AI coding productivity — the evidence

Confidence 0.80 · 6 sources · last confirmed 2026-08-30

What randomised and population-scale studies actually find when AI is put in front of software developers. The corpus contains two RCTs with opposite signs, and reconciling them is more useful than picking one.

The two randomised results

METR (2025)Cui et al. (2026)
DesignRCT, task-level randomisation3 field experiments, pre-registered
N16 developers, 246 tasks4,867 developers
PopulationExperienced OSS maintainers, ~5 years on their own reposMicrosoft, Accenture, a Fortune 100 firm
OutcomeCompletion timeCompleted tasks
Result+19% time — slower+26.08% tasks — faster (SE 10.3%)

These are compatible, and the reconciling variable is stated in Cui et al.’s own heterogeneity result: less experienced developers adopted more and gained more. METR studied the far tail of the experience distribution — maintainers who already hold the model of the system in their heads — which is exactly where this paper’s gradient predicts the smallest gain. The synthesis is a gradient claim: assistance is worth most where the marginal work is acquiring context, and least — possibly negative — where the developer already has it.

Two cautions before either number is quoted. Cui et al.’s standard error of 10.3% puts the 95% interval roughly from +6% to +46%: the finding is “positive and probably substantial,” not “26%.” And METR’s tooling vintage is February–June 2025 Cursor Pro + Claude 3.5/3.7 Sonnet in an IDE — not the cloud-agent parallelism the 2026 practitioner material is about.

The perception gap, which is the finding with the widest reach

METR’s participants forecast −24% before, estimated −20% after finishing, and were measured at +19%. After personally doing the work, they were wrong by ~39 points, in a consistent direction. Economics experts predicted −39%; ML experts −38%.

This is not a curiosity. It invalidates the instrument most organisations use. DORA 2025 reports that >80% of ~5,000 respondents believe AI increased their productivity — a belief METR shows to be systematically wrong for experienced developers. DORA’s delivery metrics are behavioural and survive this; its perception metrics should not be read as productivity evidence.

The population signal, and its split

DORA 2025: 90% AI adoption (up from 76% in 2024), 70% trust in AI-generated code — meaning 30% do not trust the tool they are using.

Relationship with AI adoption20242025
Delivery throughputnegative (~1.5% drop per 25% adoption rise)positive
Delivery stabilitynegative (~7.2% drop per 25% adoption rise)still negative

“AI adoption does continue to have a negative relationship with software delivery stability.”

Throughput turning positive means the authoring constraint has been relieved. Stability staying negative means the delivery system’s absorptive capacity — review, testing, rollback, observability — has not moved with it. DORA’s central thesis follows: “AI doesn’t fix a team; it amplifies what’s already there.”

What the productivity ledger leaves out

A completed-task count is a gross measure. Against it:

None of these refutes the productivity findings. They are different points in the same pipeline, and only the first one is flattering — which is why it is the one that gets cited.

The practical position

The defensible reading of the whole corpus: AI coding tools reliably increase output volume; whether they increase delivered value depends on the receiving system. That is DORA’s amplifier thesis, METR’s context gradient, and Carson’s own anti-thesis“I don’t get multiples of quality off of multiples of output” — all saying the same thing from three different methodologies.

automation-vs-augmentation, micro-productivity-trap, ai-generated-code-quality, agentic-pull-requests, jagged-frontier, ai-employment-effects, enterprise-ai-adoption.

Debates and supersession

  • METR vs Cui et al. — reconciled above as a context gradient. That reconciliation is this wiki’s inference; neither paper makes it. Open: a study that randomises across the experience distribution within one setting.
  • Vendor proximity. Two of Cui et al.’s three sites are Microsoft (which owns GitHub) and a large systems integrator. Randomisation and pre-registration mitigate this; they do not eliminate it.
  • No study measures net value. Every design here scores volume (tasks, time, throughput). Nobody has measured AI-assisted development against a value or defect-adjusted outcome. This is the largest gap in the concept.
  • Everything is pre-agentic. METR measures IDE assistance; Cui et al. measure Copilot. No RCT in the corpus measures cloud agent fleets, which is what the 2026 practitioner material is about.