By May 2026, Anthropic Says Claude Authored Most of Its Merged Code. Count the Review Debt

Most written, not necessarily understood

One number in Anthropic’s account of AI helping to build AI deserves a pause. The company says that, as of May 2026, Claude authored more than 80 percent of the code merged into its codebase. “Merged” means accepted, after review, into the shared collection of software the company uses. People still supplied goals, redirected work and approved changes. The demonstrated change is narrower but substantial: Anthropic reports that most accepted code was generated by Claude rather than typed by an engineer. ^[2]

On November 2, reporting on Anthropic’s revised corporate filing highlighted the company’s warning that AI-assisted research could accelerate model development faster than corresponding safety work. Anthropic said no current model had crossed its threshold for dramatically accelerated progress attributable to AI. The filing story and the company’s separate technical account point to the same present tension: the work is speeding up before the organisation claims the loop has become autonomous. ^{[1] [2]}

Anthropic also says its typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024. It openly cautions that lines of code measure quantity rather than value and probably exaggerate the true productivity gain. The supplied account does not disclose enough methodology to independently assess how authorship was divided when Claude generated code that people prompted, redirected, edited and reviewed. The figure is important evidence, but it remains the company’s measurement of its own system. ^[2]

Here is the small contradiction: code can be accepted without being fully absorbed. Approved is not identical to understood. What interests me is the liability hidden between those words. I call it review debt—the accumulated distance between what an organisation operates and what its people can still explain. Like financial debt, it can fund useful acceleration. Also like financial debt, it becomes awkward when nobody can find the statement. ^[2]

The review-debt ledger

Anthropic says human code review has already become a bottleneck. It now uses an automated Claude reviewer to inspect proposed changes for bugs and security flaws. In a retrospective analysis, the company estimated that this reviewer would have caught roughly one-third of the bugs behind previous incidents on Claude’s website. That does not prove it will prevent one-third of future incidents, but it does make a serious case for AI-assisted checking. ^[2]

A bottleneck is normally treated as an obstruction to remove. Here it may also be a point of responsibility worth preserving. Review takes time because comprehension takes time. If production expands while the time available for understanding remains fixed, scrutiny does not become efficient; some of it simply goes missing. The software may work today while the organisation gradually loses the knowledge required to modify it tomorrow.

The same pattern can appear beyond software. A teacher might receive thirty generated lesson plans, a manager fifty summaries, a council hundreds of classified submissions. Cheap production makes abundance resemble efficiency. Yet consequential outputs still need someone able to spot a false premise, reject a polished mistake and accept responsibility. The hidden cost lands on the final human, who is asked to approve at machine speed while remaining liable at human speed.

When AI checks AI

The counterargument is strong. Requiring people to inspect every line manually may preserve ceremony rather than safety. Anthropic reports that Claude completed more than 800 fixes reducing a class of API errors—failures in the software connections through which services exchange requests. It also reports large gains on experiments with fixed goals and correctness tests. Automated checking can catch mistakes that capable people miss, and refusing it merely because another AI performed the first draft would waste a useful tool. ^[2]

The interesting boundary is independence. Anthropic says one measure of whether Claude Code completed a session successfully is itself judged by a Claude model. That metric is separate from the claim that Claude authored more than 80 percent of merged code, but it exposes the circularity problem neatly. If one model does the work and another related model scores it, the result may still be useful. It is not automatically an outside check. ^[2]

I think AI checking AI is sensible when the result can be tested against something beyond either model: software tests with known conditions, source records, reproducible calculations or a person with enough time and authority to disagree. Agreement between systems is not independence when their blind spots may rhyme. Two confident answers can still make one shared mistake wearing a spare hat.

Count what comes back

A useful review-debt ledger should compare like with like. Over the same month or quarter, track the growth in AI-assisted output, the time spent reviewing consequential work and the rate of corrections after approval. Then sample whether responsible people can reconstruct important sources, assumptions and decisions. Raw review minutes do not prove quality; the revealing pattern is output climbing while review capacity, understanding or correction performance deteriorates.

Most importantly, check whether reviewers can reject generated work without being punished for lowering a productivity score. A review process without permission to say no is quality theatre, with slightly better stationery. These safeguards impose a real cost, particularly where AI can clear neglected maintenance. But visible review costs now may be cheaper than correction costs arriving later, after the people who understood the system have stopped understanding it. ^[2]

Anthropic’s claim that Claude authored most of its merged code by May does not show that people disappeared from development. It suggests their scarce contribution is moving—from producing each part toward understanding the whole. That could be an excellent trade, if organisations protect the time and authority required for it. Ask not only how much work AI completed, but how much someone can still challenge, explain and repair. Production arrives quickly. Understanding still sends an invoice. ^[2]