Eight pages in this wiki name a kind of “debt” said to accumulate in AI-assisted or agentic software development: Verification Debt, Cognitive Debt, Comprehension Debt, Trust Debt, Governance Debt, Prompt Debt, Fast-Integration Debt and Provenance Debt. Four of them come from a single multivocal literature review, Faster Code, Deeper Debt? A Multivocal Literature Review on Technical Debt and Its Early Signs in LLM-Assisted Software Development; the other four come from a synthesis paper by Koch, talks and posts by Addy Osmani, a study of an undergraduate software engineering project, and the opening instalment of one author’s column. This page describes the shapes that recur across the eight and the places where they come apart. It does not rank the terms or decide which name is the right one for any of them.
The eight cases
| Term | Where the wiki records it | What is said to accumulate | Whose ledger it is on |
|---|---|---|---|
| Verification Debt | Koch, Agentic Agile-V: From Vibe Coding to Verified Engineering in Software and Hardware Development | weak tests, hidden regressions, broad patches, unvalidated dependencies, undocumented behaviour, increased reviewer burden | the team’s unverified output and review load |
| Cognitive Debt | Osmani’s “Own the Outer Loop” keynote; a separate post on parallel agents; Comprehension Debt - the hidden cost of AI generated code. | erosion of an engineer’s own understanding of how to solve a problem | the individual engineer |
| Comprehension Debt | Comprehension Debt in GenAI-Assisted Software Engineering Projects | the gap between what a team knows about its codebase and what it needs to understand to maintain it | the team’s collective cognition, explicitly not the code |
| Trust Debt | 停止“氛围编程”（Vibe Coding），拥抱新一代软件工程 | code shipped faster than anyone can establish it deserves trust | what nobody has established about the code |
| Governance Debt | Faster Code, Deeper Debt? A Multivocal Literature Review on Technical Debt and Its Early Signs in LLM-Assisted Software Development | a continuing oversight burden for incorporated LLM output, caused by hallucination and non-determinism | oversight that does not end at merge |
| Prompt Debt | Faster Code, Deeper Debt? A Multivocal Literature Review on Technical Debt and Its Early Signs in LLM-Assisted Software Development | unclear, sensitive or undocumented prompts that reduce reproducibility and code quality | prompts that shape code but are not kept like code |
| Fast-Integration Debt | Faster Code, Deeper Debt? A Multivocal Literature Review on Technical Debt and Its Early Signs in LLM-Assisted Software Development | LLM output integrated without proper validation | the gap between adoption speed and evaluation speed |
| Provenance Debt | Faster Code, Deeper Debt? A Multivocal Literature Review on Technical Debt and Its Early Signs in LLM-Assisted Software Development | unclear ownership or missing attribution of generated code | missing records, surfacing as legal or accountability questions |
The same review names two further new categories, data debt and ethical debt. Neither has a page of its own in this wiki, so they are not counted among the eight.
Recurring shapes
One rate outrunning another
Four of the eight pages describe the mechanism as a rate mismatch, with generation or adoption on the fast side:
- Verification Debt — teams accumulate it if output volume grows faster than verification capacity.
- Trust Debt — the speed of agent-generated code outruns the capacity to establish that the code deserves trust; once generation stops being the bottleneck, human attention and review bandwidth become the binding constraint.
- Fast-Integration Debt — the problem originates in the gap between how quickly LLM output can be adopted and how quickly a developer can evaluate its downstream consequences.
- Cognitive Debt — in the parallel-agents post, letting agents generate faster than a person can understand accumulates debt that compounds across threads. Comprehension Debt - the hidden cost of AI generated code. states the same asymmetry as an inversion: a junior engineer can now generate code faster than a senior engineer can critically audit it.
The other four are not framed as a race between two rates. Comprehension Debt attributes the gap to AI tools reducing cognitive load, so that understanding which would have been built while doing the work is not built. Governance Debt rests on properties of the model — hallucination and non-determinism — which make oversight continuing rather than a one-off cost. Prompt Debt names a mismatch in how two kinds of artefact are treated, and Provenance Debt a missing record of where code came from.
Defined against ordinary technical debt
Four of the grounding pages define a term partly by what separates it from conventional technical debt, and each draws the line in a different place:
- Comprehension Debt — it resides in the collective cognition of a team rather than in the codebase.
- Comprehension Debt - the hidden cost of AI generated code. — technical debt announces itself through mounting friction and is usually a conscious tradeoff with a known location; comprehension debt breeds false confidence, because the code looks clean and the tests are green.
- Fast-Integration Debt — the problem originates not in the quality of any particular artefact but in the gap between adoption and evaluation.
- Provenance Debt — code debt surfaces as maintenance friction, while this surfaces as a legal or accountability question.
Governance Debt draws a different line: it is scoped to the oversight burden of incorporated output, and deliberately separated from the existing literature on technical debt in AI and ML systems rather than from conventional technical debt.
A cost that arrives later
Four grounding pages — three of the eight term pages and Osmani’s post — stress that the cost is not visible when the code goes in:
- Fast-Integration Debt — a practitioner account in which velocity metrics looked excellent during a sprint and the bug reports arrived weeks later, in edge cases the AI had not considered.
- Governance Debt — its defining feature is that the burden does not end at merge.
- Provenance Debt — the cost is deferred to the point where someone needs to establish who is answerable for a piece of logic or under what terms it may be distributed.
- Comprehension Debt - the hidden cost of AI generated code. — the codebase looks clean and the tests are green right up until the reckoning arrives.
Stated as unmeasured
Four of the eight pages say outright that the debt they name is not measured:
- Verification Debt — Koch offers no way to measure it directly, and lists whether evidence bundles can reduce it as an open question.
- Trust Debt — one author’s framing, with no measurement of the debt and no method for observing it.
- Fast-Integration Debt — identified inductively from grey literature and resting on practitioner reporting rather than measurement.
- Governance Debt — inductively derived from practitioner sources rather than measured.
Faster Code, Deeper Debt? A Multivocal Literature Review on Technical Debt and Its Early Signs in LLM-Assisted Software Development states the general case: neither literature it surveyed contained standardised benchmarks or datasets for evaluating technical debt in LLM-generated code. Comprehension Debt - the hidden cost of AI generated code. makes a related claim from practice — velocity metrics, DORA metrics, PR counts and coverage can all look healthy while the comprehension deficit stays invisible.
The remaining pages report evidence of other kinds without addressing measurement of the debt itself: Cognitive Debt cites a randomized trial in which engineers who worked through AI scored 50% on a comprehension quiz against 67% for those who did not; Comprehension Debt describes four accumulation patterns observed in an undergraduate project; Prompt Debt reports a study in which well-structured, specific prompts mitigated up to 87.1% of observed code smells; and Provenance Debt rests on one formal and one grey source among the review’s 104.
Where the cases come apart
Two concepts called “comprehension debt”
The same name is attached to two different concepts. Osmani’s usage — the gap between how much code exists and how much of it any human understands — is recorded on Cognitive Debt: his keynote calls it cognitive debt, the parallel-agents post calls the same dynamic comprehension debt, and Comprehension Debt - the hidden cost of AI generated code. uses “comprehension debt” for it. Comprehension Debt is the term as proposed in the undergraduate-project paper: a team’s collective knowledge falling short of what maintaining the codebase requires. The first is framed around an engineer delegating to agents, the second around a team adopting generative AI tools. Comprehension Debt lists Cognitive Debt as a related term, and neither page treats the two as one concept.
Where the debt is held
The eight differ on whose ledger the debt is on. Cognitive Debt places it in one engineer and Comprehension Debt in a team. Verification Debt places it in output and review load, Prompt Debt in inputs that are not under the controls applied to code, and Provenance Debt in records that do not exist. Trust Debt places it in what nobody has established about the code, and Governance Debt in oversight that continues after the code is merged.
How strongly the terms are backed
The four terms from Faster Code, Deeper Debt? A Multivocal Literature Review on Technical Debt and Its Early Signs in LLM-Assisted Software Development come with source counts, and those counts differ widely: Governance Debt appears in 19 grey sources and no formal ones; Fast-Integration Debt in 13 grey sources and no formal ones; Prompt Debt in 8 formal and 5 grey sources, and is reported as the primary source of LLM-specific debt in the formal literature; Provenance Debt in one formal and one grey source. The other four terms each rest on one author’s or one study’s framing: a synthesis paper for Verification Debt, a keynote and posts for Cognitive Debt, an undergraduate-project study for Comprehension Debt, and a column’s opening instalment for Trust Debt.
One causal link
Only one causal link between terms is stated in the grounding pages. Faster Code, Deeper Debt? A Multivocal Literature Review on Technical Debt and Its Early Signs in LLM-Assisted Software Development describes a domino effect in which fast-integration practices trigger cascading governance risks if left unmanaged; as Fast-Integration Debt and Governance Debt relay it, fast integration creates more code requiring oversight, which becomes governance debt and then increased long-term maintenance cost. The other pages relate their terms to neighbours only as related concepts — Trust Debt, for instance, lists Verification Debt as the closely neighbouring cost of unverified output without stating a causal order.
What each source proposes in response
The responses attached to each term differ in kind, following where each places the debt:
- Fast-Integration Debt — a human-in-the-loop model, recommended in 58 of the review’s 73 grey sources: treating LLM output as drafts requiring review and refinement.
- Governance Debt — ongoing developer training, documenting AI involvement in codebases, and prioritising long-term maintainability, alongside code quality tooling and multi-layered review.
- Verification Debt — acceptance on evidence proportionate to risk, through the risk-adaptive gates and evidence bundles of Agentic Agile-V, and keeping tests inside the agent loop.
- Comprehension Debt — practices rather than code changes: verification practices, structured retrospectives and active learning assessments.
- Prompt Debt — making prompts durable: templates, structured prompts, prompt registries and versioned documentation.
- Provenance Debt — provenance review, asking where the code came from, alongside governance checks and prompt documentation.
- Trust Debt — trust engineering, with a three-dimensional bill of materials for the AI era named as a later subject of the column.
- Comprehension Debt - the hidden cost of AI generated code. — no substitute: the post argues that tests and specs each have a ceiling as replacements for understanding, and that the comprehension work is the job.
Related pages
Review Bottleneck, Skill Atrophy, Automation Bias, Cognitive Surrender, Orchestration Tax, Vibe Coding