Software understanding in the AI era
Insights from 92 software engineers exploring how AI is changing software development, code understanding, and engineering workflows.
AI is changing how software is built. We surveyed engineers to understand how they experience that transition, and where new challenges are emerging as software creation accelerates.
92 respondents · fielded 7–20 Jul 2026 · 15 questions
About the survey
The goal was to understand how engineers experience AI-assisted development, and the challenges emerging as software creation accelerates. We fielded 15 questions. 3 of them open-ended, to 92 software engineers between 7–20 Jul 2026.
This sample skews toward students and early-career engineers, and they make up 77.2% of respondents. We’re presenting the findings with that skew in view rather than claiming they represent the profession as a whole.
Key findings
Have shipped code they didn’t fully understand
56 of 92 respondents said yes. 18 frequently, 38 occasionally. Not a single respondent said no.
Comprehension gaps aren’t an edge case in this sample. They’re closer to the norm.
Reach for AI first in an unfamiliar codebase
AI tools edge out reading documentation (53.3%) and exploring code manually (50.0%). Asking a teammate comes a distant fourth, at 27.2%.
AI has become the default entry point into unfamiliar code, ahead of every other option, including the humans on the team.
Say AI has made them understand code less deeply
At the same time, 39.1% report writing much more code with AI, and 47.8% say they now review code more carefully.
Output and review activity are both up, but depth of understanding is down for roughly a third of respondents, a tension worth designing for.
Would trust AI-generated code without review
27 respondents (29.3%) said they don’t trust it. The majority, 58 respondents, say it depends on context, or “sometimes.”
Adoption is nearly universal; blind trust is rare. Most of the real risk sits in that gap.
Want to try a codebase-understanding tool
61 of 92 said yes, 10 said maybe or conditional, and 19 said no.
A self-selected sample, but a hard-no rate near one in five is a meaningful signal of unmet need.
Main insights
01Trust in AI-generated code is conditional, not blind
Very few respondents extend blanket trust to AI output (7.6%), and just as few refuse it outright (29.3%). The prevailing stance, 58 respondents well over half the sample, is that trust depends on context. The gap that matters most is between how often people reach for AI first (67.4%) and how often they’d trust its output without checking it (7.6%).
02The hardest part is architecture and intent, not syntax
Asked what’s hardest about working in unfamiliar code, respondents overwhelmingly named file/function relationships (26.1%), architecture and flow (21.7%), and understanding someone else’s logic or intent (14.1%). Setup, tooling, and language mechanics barely registered by comparison.
03Documentation is both the top complaint and the top request
Asked what they’d change, the most common answers were an AI-powered navigator that explains code with evidence (32.6%), better or automated documentation (26.1%), and visual maps of how files and modules relate (21.7%). Respondents consistently want the “why,” not just the “what.”
04The comprehension gap widens as AI usage grows
Only 5.4% of respondents say they always understand a system before changing it; 58.7% sit at “sometimes” or worse. Meanwhile 39.1% report writing much more code with AI’s help, more output without a matching rise in confidence.
05Workflows are shifting around AI, not away from review
47.8% say they now review code more carefully, the largest single behavior shift respondents report after adopting AI tools more heavily. Teams appear to be compensating for a lower baseline of understanding with more scrutiny at review time, rather than skipping it.
Developer perspectives
Answers to the survey’s open-ended questions, unedited and attributed by role only.
“Understanding the language and all the relationships between the files”
: Student / Intern
“Not understanding the way someone chose to solve a problem”
: Student / Intern
“Lack of familiarity with the codebase and tech stack”
: Senior Engineer (5+ years)
“Reckless use of AI (mostly bad prompting combined with superficial revision of the code)”
: Student / Intern
“Interactive overview of the codebase showing architecture, module relationships, data flow, dependencies. Explain why design decisions were made.”
: Senior Engineer (5+ years)
“Understanding the codebase and structure of development”
: Software Engineer (2–5 years)
Why Crux exists
AI has accelerated how fast software gets written. The next challenge is helping teams understand, maintain, and trust the software they ship, the same gap this research set out to measure.
Crux closes that gap by turning real repositories into short interactive courses, and verifying through active recall that developers actually understand the code behind what shipped, a living understanding layer for engineering teams building alongside AI.
Gift card draw
Continue following the research
We’re continuing to study how engineering teams adapt to AI-assisted development, and building Crux to help them hold onto their understanding as that shift accelerates.