap Analysis
|
||||
|
||||
### What the guide already covers
|
||||
|
||||
| Topic | Coverage | Location |
|
||||
|-------|----------|----------|
|
||||
| AI productivity gains | ✅ General stats (Copilot, McKinsey) | `learning-with-ai.md:921-924` |
|
||||
| METR RCT (19% slower) | ✅ Present | `learning-with-ai.md:925` |
|
||||
| Vibe coding risks | ✅ Full section | Multiple locations |
|
||||
| Skill atrophy concern | ✅ Present | `learning-with-ai.md:925` |
|
||||
| AI code maintainability myth | ❌ **ABSENT** | **Gap identified** |
|
||||
| Productivity curve (habitual users) | ⚠️ Partially | `learning-with-ai.md:~100` |
|
||||
|
||||
### What this study adds
|
||||
|
||||
| Contribution | Value |
|
||||
|--------------|-------|
|
||||
| Empirical refutation of "AI code is unmaintainable" | **High** — directly debunks the most common objection |
|
||||
| 55.9% speedup for habitual users | **High** — validates learning curve section |
|
||||
| Blind review methodology | **Medium** — demonstrates scientific rigor of the finding |
|
||||
| Balance to METR 2025 results | **High** — METR = complex codebases, AI slower; this study = mixed tasks, AI faster → complete picture |
|
||||
|
||||
---
|
||||
|
||||
## Recommendations
|
||||
|
||||
**Where to integrate**: `guide/learning-with-ai.md` — section "Productivity Research" (~line 925)
|
||||
|
||||
**What to add** (1-2 lines):
|
||||
```markdown
|
||||
- **Borg et al. "Echoes of AI" RCT (2025)** — [arXiv:2507.00788](https://arxiv.org/abs/2507.00788) — Controlled experiment (151 participants, 95% professional developers, 2-phase blind design): AI users 30.7% faster (median), habitual users ~55.9% faster. **Key finding**: no significant maintainability impact for downstream developers. Directly refutes the "AI code is unmaintainable" myth. Caveat: arXiv preprint (July 2025), not yet peer-reviewed in conference proceedings.
|
||||
```
|
||||
|
||||
**Priority**: Medium-High — completes the empirical picture alongside METR 2025.
|
||||
|
||||
---
|
||||
|
||||
## Challenge Summary (technical-writer agent)
|
||||
|
||||
**Initial score:** 4/5
|
||||
**Challenged score:** 3.5 → 4/5 confirmed with corrections
|
||||
|
||||
**Key points from challenge:**
|
||||
|
||||
1. **Score justified** — but "peer-reviewed" was overstated. Corrected to "arXiv preprint."
|
||||
2. **Blind review design** (phase 2 reviewers don't know if code is AI-assisted) = most important methodological detail, absent from initial eval. Added.
|
||||
3. **55.9% habitual users** more actionable than 30.7% median — validates learning curve section.
|
||||
4. **Limitations not flagged by the post**: tâches bornées en labo ≠ 12-month production codebase drift; potential selection bias (volunteer participants likely pro-AI); knowledge debt not measured.
|
||||
5. **Risk of non-integration**: Guide would retain pro-METR bias (AI slower on complex tasks) without empirical counter-balance on maintainability.
|
||||
|
||||
---
|
||||
|
||||
## Fact-Check
|
||||
|
||||
### LinkedIn Post Claims
|
||||
|
||||
| Claim (Olivier LOVERDE's post) | Verified | Source | Notes |
|
||||
|--------------------------------|----------|--------|-------|
|
||||
| Dave Farley = co-auteur de Continuous Delivery | ✅ | Perplexity, continuous-delivery.co.uk | Co-author with **Jez Humble** (not alone — minor omission in post) |
|
||||
| 151 développeurs | ✅ | arXiv abstract | Exact |
|
||||
| 95% professionnels | ✅ | arXiv abstract | Not mentioned in post but verified |
|
||||
| Un groupe crée, un autre reprend (sans IA) | ✅ | arXiv methodology | 2-phase blind design confirmed |
|
||||
| Code IA = aucun problème de maintenance | ✅ | arXiv abstract | "No systematic maintainability advantages or disadvantages" |
|
||||
| 30% de temps gagné | ✅ | arXiv: 30.7% median | Rounded, correct |
|
||||
| 50% pour ceux qui maîtrisent | ⚠️ | arXiv: 55.9% | Slight underestimate — actual is 55.9% |
|
||||
| Devs n'ont pas débranché leur cerveau (qualifier) | ✅ | Study design | Phase 1 participants guided AI, did not use autopilot |
|
||||
|
||||
### arXiv Paper Claims
|
||||
|
||||
| Claim | Verified | Source |
|
||||
|-------|----------|--------|
|
||||
| 30.7% median reduction in completion time | ✅ | WebFetch arXiv abstract |
|
||||
| 55.9% speedup for habitual users | ✅ | WebFetch arXiv abstract |
|
||||
| No significant differences in Phase 2 | ✅ | WebFetch arXiv findings |
|
||||
| 151 participants | ✅ | WebFetch arXiv abstract |
|
||||
| 95% professional developers | ✅ | WebFetch arXiv abstract |
|
||||
|
||||
**Corrections applied:**
|
||||
- "50%" → actual figure is **55.9%** (LinkedIn slight understatement)
|
||||
- "peer-reviewed" → **arXiv preprint** (July 2025, not yet peer-reviewed in proceedings)
|
||||
|
||||
**Confidence**: High — primary source directly fetched and cross-validated.
|
||||
|
||||
---
|
||||
|
||||
## Decision finale
|
||||
|
||||
- **Score final**: 4/5
|
||||
- **Action**: Intégrer (1-2 lignes dans section Productivity Research)
|
||||
- **Confiance**: Haute (primary source verified, methodology solid, gap confirmed)
|
||||
- **Nuance à conserver**: Limitations du design labo (tâches bornées, biais de sélection, knowledge debt non mesuré)
|
||||
|
||||
---
|
||||
|
||||
## Integration Log
|
||||
|
||||
**Date integrated**: 2026-02-19
|
||||
**Post-audit corrections applied**: 2026-02-19 (technical-writer audit + Perplexity v2 check)
|
||||
|
||||
| File | Change | Line |
|
||||
|------|--------|------|
|
||||
| `guide/learning-with-ai.md` | Citation réécriture — retrait claims éditoriaux, "July 2025" → "v2 Dec 2025", restructuration factuelle | ~926 |
|
||||
| `guide/ultimate-guide.md` | Ajout blockquote nuance downstream maintainability dans section 1.7 Trust Calibration | ~1092 |
|
||||
| `guide/learning-with-ai.md` | Ajout note "On maintainability fear" dans "Why Teams Get Results" | ~151 |
|
||||
| `machine-readable/reference.yaml` | Ajout 4 entrées: productivity_rct_metr, productivity_rct_echoes, productivity_maintainability_empirical, trust_calibration_maintainability_nuance | ~94 |
|
||||
|
||||
**Corrections post-audit:**
|
||||
- "peer-reviewed" → "arXiv preprint (v2 Dec 2025), not yet published in peer-reviewed proceedings" — Perplexity confirmé
|
||||
- Retrait formulation éditoriale "directly refutes the myth" → description factuelle neutre
|
||||
- Ajout "First RCT to explicitly target maintainability of AI-assisted code" (Perplexity: arXiv v2 HTML confirmed wording)
|
||||
- Séparation bibliographie / analyse : la comparaison METR déplacée dans le corps du guide, pas dans la biblio
|
||||
|
|
@ -150,6 +150,8 @@ The pattern: **AI excels at well-defined, repeatable tasks**. It struggles with
|
|||
|
||||
The difference isn't the tool — it's the organizational discipline around it.
|
||||
|
||||
**On maintainability fear**: The concern that AI-generated code creates unmaintainable codebases is not empirically supported — downstream developers show no significant difference in evolution time or code quality (Borg et al., 2025, n=151). The real risks are skill atrophy and over-delegation, not inherent quality degradation for the next developer. ([arXiv:2507.00788](https://arxiv.org/abs/2507.00788))
|
||||
|
||||
### Implications for Learning
|
||||
|
||||
This research shapes the rest of this guide:
|
||||
|
|
@ -923,6 +925,7 @@ Sources for [§3 The Reality of AI Productivity](#the-reality-of-ai-productivity
|
|||
- **Stack Overflow 2024: AI Sentiment** — [stackoverflow.co](https://stackoverflow.co/labs/developer-sentiment-ai-ml/) — Developer attitudes toward AI tools, productivity perceptions
|
||||
- **Uplevel Engineering Intelligence (2024)** — Burnout and productivity metrics with AI coding tools
|
||||
- **METR Experienced Developer RCT (2025)** — [arXiv:2507.09089](https://arxiv.org/abs/2507.09089) — Randomized controlled trial (16 experienced devs, 246 issues, repos 1M+ lines): AI tools made developers 19% slower on familiar codebases, despite perceiving themselves 20% faster (39-point perception gap). Strongest evidence for skill atrophy risk in experienced developers.
|
||||
- **Borg et al. "Echoes of AI" RCT (2025)** — [arXiv:2507.00788](https://arxiv.org/abs/2507.00788) — 2-phase blind RCT (151 participants, 95% professional developers): AI users 30.7% faster (median), habitual users ~55.9% faster. Phase 2: downstream developers evolving AI-generated code showed no significant difference in evolution time or code quality vs. human-generated code. First RCT to explicitly target maintainability of AI-assisted code. Co-authored by Dave Farley ("Continuous Delivery"). Note: arXiv preprint (v2 Dec 2025), not yet published in peer-reviewed proceedings.
|
||||
- **DORA/Google DevOps Research (2024)** — AI tool adoption impact on team performance
|
||||
|
||||
### Practitioner Perspectives
|
||||
|
|
|
|||
|
|
@ -1091,6 +1091,8 @@ Research consistently shows AI code has higher defect rates than human-written c
|
|||
|
||||
**Key insight**: AI produces code faster but verification becomes the bottleneck. The question isn't "does it work?" but "how do I know it works?"
|
||||
|
||||
> **Nuance on downstream maintainability**: A 2-phase blind RCT (Borg et al., 2025, n=151 professional developers) found no significant difference in the time needed for downstream developers to evolve AI-generated vs. human-generated code. The defect rates above are real — but they do not systematically translate into higher maintenance burden for the next developer. The risk is more narrowly scoped than commonly assumed. ([arXiv:2507.00788](https://arxiv.org/abs/2507.00788))
|
||||
|
||||
### The Verification Spectrum
|
||||
|
||||
Not all code needs the same scrutiny. Match verification effort to risk:
|
||||
|
|
|
|||
|
|
@ -91,6 +91,11 @@ deep_dive:
|
|||
learning_embracing_ai: "guide/learning-with-ai.md:518"
|
||||
learning_30day_plan: "guide/learning-with-ai.md:710"
|
||||
learning_red_flags: "guide/learning-with-ai.md:770"
|
||||
# Productivity Research RCTs
|
||||
productivity_rct_metr: "guide/learning-with-ai.md:925" # METR 2025: experienced devs 19% slower on large codebases despite perceiving 20% faster
|
||||
productivity_rct_echoes: "guide/learning-with-ai.md:926" # Borg 2025: 30.7% faster (median), ~55.9% habitual users, no maintainability impact downstream
|
||||
productivity_maintainability_empirical: "guide/learning-with-ai.md:926" # Empirical data on "AI code is unmaintainable" claim — blind RCT shows no significant difference
|
||||
trust_calibration_maintainability_nuance: "guide/ultimate-guide.md:1092" # Nuance: defect rates ≠ maintenance burden (Borg et al. 2025)
|
||||
learning_mode_template: "examples/claude-md/learning-mode.md"
|
||||
learn_quiz_command: "examples/commands/learn/quiz.md"
|
||||
learn_teach_command: "examples/commands/learn/teach.md"
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue