Stack Overflow Developer Survey 2025: AI Edition

The Adoption Paradox

Core finding: AI tool usage has reached 84%, but enthusiasm is waning—the honeymoon is over.

The 2025 Stack Overflow Developer Survey captures a pivotal moment in the AI-developer relationship. On the surface, adoption metrics look stellar: 84% of developers now use or plan to use AI tools, up from 76% in 2024. Daily usage has become the norm, with 51% of professional developers reaching for AI tools every single day.

But beneath these impressive adoption numbers lies a more complicated story. Developer sentiment toward AI has dropped to 60% favorable—down significantly from the 70%+ enthusiasm levels of 2023-2024. The initial excitement has given way to a more measured, often frustrated, relationship with these tools.

This paradox—rising usage paired with declining sentiment—suggests developers have moved past the "wow" phase and into the pragmatic reality of working with AI tools daily. They're using AI more because it's useful, not because they love it.

Trust in Crisis

Core finding: 46% of developers actively distrust AI accuracy—only 3% highly trust it.

Perhaps the most concerning finding is the trust deficit that has emerged. When asked about AI tool accuracy, the numbers paint a stark picture:

• 46% actively distrust AI outputs
• 33% trust AI outputs
• Only 3% report "highly trusting" AI-generated results

The skepticism runs deepest among those who know the most. Experienced developers are the most cautious, with just 2.6% highly trusting AI outputs and a full 20% reporting high distrust. This isn't technophobia—it's informed caution born from extensive exposure to AI's limitations.

This trust gap has real workflow implications. Developers report spending significant time verifying, testing, and correcting AI outputs—potentially eroding the productivity gains these tools promise.

The Almost Right Problem

Core finding: 66% of developers are frustrated by AI solutions that are "almost right, but not quite."

The survey identifies the specific pain points driving developer frustration, and one issue towers above the rest: 66% cite AI solutions that are "almost right, but not quite" as their primary complaint.

Top AI frustrations:

The "almost right" problem is particularly insidious. These near-misses often look correct at first glance, passing initial review only to cause subtle bugs that surface later. The time spent hunting down these issues can exceed the time saved by using AI in the first place.

Even more concerning: 20% report decreased confidence in their own problem-solving abilities. This suggests potential skill atrophy as developers increasingly defer to AI for solutions they might have reasoned through themselves.

AI Agents: Still Early Days

Core finding: Only 31% use AI agents, and 38% have no plans to adopt them.

While AI coding assistants have reached mainstream adoption, AI agents remain a frontier technology. The survey reveals that only 30.9% of developers currently use AI agents in any capacity:

• 14.1% use agents daily
• 9% use agents weekly
• 7.8% use agents monthly or less

More telling is the resistance: 37.9% have no plans to use AI agents at all. This isn't a "wait and see" posture—it's active rejection of the technology in its current form.

Among those who do use agents, the results are generally positive:

• 69% agree agents increased productivity
• 70% report reduced time on specific tasks
• 64% say agents helped automate repetitive work
• 62% credit agents with accelerating learning

However, only 17% report improved team collaboration—suggesting agents remain individual productivity tools rather than team-level force multipliers.

The concerns are substantial: 87% worry about accuracy and 81% about security/privacy when using AI agents.

The Tool Landscape

Core finding: ChatGPT dominates usage at 82%, but Claude is the most admired LLM.

The AI tool ecosystem has clear leaders, though market share and developer affinity don't perfectly align.

AI Assistant Usage:

LLM Usage (among professionals):

Despite ChatGPT's commanding usage lead, Claude Sonnet ranks as the most admired LLM. This gap between usage and admiration suggests developers often use ChatGPT out of habit or accessibility while viewing Claude as technically superior.

For AI agent orchestration, Ollama leads at 51%, followed by LangChain (33%) and LangGraph (16%)—indicating a preference for open-source, locally-run solutions.

Workflow Integration

Core finding: Developers resist AI for high-stakes tasks—76% won't use it for deployment.

The survey reveals clear boundaries developers have drawn around AI in their workflows. High-responsibility tasks remain firmly human-controlled:

Tasks where developers WON'T use AI:

• 76% — Deployment and monitoring
• 69% — Project planning
• 66% — Predictive analytics
• 65% — Security-critical operations

Where developers DO use AI:

• 54% — Searching for answers/documentation
• 36% — Generating synthetic test data
• 33% — Learning new technologies
• 31% — Documenting code
• 17% — Writing production code

The pattern is clear: developers embrace AI for learning, exploration, and auxiliary tasks while maintaining human control over anything that ships to production or affects system reliability.

Only 17% use AI for writing production code—a surprisingly low figure given the hype around AI coding assistants. The majority use these tools for understanding and supporting their work rather than generating the core deliverables.

Productivity Reality Check

Core finding: 52% report productivity gains, but 41% see minimal or no workflow change.

The productivity picture is mixed. While advocates tout transformational gains, the survey data tells a more nuanced story:

Workflow impact from AI tools:

• 16% — Changed workflow "to a great extent"
• 35% — Changed workflow "somewhat"
• 41% — Minimal or no change
• 8% — Negative impact on workflow

So while 52% report positive productivity effects, a substantial 41% see minimal or no change to their workflows despite adopting AI tools. This suggests that for many developers, AI tools are nice-to-have additions rather than fundamental workflow transformations.

The "great extent" category at only 16% indicates that truly transformational AI-assisted development remains the exception rather than the rule—even among those actively using these tools.

The Human Element

Core finding: Even if AI handles most coding, developers still want humans for trust, ethics, and learning.

The survey asked developers when they'd still want human help even in a hypothetical future where AI handles most coding tasks. The responses reveal what developers truly value:

When developers want human help:

• 75% — When they don't trust the AI's answer
• 62% — For ethical or security concerns
• 61% — To fully understand something
• 58% — To learn best practices
• 52% — For complex architectural decisions

These responses illuminate the limits of AI as developers perceive them. Trust verification, ethical judgment, deep understanding, and learning—these remain fundamentally human needs that AI cannot satisfy.

The strong preference for human guidance on ethics (62%) and security (62%) aligns with the workflow data showing developers keep AI away from high-stakes decisions. It's not just about capability—it's about accountability and judgment.

Job Security Concerns

Core finding: 64% believe AI poses no job threat—but anxiety is growing.

Developer sentiment on job security has shifted subtly but meaningfully:

• 64% believe AI poses no threat to their job
• 36% express some level of concern

While 64% still feel secure, this is down from 68% in 2024. The 4-point shift suggests growing uncertainty as AI capabilities expand and economic pressures mount in the tech industry.

The concern isn't uniform across experience levels. Early-career developers—who show the highest AI adoption rates—also report higher anxiety about AI's impact on job prospects. Senior developers, despite being more skeptical of AI capabilities, feel more secure in their positions.

This creates an interesting dynamic: those most enthusiastic about AI tools are also most worried about what that enthusiasm might mean for their careers.

Generational Divide

Core finding: Early-career developers lead adoption at 55% daily usage; seniors are most skeptical.

The survey reveals a clear generational divide in AI attitudes and usage:

Daily AI tool usage by experience:

• Early-career developers — 55.5% daily usage
• Mid-career developers — ~50% daily usage
• Experienced developers — 47.3% daily usage

Trust levels by experience:

• Early-career — Higher trust, lower skepticism
• Experienced — Only 2.6% highly trust; 20% highly distrust

This pattern makes sense: early-career developers have less historical context for how software development "should" work and may be more willing to accept AI as a natural part of the process. Experienced developers have seen tools come and go, understand the complexity AI often glosses over, and have developed instincts that AI sometimes contradicts.

The favorability gap is notable too: professional developers show 61% favorable sentiment versus 53% among those learning to code. Counterintuitively, those with more experience are slightly more positive—perhaps because they're better at using AI effectively while understanding its limits.

Conclusion

The 2025 Stack Overflow Developer Survey paints a portrait of an industry in transition. AI tools have become ubiquitous—84% adoption is remarkable by any standard—but the relationship has matured from infatuation to something more complicated.

The trust crisis (46% distrust), the "almost right" problem (66% frustrated), and the resistance to AI agents (38% no plans to adopt) all point to a developer community that has learned through experience where AI helps and where it falls short.

The clearest signal may be in how developers deploy these tools: eagerly for learning and exploration, cautiously for code generation, and not at all for deployment and production decisions. They've drawn boundaries that make sense given AI's current capabilities—and those boundaries are worth noting for anyone building or implementing AI developer tools.

Perhaps most telling: even imagining a future where AI handles most coding, developers still overwhelmingly want human help for trust verification, ethics, and genuine understanding. The human element isn't going away—it's just being redefined.

This analysis is based on Stack Overflow's 2025 Developer Survey, with over 65,000 developer respondents from the global developer community. The full survey results are available at the link above.