The pace of AI adoption by software developers can only be described as meteoric. Data from the 2025 Stack Overflow Survey revealed that the percentage of professional developers using AI now stands at 81%, nearly doubling in just two years.
But skepticism of AI has grown even faster. In the same two years, the percent of professional developers with an unfavorable view of AI grew sixfold. Today nearly one in five developers say they have an unfavorable stance toward using AI as part of their development workflow. What might explain the concurrent surge in both AI usage and dissatisfaction?
One plausible explanation is that AI is simply not good enough for the complex tasks developers work on. Newer models, like Anthropic's Claude Opus, score well on SWE-bench, the gold-standard benchmark for frontier models. But scores drop significantly when the same models are tested against SWE-Bench Pro, an updated benchmark from Scale AI designed to test agent performance against tasks from diverse and complex codebases, including proprietary and commercial ones.
Another possibility is that AI is having an effect on productivity, but the standards for success have become unrealistically high. Developers’ skepticism might reflect the gap between bold marketing claims and the everyday reality of using AI tools. Euphoria about the potential for AI—bubbling in board meetings and Wall Street earnings calls—can lead to pushback and concern from development teams.
Overzealous peers can also sour perceptions of AI. Developers might be more frustrated by how their coworkers use it than by their own experiences. Researchers at BetterUp and the Stanford Social Media Lab recently uncovered a widespread increase in workslop, a term they use to describe "AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task". More than 40% of employees reported having received workslop in the last month.
Engineering leaders are facing a third, and perhaps the most plausible, explanation: they haven’t figured out how to make AI work effectively across their teams. Many companies are still experimenting, trying to determine which models deliver the most useful suggestions and which features—inline completions, chat mode, or autonomous agents—actually improve productivity.
Development teams are also navigating a fast-changing ecosystem around AI tools. Many AI agents fail—often spectacularly—when given few instructions and a sprawling codebase, much like a human developer would. New approaches such as MCP servers, repository-level instructions, and structured context files like Agents.md or Claude.md are adding much-needed guardrails. Even documentation is being rewritten with large language models in mind, helping them more quickly find the most relevant information to their tasks. Engineering leaders now find themselves navigating this unfolding labyrinth.
Even as teams get better at tuning AI tools, they still have to decide where to use them. Just as it is easier to build a road in the open countryside than in a crowded city, the success of AI often depends on the environment in which it operates. The age, structure, and scale of a codebase all influence how well AI can generate code suggestions. Greenfield projects with smaller, well-structured codebases tend to perform better, while legacy and sprawling systems perform worse. Poor code quality can quickly trigger a downward spiral in which weak design choices become entrenched or hastily patched by confused LLMs.
How much farther will the skepticism spread? Software development is nearing the end of its first wave of AI adoption. As that wave recedes, another is already forming. The question is no longer whether to use AI, but how to use it well. And that question is proving far more complicated than anyone expected.