An announced shooter is currently being discussed on the gaming platform Steam, which its development team says was largely created with artificial intelligence. The reactions from players are strong. The case is interesting beyond the gaming industry because it touches on a question that now affects every software project: how much machine-generated code can a product withstand without quality and credibility suffering?
What lies behind the term vibe coding
Vibe coding is the term used for a way of working in which developers describe an outcome in natural language and have the code largely generated by an AI model, without tracing every line in detail. The appeal is obvious: prototypes are created in hours instead of weeks, and even small teams can cover functional scopes that previously required several specialists.
This works surprisingly well as long as it is about proof of feasibility, internal tools or clearly defined sub-tasks. It becomes problematic where the prototype becomes a product that is to be paid for, operated and maintained over years. It is precisely at this threshold that the criticism in the current case ignites.
Why the audience reacts
What is remarkable is less the technology than the reaction. A considerable part of the criticism is not directed against AI as such, but against the expectation of paying for a result whose creation is perceived as largely automated. In such situations, users assess not only the product but also the effort and care they assume lies behind it.
This perception cannot be argued away, but it can be influenced. Anyone who discloses where AI was used and where humans checked, corrected and took responsibility removes the basis for a large part of the criticism. Anyone who conceals the share and is only confronted with it afterwards loses twice: in product quality and in trust.
The typical weak points of generated code
From project practice, some recurring patterns can be named that apply regardless of the specific case:
- Superficial correctness: Generated code often runs straight away, but covers special cases, error handling and unusual inputs only incompletely.
- Missing architecture: Individually sensible building blocks do not add up to a viable system. Responsibilities overlap, logic is implemented multiple times.
- Security gaps: Input validation, permission checks and the handling of credentials are among the areas in which models reproduce known error patterns.
- Maintenance debt: Code that nobody in the team really understands is difficult to extend later. Every change becomes a risk.
- Legal ambiguities: The origin and licence status of generated components are not always clear – an issue that comes up at the latest in tenders and audits.
What helps in practice
The way out is not to banish AI tools from the development process. Used productively, they demonstrably save time. What matters is that they are embedded in a process that catches errors before users notice them.
- Mandatory review without exception: Every generated line goes through the same review as hand-written code. Whoever approves code takes responsibility for it – regardless of who or what wrote it.
- Automated review layers: Static analysis, dependency scans and robust test coverage catch some of the typical weaknesses before a human has to look.
- Architecture remains a human matter: Data model, interfaces and module boundaries are defined in advance. AI fills in structures, it does not invent them.
- Defined areas of use: For tests, migration scripts or boilerplate, the benefit is high and the risk low. Stricter rules apply to authentication, payment processing or data protection logic.
- Traceability: The team should document which components were created with AI support. This makes later troubleshooting easier and answers clients' questions before they are asked.
Transparency is a product feature
The case on Steam shows that communication about the creation process is now part of the product. Target audiences develop a sense of when a result was cobbled together, and they say so publicly. In review portals, app stores or forums, a narrative emerges within days that a team finds difficult to recapture.
For companies, this means: the question of whether AI was used is less delicate than the question of whether someone reviewed the result.
A clear statement on this – internally to management, externally to customers – is more robust than silence.
Context for web and software projects
For website and application projects, the focus of the work shifts: writing code is not the bottleneck, but rather its review, structuring and long-term maintenance. Anyone introducing AI support should therefore invest in reviews, tests and architecture decisions in the same step, otherwise the effort saved is merely shifted to the maintenance phase. And anyone who openly states the AI share in a product prevents this from later becoming a question of trust.