Research
The evidence behind drift.
The evidence behind drift.
ReWeaver-driven and independent research on AI-assisted development: how drift forms, what it costs, and why human review still matters. Curated and updated as the field moves.
ORIGINAL RESEARCH
ORIGINAL RESEARCH
ORIGINAL RESEARCH
AI involvement and code drift: A provenance-labeled comparison of drift density and severity across production-readiness dimensions.
AI involvement and code drift: A provenance-labeled comparison of drift density and severity across production-readiness dimensions.
AI involvement and code drift: A provenance-labeled comparison of drift density and severity across production-readiness dimensions.
ReWeaver AI, 2026
BACKGROUND: Prior work on drift in AI-generated code has largely measured freshly generated components in isolation, leaving open the question of what AI involvement does to real, shipping codebases. Because fresh generation (greenfield) and mature production (brownfield) code differ in age, review history, and accumulated debt, a like-for-like comparison requires real code on both sides of the ledger. This study samples repositories labeled by degree of AI involvement in advance, and scans them on identical instrumentation, so that drift is compared across provenance groups on real code, rather than between freshly generated and mature corpora.
METHODS: Twenty-one open-source repositories, selected by GitHub search rules with six added by hand, were labeled by authorship provenance before scanning into three groups: human-led (n=6), AI-assisted (human + AI; n=10), or AI-led (n=5). Inclusion in an AI group means that AI was used. Full or partial authorship could not be guaranteed. Using ReWeaver, a multi-dimensional deterministic code drift scanning engine, we scanned repositories at the application root (a full walk of up to 2,000 files per scan root). Analysis included whole-repository Production Drift Ratio (PDR; a measure of drift frequency weighted by remediation time) scores, and line-level drift density, as well as breakdowns on nine dimensions of production readiness: user experience, design consistency, accessibility, security & privacy, reliability, architecture, maintainability, testability, and AI code governance.
RESULTS: Both PDR and drift density rose monotonically with the degree of AI involvement (M PDR: human-led = 0.773; AI-assisted = 0.830; AI-led = 0.835). Drift density followed a similar pattern (human-led = 4.42%; AI-assisted = 6.39%; AI-led = 10.15%). Added drift in AI-led code concentrated in design-facing dimensions: User Experience (+0.34 PDR), Accessibility (+0.17), and Design Consistency (+0.10). Maintainability, the largest single source of drift in the human-led baseline, was roughly equivalent in PDR between tiers.
CONCLUSION: On real code compared on identical instrumentation, drift was found to grow with AI involvement, specifically where design intent lives.
BACKGROUND: Prior work on drift in AI-generated code has largely measured freshly generated components in isolation, leaving open the question of what AI involvement does to real, shipping codebases. Because fresh generation (greenfield) and mature production (brownfield) code differ in age, review history, and accumulated debt, a like-for-like comparison requires real code on both sides of the ledger. This study samples repositories labeled by degree of AI involvement in advance, and scans them on identical instrumentation, so that drift is compared across provenance groups on real code, rather than between freshly generated and mature corpora.
METHODS: Twenty-one open-source repositories, selected by GitHub search rules with six added by hand, were labeled by authorship provenance before scanning into three groups: human-led (n=6), AI-assisted (human + AI; n=10), or AI-led (n=5). Inclusion in an AI group means that AI was used. Full or partial authorship could not be guaranteed. Using ReWeaver, a multi-dimensional deterministic code drift scanning engine, we scanned repositories at the application root (a full walk of up to 2,000 files per scan root). Analysis included whole-repository Production Drift Ratio (PDR; a measure of drift frequency weighted by remediation time) scores, and line-level drift density, as well as breakdowns on nine dimensions of production readiness: user experience, design consistency, accessibility, security & privacy, reliability, architecture, maintainability, testability, and AI code governance.
RESULTS: Both PDR and drift density rose monotonically with the degree of AI involvement (M PDR: human-led = 0.773; AI-assisted = 0.830; AI-led = 0.835). Drift density followed a similar pattern (human-led = 4.42%; AI-assisted = 6.39%; AI-led = 10.15%). Added drift in AI-led code concentrated in design-facing dimensions: User Experience (+0.34 PDR), Accessibility (+0.17), and Design Consistency (+0.10). Maintainability, the largest single source of drift in the human-led baseline, was roughly equivalent in PDR between tiers.
CONCLUSION: On real code compared on identical instrumentation, drift was found to grow with AI involvement, specifically where design intent lives.
BACKGROUND: Prior work on drift in AI-generated code has largely measured freshly generated components in isolation, leaving open the question of what AI involvement does to real, shipping codebases. Because fresh generation (greenfield) and mature production (brownfield) code differ in age, review history, and accumulated debt, a like-for-like comparison requires real code on both sides of the ledger. This study samples repositories labeled by degree of AI involvement in advance, and scans them on identical instrumentation, so that drift is compared across provenance groups on real code, rather than between freshly generated and mature corpora.
METHODS: Twenty-one open-source repositories, selected by GitHub search rules with six added by hand, were labeled by authorship provenance before scanning into three groups: human-led (n=6), AI-assisted (human + AI; n=10), or AI-led (n=5). Inclusion in an AI group means that AI was used. Full or partial authorship could not be guaranteed. Using ReWeaver, a multi-dimensional deterministic code drift scanning engine, we scanned repositories at the application root (a full walk of up to 2,000 files per scan root). Analysis included whole-repository Production Drift Ratio (PDR; a measure of drift frequency weighted by remediation time) scores, and line-level drift density, as well as breakdowns on nine dimensions of production readiness: user experience, design consistency, accessibility, security & privacy, reliability, architecture, maintainability, testability, and AI code governance.
RESULTS: Both PDR and drift density rose monotonically with the degree of AI involvement (M PDR: human-led = 0.773; AI-assisted = 0.830; AI-led = 0.835). Drift density followed a similar pattern (human-led = 4.42%; AI-assisted = 6.39%; AI-led = 10.15%). Added drift in AI-led code concentrated in design-facing dimensions: User Experience (+0.34 PDR), Accessibility (+0.17), and Design Consistency (+0.10). Maintainability, the largest single source of drift in the human-led baseline, was roughly equivalent in PDR between tiers.
CONCLUSION: On real code compared on identical instrumentation, drift was found to grow with AI involvement, specifically where design intent lives.
Read more…
Measuring the frequency and severity of drift in human-authored and AI-generated code.
Measuring the frequency and severity of drift in human-authored and AI-generated code.
Measuring the frequency and severity of drift in human-authored and AI-generated code.
ReWeaver AI, 2026
BACKGROUND: Existing research on AI-generated code quality has focused primarily on aggregate metrics that measure how often drift occurs without accounting for its impact or remediation cost. This study argues that frequency alone is an insufficient and potentially misleading measure of production readiness risk.
METHODS: Using ReWeaver, a multi-dimensional deterministic code drift scanning engine, we compared five leading AI coding tools against a human-authored reference across eight dimensions of production readiness, calculating both drift frequency and the Production Drift Ratio (PDR), a severity-weighted metric that converts drift into estimated remediation cost.
RESULTS: Drift was present across all code tested. AI tools produced higher drift frequency than the human baseline on five of eight dimensions. However, differences were not statistically significant. PDR scores diverged substantially from frequency rankings across every dimension reaching statistical significance (z = −5.43, p < .001). The divergence was most pronounced in the Security & Privacy dimension, where AI tools produced three times the human drift frequency but a PDR 22x higher. No tool performed uniformly well across all dimensions.
CONCLUSION: Frequency tallies what went wrong. The PDR quantifies what it will cost to fix it.
BACKGROUND: Existing research on AI-generated code quality has focused primarily on aggregate metrics that measure how often drift occurs without accounting for its impact or remediation cost. This study argues that frequency alone is an insufficient and potentially misleading measure of production readiness risk.
METHODS: Using ReWeaver, a multi-dimensional deterministic code drift scanning engine, we compared five leading AI coding tools against a human-authored reference across eight dimensions of production readiness, calculating both drift frequency and the Production Drift Ratio (PDR), a severity-weighted metric that converts drift into estimated remediation cost.
RESULTS: Drift was present across all code tested. AI tools produced higher drift frequency than the human baseline on five of eight dimensions. However, differences were not statistically significant. PDR scores diverged substantially from frequency rankings across every dimension reaching statistical significance (z = −5.43, p < .001). The divergence was most pronounced in the Security & Privacy dimension, where AI tools produced three times the human drift frequency but a PDR 22x higher. No tool performed uniformly well across all dimensions.
CONCLUSION: Frequency tallies what went wrong. The PDR quantifies what it will cost to fix it.
BACKGROUND: Existing research on AI-generated code quality has focused primarily on aggregate metrics that measure how often drift occurs without accounting for its impact or remediation cost. This study argues that frequency alone is an insufficient and potentially misleading measure of production readiness risk.
METHODS: Using ReWeaver, a multi-dimensional deterministic code drift scanning engine, we compared five leading AI coding tools against a human-authored reference across eight dimensions of production readiness, calculating both drift frequency and the Production Drift Ratio (PDR), a severity-weighted metric that converts drift into estimated remediation cost.
RESULTS: Drift was present across all code tested. AI tools produced higher drift frequency than the human baseline on five of eight dimensions. However, differences were not statistically significant. PDR scores diverged substantially from frequency rankings across every dimension reaching statistical significance (z = −5.43, p < .001). The divergence was most pronounced in the Security & Privacy dimension, where AI tools produced three times the human drift frequency but a PDR 22x higher. No tool performed uniformly well across all dimensions.
CONCLUSION: Frequency tallies what went wrong. The PDR quantifies what it will cost to fix it.
Read more…
RELEVANT RESEARCH
RELEVANT RESEARCH
RELEVANT RESEARCH
Drift
The hermeneutics of drift: Mapping the human–AI–institutional collapse across the Drift Triangle
The hermeneutics of drift: Mapping the human–AI–institutional collapse across the Drift Triangle
Katz, 2025
Examines how meaning breaks down in human-AI communication before performance failures become visible, and introduces the Drift Triangle Model to describe that breakdown.
Examines how meaning breaks down in human-AI communication before performance failures become visible, and introduces the Drift Triangle Model to describe that breakdown.
ResearchGate →
Technical Debt
Beyond productivity: Evaluating the hidden costs of generative AI in software development
Beyond productivity: Evaluating the hidden costs of generative AI in software development
Anderson et al., 2025
A dynamic simulation model of AI-assisted coding over time finds an inverted-U productivity curve: a sweet spot where AI assistance helps, and a threshold beyond which technical debt undermines maintainability and profitability. Greenfield, fast-moving teams tolerate more AI assistance; legacy codebases and less experienced teams decline faster.
A dynamic simulation model of AI-assisted coding over time finds an inverted-U productivity curve: a sweet spot where AI assistance helps, and a threshold beyond which technical debt undermines maintainability and profitability. Greenfield, fast-moving teams tolerate more AI assistance; legacy codebases and less experienced teams decline faster.
SSRN →
Debt behind the AI boom: A large-scale empirical study of AI-generated code in the wild
Debt behind the AI boom: A large-scale empirical study of AI-generated code in the wild
Liu et al., 2026
Static analysis across 302.6k verified AI-authored commits in 6,299 GitHub repositories finds 484,366 distinct issues introduced, 89.3% of them code smells, with more than 15% of every tool's commits introducing at least one. 22.7% of tracked issues were still present at the latest revision.
Static analysis across 302.6k verified AI-authored commits in 6,299 GitHub repositories finds 484,366 distinct issues introduced, 89.3% of them code smells, with more than 15% of every tool's commits introducing at least one. 22.7% of tracked issues were still present at the latest revision.
arXiv →
Human in Control / Human in the Loop
Why human guidance matters in collaborative vibe coding
Why human guidance matters in collaborative vibe coding
Hu et al., 2026
Across 20 experiments and 737 participants, humans gave uniquely effective high-level instructions for vibe coding. AI-provided instructions often caused performance collapse; hybrid setups, humans instructing and AI evaluating output, performed best.
Across 20 experiments and 737 participants, humans gave uniquely effective high-level instructions for vibe coding. AI-provided instructions often caused performance collapse; hybrid setups, humans instructing and AI evaluating output, performed best.
arXiv →
Beyond autonomy: The non-zero probability of human necessity in AI problem-solving
Beyond autonomy: The non-zero probability of human necessity in AI problem-solving
Katz, 2025
Establishes that the probability of human necessity in AI problem-solving remains non-zero, underscoring the essential role of human-in-the-loop systems for AI reliability.
Establishes that the probability of human necessity in AI problem-solving remains non-zero, underscoring the essential role of human-in-the-loop systems for AI reliability.
ResearchGate →
State of the Field
2026 AI Engineering Report: The Acceleration Whiplash
2026 AI Engineering Report: The Acceleration Whiplash
Faros AI, 2026
Faros AI →
Gartner predicts 40% of enterprise apps will feature task-specific AI agents by 2026, up from less than 5% in 2025
Gartner predicts 40% of enterprise apps will feature task-specific AI agents by 2026, up from less than 5% in 2025
Gartner, 2025
Gartner →
AI and Productivity
Measuring the impact of early-2025 AI on experienced open-source developer productivity
Measuring the impact of early-2025 AI on experienced open-source developer productivity
Becker et al., 2025
METR →
The productivity-reliability paradox: Specification-driven governance for AI-augmented software development
The productivity-reliability paradox: Specification-driven governance for AI-augmented software development
Farrag, 2026
arXiv →
How much more productive are AI-powered developers? Large sample productivity data
How much more productive are AI-powered developers? Large sample productivity data
Harding, 2025; GitClear
GitClear →
Stanford study of 100,000 developers on engineering productivity
Stanford study of 100,000 developers on engineering productivity
Stanford Software Engineering Productivity Research Group, 2025
Stanford →
AI-assisted programming may decrease the productivity of experienced developers by
increasing maintenance burden
AI-assisted programming may decrease the productivity of experienced developers by increasing maintenance burden
AI-assisted programming may decrease the productivity of experienced developers by
increasing maintenance burden
Xu et al., 2025
arXiv →