The AI Code Debt Bomb is About to Explode

An AI assistant introduced a crash-on-contact bug into a production repo — and nobody fixed it for 42 days, because the author was a chat session that no longer exists. This is the forensic story of what AI-generated code becomes six months after you ship it.

A longitudinal study out of Singapore Management University followed 300,000+ AI-authored commits across 6,000+ repositories and tracked every issue forward in time. The verdict: more than 1 in 5 AI-introduced issues are never fixed, security issues survive at nearly twice the rate of cosmetic ones, and a measurable slice of AI code ships pre-aged — pinned to superseded versions and hallucinated packages. Three forensic exhibits: the debt that won't die, comprehension debt, and dependencies that were born old. The honest counter-evidence is in here too — because this channel does forensics, not fashion.

☕ Support Macro Lens on Patreon → https://www.patreon.com/c/MacroLens
Independent, ad-light deep-dives. Patreon keeps them that way.

━━━━━━━━━━━━━━━━━━
RELATED VIDEO
━━━━━━━━━━━━━━━━━━
This video was the autopsy. If you want the system that keeps your code off the table — how to work with AI assistants so what you ship today isn't a nightmare in month six — watch:
▶ Stop Vibe Coding: The Right System for AI-Assisted Development

━━━━━━━━━━━━━━━━━━
CHAPTERS
━━━━━━━━━━━━━━━━━━
0:00 The 42-day bug nobody owned
0:27 Following AI code home: the 3 exhibits
1:14 Why software has a half-life
1:37 Exhibit 01: The debt that won't die
2:53 Exhibit 02: The readers who never existed
4:20 Exhibit 03: Born old
5:09 Slopsquatting: when hallucination becomes an attack surface
6:07 The invoice: throughput vs. stability (DORA)
6:23 The number that cuts the other way
6:47 It doesn't die faster. It dies alone.
7:11 The fix: stop shipping orphaned code

━━━━━━━━━━━━━━━━━━
CONCEPTS & SOURCES MENTIONED
━━━━━━━━━━━━━━━━━━
- Singapore Management University longitudinal study (300K+ AI-authored commits, 6K+ repos) — the spine of this video
- GitClear — 211M changed lines analyzed (2020–2024): moved/refactored code down, copy-paste up, duplicated blocks 8x
- Anthropic study — engineers learning with AI scored ~17 points lower on comprehension
- Comprehension debt — term coined by Addy Osmani
- USENIX Security 2025 — package hallucination: ~20% of LLM-recommended packages don't exist
- Slopsquatting — the supply-chain attack surface created by hallucinated package names
- Veracode GenAI Code Security Report — models chose the insecure option ~45% of the time
- DORA report — AI adoption raises delivery throughput, degrades delivery stability
- Erik Bernhardsson, "Git of Theseus" — the ~3.3-year half-life of a typical codebase

(Sources are named for transparency; look them up directly — this description contains no affiliate or tracking links.)

━━━━━━━━━━━━━━━━━━
FAQ
━━━━━━━━━━━━━━━━━━
What is the half-life of AI-generated code? It isn't shorter than normal code — the classic codebase half-life is ~3.3 years. The problem isn't the decay curve; it's that AI code decays without a maintainer who understands it.

Why doesn't AI-generated technical debt get fixed? Fixing code requires comprehension, and the context often lived in a chat session that no longer exists. The Singapore study found more than 1 in 5 AI-introduced issues are never resolved.

What is comprehension debt? A term from Addy Osmani: code gets written faster than anyone reads or understands it, so the ability to maintain it erodes — the bill comes due around month six.

What is slopsquatting? An attack where someone registers a package name an AI repeatedly hallucinates and fills it with malware. A USENIX study found ~20% of AI-recommended packages don't exist, and 43% of those names recur consistently.

Is AI-generated code worse than human code? Not uniformly. The same study found AI commits removed more code smells than they added. AI is an amplifier: it subtracts debt where mistakes are cheap and adds it where mistakes are expensive.

#AICoding #TechnicalDebt #SoftwareEngineering #CodeQuality #AICodeReview