AN INTERACTIVE RESEARCH NOTE · 2025 DATA
The AI coding
paradox.
Everybody is using it.
Nobody agrees what it does.
READING TIME05 MIN
SIGNALS COMPARED03
LAST REVIEWED10 AUG 2026
?
01 / ADOPTION
84%
of developers use or plan to use AI tools in their development process.
USAGE IS NO LONGER THE QUESTIONSO / 2025
AI crossed into the default workflow.
Adoption rose eight points in one year. Among professional developers, daily use alone reached 51%.
202476%
202584%
1 Stack Overflow Developer Survey, 33,662 responses to this question.
02 / CONFIDENCE
Use went up.
Use went up.
Confidence did not.
60%favorable
Positive sentiment fell from above 70% in 2023–24.
WHAT DEVELOPERS WORRY ABOUT
87%Accuracy
81%Security & privacy
Concern can coexist with habitual use. That tension is the first clue that adoption is a poor proxy for value.
2 Stack Overflow Developer Survey 2025. Concern questions: 28,930 responses.
03 / PERCEPTION
The felt
The felt
speed-up.
Self-reporting captures something real: flow, confidence, reduced friction. It does not automatically measure elapsed time or quality.
52%report a positive productivity effect
70%of agent users say specific tasks take less time
Survey signal
perceived effect
3 Stack Overflow Developer Survey 2025; the 70% figure is among respondents using AI agents.
04 / MEASUREMENT
Then somebody
Then somebody
started a clock.
RANDOMIZED CONTROLLED TRIAL · FEB–JUN 2025
METR ↗
OBSERVED+19%longer
With early-2025 AI tools, 16 experienced open-source developers took longer across 246 real issues in repositories they knew well.
FASTERSLOWER
Before tasks, developers expected−24%
Afterward, developers believed−20%
Measured implementation time+19%
Read narrowly. This was one population, working on mature codebases with early-2025 tools. It is evidence about that setting—not a universal verdict on AI coding.
05 / MACRO SIGNAL
Meanwhile,
Meanwhile,
the graph went up.
GitHub recorded its busiest year. The timing overlaps with agent adoption, but activity is not the same thing as productivity—and correlation is not attribution.
OUTPUT ≠ OUTCOMEOUTPUT ≠ OUTCOMEOUTPUT ≠ OUTCOMEOUTPUT ≠ OUTCOME
4 GitHub Octoverse 2025. GitHub describes correlations and does not claim AI caused the increases.
06 / THE 2026 UPDATE
Tools changed.
Tools changed.
The experiment got harder.
METR’s follow-up saw a directionally better result with late-2025 tools—but selection effects made the estimate unreliable.
ESTIMATED CHANGE IN COMPLETION TIME← FASTER SLOWER →
Returning developers
−18% (−38 to +9)
Newly recruited
−4% (−15 to +9)
Both intervals cross zero. Developers who did not want to work without AI increasingly opted out—exactly the kind of selection that can distort the answer.
5 METR methodology update, February 2026. Negative values indicate estimated speed-up.
07 / A BETTER QUESTION
Stop asking
Stop asking
“Does AI work?”
Ask where, for whom, and against which outcome.
- 01Cycle time
From accepted task to review-ready change.
- 02Rework
Review rounds, reverted changes and defects.
- 03Task shape
Greenfield, mature code, debugging or exploration.
- 04Human experience
Skill, context, fatigue and perceived flow.
THE TAKEAWAY
AI adoption is settled. AI productivity is a measurement problem.