Review Debt: AI Made Work Cheap. Trust Is Still Expensive.
September 17, 2026 by Asif Waliuddin
Machine-speed generation is creating a new enterprise liability: the widening gap between what organizations promote as trustworthy and what has actually been understood, challenged, and validated.
On July 18, 2026, one of my internal review-loop sensors told me an artifact was grounded. The status was green, the workflow had completed, and the system had done exactly what it was designed to do. Then I looked underneath it.
The core evidence ledger was empty.
Nothing dramatic had happened. No adversary broke in, no model produced a spectacular hallucination, no one waved off a flashing red warning. The warning was green, which is the more dangerous failure mode, because green is the state everyone downstream trusts without looking again. The workflow had completed. The evidence had not.
I changed the behavior so empty evidence now fails closed. But the fix mattered less than what the incident exposed: a category error that is becoming structural inside AI-enabled organizations. I had been treating movement through a workflow as if it were movement toward justified trust, and I had built a sensor sophisticated enough to make that mistake for me.
AI has made it extraordinarily cheap to create work. It has not made judgment cheap in the same way, and that gap is what I mean by review debt.
Review debt is not technical debt, though the word invites the comparison. Technical debt is a compromise you make on purpose: ship faster now, pay the repair cost later. Review debt is quieter, and it is this: the growing distance between what an organization promotes as trustworthy and what a qualified reviewer, or an independent verifier, has actually understood, challenged, and validated. That distance matters because modern AI systems produce far more than drafts. They produce code, specifications, decisions, incident analyses, and memory artifacts that become inputs to other systems almost immediately. A draft nobody has reviewed is at least honest about its status; a draft marked "approved" without real validation is worse, because the debt has acquired authority. Once it lands in a codebase, a knowledge base, or an agent's memory, whoever meets it later stops treating it as a hypothesis and starts treating it as context. And context is where last week's uncertainty quietly becomes this week's assumption.
What the public numbers actually measure
I want to be careful here, because this claim is easy to overstate. There is no clean public dataset for "trust." What exists is GitHub's 2025 Octoverse report, which counts commits, pull requests created, comments on issues and pull requests, and comments on commits. Those are activity signals. None of them measures whether anyone understood the work.
The counts are striking. GitHub reported more than 986 million commits in 2025, up about 25 percent year over year, with pull requests created rising about 20 percent. Over the same period, comments on issues and pull requests stayed essentially flat, and comments on commits fell 27 percent. GitHub is explicit that these are observational signals rather than causal claims, and that more work is needed to understand what AI is doing to the software workflow. I take that caveat seriously, and I am about to lean on it. (GitHub Octoverse 2025)
Here is what I read into it, as my inference and not as something the data proves. Production rose sharply while the commentary that usually accompanies scrutiny did not rise with it. Comments are an imperfect stand-in for review attention; people review without commenting, and comment without reviewing. But of the four signals, comments on commits sit closest to "someone stopped and examined this specific change," and it is the one that fell hardest. When output climbs by a quarter and the nearest proxy for scrutiny drops by more than a quarter, I doubt the two curves are still tracking each other.
There are innocent readings. Maybe tooling absorbed the commentary that once lived in comment threads, or review simply moved somewhere the report cannot see. My hypothesis, and it is only a hypothesis, is simpler: the bottleneck did not vanish when generation got cheap, it moved. The expensive step used to be writing the work; now it is deciding whether the work has earned trust, and that did not get cheaper because the writing did.
A study of 860 Microsoft developers, published in April 2026, reached a compatible picture from another direction. As code generation accelerated, the developers surveyed asked for AI systems that push quality signals earlier in the workflow. The guardrails they wanted were specific: explicit authority scoping, provenance, uncertainty signaling, and least-privilege access. The authors named the pattern bounded delegation, letting AI absorb the assembly work around a craft while keeping its authority precise and visible. (arXiv 2604.07830, April 2026) That is an operating model for using much more of AI without pretending verification came free.
There are four kinds of review debt
Queue debt
The obvious form: work waits for qualified validation faster than the organization can absorb it. What makes it dangerous is the aging. Waiting work grows stale while more keeps arriving, until reviewers stop asking what deserves deep inspection and start asking what they can survive before lunch.
Bypass debt
Work that reaches a trusted state without the validation that state implies. A document slides from draft to reviewed to approved to published while the evidence beneath it never changes. A workflow can carry an artifact through five states without moving it a step closer to being true.
Provenance debt
This shows up when an organization cannot reliably answer basic questions about an artifact: who or what produced it, from which inputs, under which model or agent, at what time, and on what evidence.
A preprint from June 23, 2026 shows how hard that gets even in public repositories. Researchers who analyzed more than 180 million repositories found that no single detection method captured more than a fraction of coding-agent activity. In one snapshot, a bot-account lookup recovered just 3.3 percent of the Claude Code commits their multi-method approach identified, a thirtyfold recall gap. An independent pull-request census missed 79 percent of the Claude Code adopters that commit-level detection found, and essentially all of the Codex adopters in the same comparison. (arXiv 2606.24429, June 2026)
The paper is about attribution, not review quality, and I will not borrow more from it than it earns. But the consequence is immediate: if you cannot reliably see which work was generated or mediated by agents, you have no trustworthy denominator for measuring review debt at all. You cannot govern what your telemetry does not know exists.
Comprehension debt
The most deceptive form, because it survives a completed review. Someone opened the pull request, scanned the document, and clicked approve; the artifact satisfied the workflow. Whether anyone formed a defensible understanding of the change, tested its assumptions, or checked the evidence beneath it is a separate question the green checkmark does not answer. The real test is whether the approver, or an independent verifier, could explain why the artifact is safe to promote. A review event and review comprehension are not the same thing, and the gap between them is where the honor system lives.
Why the dashboard looks fine
Most enterprise dashboards are very good at counting activity: artifacts generated, tickets closed, pull requests merged, cycle time reduced, tokens consumed, agent runs completed, automation rate climbing. Those are useful operating signals. They are just not evidence that the work is understood or trustworthy.
A system can push several of those numbers the right way while review debt gets worse underneath them. Faster cycle time might mean better engineering, or reviewers who stopped pushing back. More merged work might mean more value shipped, or more unexamined state entering the system. Higher agent utilization might mean leverage, or a production rate I lack the capacity to validate. The dashboard cannot tell me which reading is right unless I instrument the trust layer itself, and few teams I have seen do.
Three metrics I would stand up now
These three measure the distance between finishing a workflow and being confident in the result.
- Review queue depth and age. How much work is waiting for qualified validation, and for how long, segmented by consequence and destination. A two-day-old draft is not the same liability as a two-day-old artifact about to enter shared memory or production.
- Unreviewed promotion rate. The share of artifacts that cross a trust boundary without the independent validation that boundary requires. The boundary might be a merge, a publication, an entry into shared or agent memory, a customer delivery, or an executive decision record. Where it sits varies by organization; every organization has one.
- Provenance-unknown ratio. How much promoted work cannot say who or what created it, from which inputs, under which authority, and on what evidence. Where provenance is missing, the artifact should not inherit the trust of one whose production and validation path can be reconstructed.
None of these is a vanity metric. Each measures how far the work has traveled past the point where anyone could still justify trusting it.
The answer is not "put a human on everything"
Human review does not scale to machine-speed output, and it is not sacred just because it is human. People skim, they fatigue, they anchor on confident prose and click approve. The goal is to make trust promotion proportional to consequence and backed by independent evidence. Nobody needs a person to eyeball every token a machine emits.
Different artifacts earn trust in different ways. Some can be verified automatically against deterministic tests. Some need an independent model reasoning from genuinely separate context. Some call for provenance checks, policy checks, evidence-entailment checks, or a controlled adversarial pass. And some deserve slow, expensive expert attention, because the cost of being wrong is large and hard to reverse. The unit of design I care about is evidence in the loop.
Practice, not product
I run this on my own portfolio, so I will be specific about what I have and what I do not. Across bounded parts of my internal operating system, I can watch queue state, age, stalls, failures, and provenance. What I do not yet have is a single portfolio-wide, risk-weighted review-debt number that tells me the whole truth, and I would rather say that plainly than imply the problem is solved.
The July 18 sensor is why I stay cautious about my own claims. It proved that even a system I built to watch review quality could promote its own green status as if that status were the evidence. So I made the smallest change that would have caught it: an empty evidence ledger no longer gets to call itself grounded. It is a tiny rule, and it stands in for a principle I keep relearning, that trust has to be earned by evidence and cannot be inherited from workflow state.
The next advantage is digestion
AI changed the economics of production. On my own portfolio I generate artifacts at a rate that would have taken a team a few years ago, and that leverage is real. But an organization can only digest so much of what it produces, and generation got cheap far faster than digestion did. When the two diverge, the place looks faster from the outside while its understanding of itself thins, because more moves through it than anyone genuinely knows.
The bill arrives later, and somewhere specific: a codebase nobody wants to touch, a memory store seeded with things no one checked, an approval no one can reconstruct, a team afraid to change code it no longer owns. Review debt is what accumulates in the quiet before any of that shows.
The organizations that handle this well will draw explicit promotion boundaries, keep provenance, automate what can be independently verified, spend scarce expert attention on the decisions that are genuinely consequential, and measure the debt while it is still small enough to pay down. I am somewhere in the middle of building that myself, one fail-closed sensor at a time. What I am sure of is what I learned on July 18: a green status is a claim, not a proof, and the moment I let synthetic work become organizational memory without evidence, I have stopped running the system and started trusting it to run itself.
This article accompanies Episode 2 of AI Unveiled (The Honor System): "Review Debt: The Bottleneck Your Dashboard Is Hiding."