Generative AI has entered developers' daily workflows, but 'using AI' and 'trusting AI output' are developing at different paces. This report primarily draws on public data from the Stack Overflow 2025 Developer Survey, Google Cloud / DORA 2024 Report, DORA 2024 AI Preview, DORA 2025 Report, and the Sonar Developer Survey Report to observe how developers use AI for code writing, explanation, documentation, technical search, and code review, and why these outputs still require human confirmation. It should be noted that DORA 2025 is primarily used in this research corpus for trend judgments such as 'AI as an organizational capability multiplier,' and therefore is utilized to explain the direction of change, without additionally citing quantitative data not fully disclosed in the reviewed materials.
Research Background
Over the past two years, the discussion around developer use of generative AI has gradually shifted from 'whether to use' to 'how to use.' Code completion, technical Q&A, documentation generation, code explanation, and information summarization have become common use cases. However, software development differs significantly from general content generation: whether a piece of code is usable depends not only on syntactic correctness but also on business context, dependency environments, test results, security requirements, and deployment accountability. Therefore, as AI becomes increasingly embedded in development workflows, the new questions are no longer just about generation speed, but about who verifies the generated results, how they are verified, and whether local efficiency gains can truly translate into stable software delivery performance. I observe that several public surveys from 2024–2025 are collectively revealing a change: AI usage frequency continues to rise, but human judgment has not exited the workflow as a result.
AI Has Entered Daily Workflows, but Adoption Growth Has Not Translated into Higher Trust
The scope of developer use of AI continues to expand. According to the Stack Overflow 2025 Developer Survey, the proportion of respondents currently using or planning to use AI tools rose from 76% the previous year to 84% in 2025; among professional developers, 51% report using AI tools daily. This indicates that AI is no longer just a supplementary tool occasionally tried by some developers, but is increasingly part of high-frequency work scenarios.
However, increased adoption has not brought a corresponding improvement in attitudes. The same survey shows that the proportion of developers with a positive attitude towards AI tools dropped from over 70% in 2023 and 2024 to 60% in 2025. This set of changes illustrates that 'willingness to use' and 'willingness to directly accept output' are two different questions. Developers can use AI frequently while remaining cautious about the results.
This discrepancy is related to the nature of software development tasks. Developers may ask AI to explain unfamiliar code, organize technical materials, draft documentation, or quickly generate an implementation approach, but these scenarios do not imply that final judgment authority has been delegated to AI. Especially when output is intended for the codebase, production environment, or impacts system behavior, developers need to re-examine context, dependencies, edge cases, and potential risks. AI, therefore, is more like a high-frequency information and generation tool rather than a decision-making entity that automatically earns trust.
What Truly Increases Verification Costs Are the Many 'Almost Right but Not Quite' Outputs
A direct reason human verification persists is that AI output issues are often not obvious errors but rather superficially plausible, locally correct content that deviates in details or context. According to the Stack Overflow 2025 Developer Survey, 46% of developers do not trust the accuracy of AI tool output, higher than the 33% who do; only 3% report a high level of trust.
This trust gap is also evident in developers' actual frustrations. The 2025 survey shows that 66% of developers cited 'AI solutions were almost right, but not quite correct' as a major frustration, while another 45.2% stated that debugging AI-generated code takes more time. The key issue reflected here is not whether AI can generate answers, but how much verification work remains before the generated results are 'ready for production.'
Completely wrong code is often easily exposed by tests or runtime results, whereas almost-right output requires developers to make further judgments: whether the interface matches the current version, whether exception handling is complete, whether the logic misses edge cases, whether the code aligns with the existing architecture, and whether the model misunderstood the context. The generation action might take seconds, but the verification action may still involve reading, testing, modifying, and re-confirming. Therefore, when observing AI's impact on development efficiency, it is necessary to separate 'generation speed' from 'verification cost' and not equate code production speed with overall efficiency.
| Metric | Data | Year | Source |
|---|---|---|---|
| Use or plan to use AI | 84% | 2025 | Stack Overflow Developer Survey |
| Professional developers using AI daily | 51% | 2025 | Stack Overflow Developer Survey |
| Positive attitude towards AI | 60% | 2025 | Stack Overflow Developer Survey |
| Do not trust AI output accuracy | 46% | 2025 | Stack Overflow Developer Survey |
| Trust AI output accuracy | 33% | 2025 | Stack Overflow Developer Survey |
| 'Almost right' as major frustration | 66% | 2025 | Stack Overflow Developer Survey |
AI Can Improve Some Individual Work Metrics, but Local Productivity Gains Do Not Automatically Lead to Delivery System Improvements
Another set of data worth distinguishing comes from the Google Cloud 2024 DORA Report. The report shows that over 75% of respondents already rely on AI in their daily work, with use cases including code writing, code explanation, and information summarization. As AI use increases, some local work metrics show positive associations: for every 25% increase in AI adoption, documentation quality improves by approximately 7.5%, code quality by approximately 3.4%, and code review speed by approximately 3.1%.
However, the same study also observed that software delivery throughput decreased by approximately 1.5% and delivery stability by approximately 7.2%. This data does not necessarily mean AI itself causes deterioration in software delivery, but rather highlights that improvements at the individual task level and organizational software delivery performance operate at different levels and cannot be directly equated.
DORA's observations focus on the complete software delivery environment, which also includes testing capability, version control, team collaboration, small-batch delivery, platform engineering, and workflow stability. Even if AI helps developers produce code faster, if the rate at which new code enters the system exceeds the capacity of testing, review, or release processes, the organization's overall delivery performance may not improve correspondingly. The 2024 DORA data also shows that 39% of respondents have little or no trust in AI-generated code, further illustrating that enhanced generation capabilities do not automatically eliminate quality control needs.
The DORA 2025 public summary continues in this direction: AI acts more like an amplifier of an organization's existing capabilities. When foundational engineering capabilities are stable, AI can amplify efficiency gains; when processes, testing, or collaboration have issues, AI may also allow pre-existing problems to propagate faster into downstream stages. Therefore, studying developer AI usage requires observing both individual efficiency and organizational systems, rather than merely comparing the speed of 'human-written code' versus 'AI-written code.'
The Closer to High-Accountability Tasks, the More Developers Tend to Retain Human Decision-Making, but Verification Processes Remain Incompletely Standardized
Developers' acceptance of AI also varies with task accountability. According to the Stack Overflow 2025 Developer Survey, 76% of developers do not plan to use AI for deployment and monitoring, and 69% do not plan to use it for project planning. These tasks are typically directly related to system stability, resource allocation, incident response, and long-term technical decisions, so developers have higher requirements for accuracy, accountability, security, and privacy.
The human role is equally prominent in information confirmation. When developers do not trust AI answers, 75.3% state they still seek help from humans. This indicates that even though AI has become an entry point for technical information retrieval, human colleagues, expert judgment, and existing verification mechanisms still serve a confirmation role.
However, 'knowing verification is needed' does not mean the verification process is standardized. The Sonar Developer Survey of 1,149 technology and software engineering professionals globally who used AI in their work over the past year found that fewer than half of developers always verify AI-generated code before committing it to the codebase. In other words, there remains a gap between risk awareness and actual execution.
This further shifts the research question: the focus is no longer just on 'whether developers use AI,' but on who is responsible for confirming AI output before it enters the codebase, testing environment, and production, whether the organization mandates verification steps, and whether verification results are traceable. For high-accountability tasks, AI's role currently remains closer to auxiliary generation and information support rather than independently bearing decision-making responsibility.
| Scenario / Metric | Data | Observation Implication | Source |
|---|---|---|---|
| Do not plan to use AI for deployment and monitoring | 76% | Lower acceptance in high-accountability stages | Stack Overflow 2025 |
| Do not plan to use AI for project planning | 69% | System-level decisions still prefer human involvement | Stack Overflow 2025 |
| Seek human help when distrusting AI | 75.3% | Humans still play a key confirmation role | Stack Overflow 2025 |
| Always verify AI code before committing | Less than half | Gap between verification awareness and process execution | Sonar Developer Survey |
Trend Observations / Industry Implications
Looking at Stack Overflow, DORA, and Sonar data together, the change in developer use of generative AI from 2024–2025 is not a direct shift from 'not using' to 'complete dependence,' but a gradual entry into a hybrid phase of 'high-frequency usage + continuous verification.' Developer use of AI is shifting from 'whether to adopt' to 'how to verify': adoption rates continue to rise, but this has not simultaneously eliminated concerns about accuracy and accountability.
Early evaluations of AI coding tools focused more on whether models could generate code, answer technical questions, or reduce writing time. As tools increasingly integrate into real workflows, the evaluation dimensions are expanding. Whether sources are clear, whether answers can be traced back, whether code has been tested, whether version information is correct, and who is accountable for production outcomes – these factors are beginning to influence, alongside generation speed, whether AI output is truly usable.
DORA 2024 data further reminds enterprises that individual productivity and organizational delivery performance cannot be explained using the same metrics. Developers completing documentation, code explanation, or code review faster does not mean the entire software system will see proportional increases in throughput or stability. Organizations still need testing, small-batch delivery, code review, stable workflows, and clear accountability mechanisms to absorb the additional output generated by AI.
For technical information retrieval, this change is equally important. Developers need not just a system that can 'answer,' but also one that helps them determine what information the answer is based on, whether it applies to the current version and environment, and whether key conclusions can be independently verified. As AI becomes the entry point for more technical questions, verifiability is becoming part of information quality.
Research Implications
Based on this public data, when enterprises observe the use of generative AI, they should not merely count whether employees use AI, but also examine how AI output is verified after entering business processes. For content, data, and technical teams, three key aspects warrant focused examination: whether information sources are clear, whether key facts can be traced back, and whether there is human confirmation before AI output enters publishing, codebases, or decision-making processes.
This also has direct implications for GEO. For technical, highly specialized, and high-accountability information to consistently serve as a reliable source for AI systems, mere content coverage is insufficient. Clear sourcing, stable data definitions, verifiable facts, and well-structured information are more conducive for users to confirm the basis of the content after seeing an AI answer.
From the perspectives of AIPO's content planning & platform distribution and data analysis, enterprises can go beyond simply observing whether their brand is mentioned by AI. They can further analyze: what information AI answers cite, how different platforms interpret the same brand facts, whether key data points are consistent, and whether users can trace from the answer back to clear information sources. Such observations help place 'AI visibility' and 'information verifiability' within the same analytical framework.
Now that AI has become an entry point for search, research, and content access, businesses also need to understand whether their brand information appears across different AI platforms and what public information supports those answers. aipogeo can show you your real visibility across 6 major AI engines in just 1 minute.
Related Questions
Why do developers use AI heavily but still not fully trust AI-generated code?
High-frequency usage and direct trust are not the same thing. The Stack Overflow 2025 survey shows that while AI usage continues to rise, the proportion of developers distrusting accuracy remains higher than those expressing trust, indicating developers are more inclined to treat AI as an assistive tool rather than a verification-free source of results.
How many developers are already using AI in their daily work in 2025?
According to the Stack Overflow 2025 Developer Survey, 84% of respondents are currently using or planning to use AI tools, and 51% of professional developers report using AI tools daily.
What are common developer frustrations with AI coding tools?
The Stack Overflow 2025 survey shows that 66% of developers cited 'solutions were almost right, but not quite correct' as a major frustration, and 45.2% stated that debugging AI-generated code takes more time.
Can AI coding tools actually improve software development efficiency?
DORA 2024 observed an association between increased AI use and improvements in documentation quality, code quality, and code review speed, but the same research did not observe synchronized improvements in organizational-level software delivery metrics. Therefore, individual task efficiency and the complete delivery system need to be evaluated separately.
Why is the willingness to use AI noticeably lower for tasks like deployment and monitoring?
Deployment and monitoring directly impact system stability, security, and accountability. The Stack Overflow 2025 survey shows that 76% of developers do not plan to use AI for deployment and monitoring, reflecting a stronger need for human involvement in high-accountability tasks.
Why does AI-generated code still require human verification?
AI output may be syntactically valid, but can still have issues related to context, dependencies, versions, edge cases, or security. The Sonar survey also shows that fewer than half of respondent developers always verify AI-generated code before committing it, indicating that establishing a consistent verification process remains an important topic in current workflows.