> **来源：[研报客](https://pc.yanbaoke.cn)** # **State of Code Developer Survey Report Summary** ## **Core Content** This report, based on a survey of 1,149 professional developers, explores the current state of AI in software development, focusing on usage patterns, effectiveness, and the emerging challenges and opportunities. --- ## **Main Findings** ### **AI Adoption and Usage** - **AI is now a daily tool**: 72% of developers who have tried AI coding tools use them every day. - **AI-generated code is widespread**: 42% of developers' code is currently AI-assisted or generated, with predictions that this will rise to over 50% by 2027. - **AI is used across all project types**: From prototypes (88%) to mission-critical services (58%), developers are integrating AI into various stages of development. - **Top AI tools**: GitHub Copilot (75%) and ChatGPT (74%) are the most widely used AI coding tools, followed by Claude (48%), Gemini/Duet AI (37%), and Cursor (31%). ### **Effectiveness vs. Adoption** - A **gap exists between AI usage and effectiveness**: While AI is used frequently, developers report that it is only **effectively** used for specific tasks like writing documentation (74% effective), explaining code (66% effective), and generating tests (59% effective). - **Code review and debugging are less effective**: Only 47% and 44% of developers find AI effective for these tasks, respectively. ### **Developer Trust in AI** - **High skepticism**: 96% of developers do not fully trust that AI-generated code is functionally correct. - **Verification is critical**: Only 48% of developers always check their AI-assisted code before committing, indicating a **verification bottleneck**. - **Effort required for quality**: 61% of developers agree that getting good code from AI requires significant effort in prompting and fixing. ### **Skills in the AI Era** - **Reviewing and validating AI code** is the most important new skill for developers (47%). - **Efficient prompting** is also highly valued (42%). - Other important skills include translating domain knowledge into code requirements (27%), architecting systems (25%), and identifying security risks (24%). ### **Tool Access and Adoption** - **Shadow adoption is common**: Over 50% of developers use ChatGPT through personal accounts, and 63% use Perplexity that way. - **Formal adoption varies**: GitHub Copilot and Amazon Q Developer are more likely to be used through work-sanctioned accounts (78% and 72% respectively). - **Tool choice differs by company size**: SMBs prefer ChatGPT, Claude, and JetBrains, while mid-sized and enterprise developers lean toward infrastructure and deployment automation. --- ## **AI Agents and Their Role** ### **Agentic AI is gaining traction** - **64% of developers** have started using AI agents, with 25% using them regularly. - AI agents are most effective for: - **Creating code documentation** (68%) - **Automated test generation and execution** (61%) - **Automated code review** (57%) ### **Less common use cases** - **Security vulnerability patching** (28%) - **Automated debugging** (44%) - **Deployment pipeline management** (33%) --- ## **The New Developer Toil** ### **AI reduces toil, but does not eliminate it** - **75% of developers** believe AI reduces the amount of time spent on toil work. - However, developers still spend **23-25% of their workweek** on toil tasks, such as managing technical debt and debugging legacy code. - **Toil is shifting**: Less frequent AI users face traditional toil (e.g., understanding code), while frequent users encounter new toil (e.g., managing technical debt). --- ## **AI’s Impact on Key Metrics** | Impact Area | Positive Impact (%) | |---------------------------|---------------------| | Developer productivity | 89% | | Time-to-market | 70% | | Feature/fix release frequency | 60% | | Code quality | 58% | | Code maintainability | 56% | | End-user experience | 47% | | Technical debt | 47% | | Rework / patch costs | 42% | | Defect rates | 39% | | Vulnerability rates | 34% | | Outage frequency | 25% | | Outage severity | 24% | --- ## **Key Takeaways** - AI is a **daily part of the development workflow**, but **effectiveness is uneven**. - **Verification is a major bottleneck**, with developers spending significant time reviewing and fixing AI-generated code. - **Trust remains low**, especially for code that is critical or mission-focused. - **AI agents** are beginning to automate key tasks like documentation and testing, but not yet complex maintenance or security tasks. - **Tool sprawl** is growing, with developers using **multiple AI tools** daily, often through **personal accounts**, creating risks in terms of security and compliance. - The **shift in toil** suggests that AI is not reducing overall workload but **changing its nature**, requiring new skills and approaches in code review and maintenance. --- ## **Recommendations for Engineering Leaders** - Implement **systematic verification processes** to ensure code quality and security. - Address the **"bring your own AI" (BYOAI)** culture by promoting governance and secure tool access. - Focus on **training developers** in reviewing and validating AI-generated code. - Monitor and support **tool adoption trends**, especially for smaller organizations and junior developers.