What the web community is saying about AI
AI is everywhere right now, and it’s hard for web developers to separate useful signal from the noise. Rather than guess what developers need in this space, the web team spent the last few months talking directly to practitioners, including Google Developer Experts, about how they’re actually using AI today. The focus was on two areas: building user-facing features and boosting productivity in daily workflows.
The conversations were small, but the findings point to broader trends. The team used those insights, along with other developer landscape research, to shape a new AI collection for web.dev, containing overviews, codelabs, demos, and resources. What follows is a summary of what developers shared about their experiences with generative AI, the gaps they've discovered, and where they still need help.
How developers are using generative AI to work faster
Interest in generative AI is widespread, but usage today is highly practical. Developers told us they're leveraging code generation tools and LLM-based chat to speed up their work, with just about everyone either using these tools themselves or knowing someone who is.
What AI does well for coding tasks
According to feedback, tools like Gemini and Copilot deliver real value on specific, routine tasks: writing standard unit tests, basic autocomplete, and simpler functions that don't require wide project context. For complex, algorithmic work—or functions needing deep codebase awareness—the tools fall short.
More senior developers raised concerns about the long-term health of codebases. They flagged problems such as code duplication and maintenance burden, supported by data pointing to AI’s downward pressure on code quality. A common worry is that less experienced developers may not detect subtle errors, or may lack the judgment to validate what generative AI produces. Another gap: writing accessible components requires specific domain knowledge that current generation tools often get wrong.
LLMs as a learning tool
ChatGPT and Gemini are firmly in place as a learning aid. Developers ask LLMs to explain sorting algorithms, get up to speed in new languages, or fill in specific gaps in their knowledge. The appeal? The fast back-and-forth, and the lack of judgment—ask a question, and the model just delivers the answer.
The caveat applies here as well: for newer developers, there needs to be baseline competency to catch hallucinations and detect when a model generates incorrect responses.
Company policy lags behind developer behavior
Many of the people we spoke with say their companies haven’t set policies around generative AI for developer productivity. Most adoption is driven from the bottom up, by developers experimenting on their own. Where policies do exist, they largely discourage use. The leading reason is fear of leaking company IP to third parties.
"My company generally misunderstands what AI means, so they're not creating the right policies."
Attitudes do shift when companies engage directly with AI vendors to understand data usage and risks. Businesses are far more comfortable encouraging adoption when they secure enterprise accounts and partnerships that promise data protection.
Building user-facing AI features
On the product side, responses to "AI/ML" were heavily skewed toward generative AI. Developers genuinely want to use it to improve user experiences, but aren't clear on what those experiences look like—or what tools are dependable enough for production.
Where developers have shipped or are currently building features, the most common use case is generating answers to user questions, often via chatbots or one-off interfaces. For all of them, output quality is the number one concern. Accuracy is a persistent battle, particularly when the output is user-facing.
"Demos with AI are wild. Everytime I demo my project, the output is completely different."
Testing is a major pain point. Developers are writing test suites that try to validate outputs across prompts, but there are no established methods or tools. Manual review remains the default because responses are non-deterministic and our existing systems just aren't built to handle them. Cost calculation is also in the mix: running generative AI models isn't cheap, and teams are carefully weighing spend against user value.
Ready-made models, not bespoke training
Most teams we talked with default to standard models and APIs, a clear calculation around optimizing for time-to-market and limited engineering resources. There is interest and awareness in advanced techniques like Retrieval Augmented Generation (RAG) and fine-tuning, but the practitioners we spoke with prefer concrete tooling over developing expertise in ML.
"I want to stay in the web development space. I don't want to become an ML engineer."
Technologists are least interested in dropping down into ML engineering when it comes to shipping faster for their web apps. For the rare projects that require optimized models, developers would rather lean on other teams or stick with default products.
Privacy and security shape model choice
For developers building in data-sensitive verticals like healthcare, data privacy and security outweigh other concerns. Approaches involving on-device AI remain mostly unexplored, offering promise given strict requirements around controlling third-party data exposure. There’s broad recognition that cloud APIs increase external party visibility and that self-contained, on-device options might better meet demands.
AI content that speaks to web developers
Based on what the team heard, the near-term web.dev releases are designed to give developers what they asked for: high-level AI concepts, examples of productivity gains from generative AI, practice building with existing tools and APIs, and ways to collaborate effectively with ML engineers.
Many web developers want to keep their focus firmly on the web—on building with performant, safe tools designed to stay reliable. We encourage everyone who wants to explore deeper to directly build the tools and APIs developers need. The conversation doesn’t end here; web.dev AI content is an evolving area. Keep checking in with the AI collection as more resources arrive over the coming months.
If you have ideas or experiences to share on AI code output quality, privacy and security, cost, and which tools are worth a web developer's time, reach out for office hours.
