The intersection of artificial intelligence and software development has become one of the most dynamic and debated frontiers in technology. From the promise of “vibe coding” to the sobering realities of implementation, AI tools are reshaping how programmers create, debug, and think about code. This transformation is particularly visible within game development, a field known for both its complex technical challenges and rapid innovation cycles. To understand the true impact, beyond the marketing hype, we must turn to the developers themselves—those who wield these tools daily to build real-world projects. Their experiences reveal a nuanced landscape where AI offers remarkable benefits in specific contexts while presenting significant limitations in others, fundamentally altering workflow priorities rather than simply replacing human ingenuity.
AI as a Catalyst for Speed and Understanding
For many studios, AI has become an integral part of the development toolkit, primarily by accelerating non-critical and exploratory work. Kristinn Þór Sigurbergsson, director of gameplay engineering at CCP, the studio behind Eve Online, reports that his team uses AI tools extensively. One of the most significant advantages is in navigating and understanding large, mature codebases. Tools like Cursor or Claude Code allow developers to get up to speed remarkably quickly through enhanced codebase navigation, summarisation, and logic tracing across files. This capability reduces the initial learning curve for complex projects.
Perhaps the most transformative application, according to Sigurbergsson, is in the realm of non-production code. Writing small scripts for data generation, issue investigation, or automating one-off tasks has seen a dramatic shift in efficiency. “Something that might have taken half a day can now be done in minutes,” he notes. This changes the economic proposition of such tasks, freeing developer time for more critical thinking. Furthermore, AI aids in prototyping features or behaviours, serving as an effective communication and exploration tool even if the output is rarely production-ready.
This speed boost alters the development cycle itself. Developers now spend more time in planning and review stages, with less time dedicated to manual implementation. The lowered “cost of being wrong” encourages greater bravery in the design phase, as iterations can be executed faster. Indie veteran Cliff Harris of Positech describes using Anthropic’s Claude as “life-changing,” emphasizing how it serves as a partner for learning obscure algorithms and optimisations, accelerating his decades-long experience.
The Supporting Role: Debugging and Refactoring
AI also plays a substantial supporting role in more routine coding tasks. Garry Newman of Facepunch Studios highlights its utility in explaining concepts and refactoring code. Instead of manually copying and pasting across numerous files, he can leverage AI to perform these tedious operations, viewing it as an evolution of traditional coding practices. Many experienced coders use AI for quick references, debugging suggestions, and as a sounding board, especially as reasoning models with low hallucination rates improve.
The Limits of AI: Hallucination and Loss of Understanding
Despite these advantages, developers report significant and often frustrating limitations. The most cited problem is the phenomenon of “hallucination”—where AI models confidently generate incorrect or nonsensical code. Adam Grimley, senior programmer at Huey Games, uses AI primarily for brainstorming outside his knowledge base but always verifies its suggestions against human-written sources, a process he describes as slow. Alex Darby, a veteran technical director, found AI tools immensely unhelpful due to this unreliability; the time spent verifying and correcting nonsensical outputs often exceeded the time it would have taken to write the code manually.
Beyond accuracy, there is a deeper concern about the erosion of understanding. Chet Faliszek of Stray Bombay emphasizes that learning a language or system involves more than getting a functional end output; it’s about grasping the underlying logic to innovate and improve. Bram Ridder of Kythera AI similarly cautions that over-reliance on AI for code generation can deprive developers of the crucial learning process, leading to a fragile knowledge foundation.
Architectural Weakness and Workflow Disruption
The nature of AI-generated code itself is another point of contention. Jem Frisby, a backend web developer at Failbetter Games, criticises its output as often having “poor architecture,” being brittle, and utterly disinterested in performance. Worse, she notes that it forces the developer to adapt to its style, disrupting collaborative software practices. John Ogden, CTO at Huey Games, agrees that AI fails at the architectural level, noting it cannot replace the holistic, general intelligence of a human programmer who builds and maintains a mental model of a system. The “worst case,” he suggests, is a large blob of AI code that requires human debugging, stripping away the intuitive problem-solving that comes from deep system ownership.
The Context Window Problem and Future Frameworks
A fundamental technical barrier to wider adoption in complex projects like games is the context window limitation. Paul Kilduff-Taylor of Mode 7 Games observes that current commercial AI systems lack context windows large enough to encompass an entire Unity project. This explains why many “AI-made game” demonstrations are simplistic prototypes built on lightweight frameworks. The inability to understand a full project’s architecture scales poorly, making AI unreliable for larger, integrated systems.
The future of AI in coding, therefore, may depend on the development of appropriate frameworks and workflows. Darby suggests that only companies with massive suites of automated tests and continuous delivery processes—common in tech but less so in games—could effectively leverage AI at a company-wide level, as their code is more modular and self-contained. Kilduff-Taylor concludes that context, scaffolding, and secondary systems are key to unlocking AI’s potential, and the right framework for games does not yet exist.
Ultimately, the real-world use of AI in coding today paints a picture of a powerful but flawed assistant. It excels as a accelerator for exploration, prototyping, and automating trivial tasks, shifting developer focus from typing to thinking. However, it struggles as a creator of reliable, architecturally sound production code and risks undermining deep technical understanding if used uncritically. The technology is not a replacement for programmers but a tool that, when applied judiciously, can make them faster and more daring, while demanding heightened scrutiny and review. The journey from hype to mature integration hinges on overcoming current limitations in reliability, context, and workflow adaptation, a challenge the industry continues to navigate.
