Code Got Cheap. Quality Didn't: Why "AI Makes Software Worthless" Gets the Cost Structure Wrong

The rapid advancement of artificial intelligence has made generating code remarkably cheap, but it has not made defining quality, proving conformance, or building trust in software cheap. The cost structure of software is changing: while implementation is now abundant, explicit definitions, validation, traceability, and operational evidence remain scarce.

AI can produce candidate implementations swiftly, but a product's value does not disappear simply because its code can be easily reproduced. The misconception that "software value goes to zero in the AI era" arises from conflating the cost of producing an artifact with the value of that artifact.

Software development cost consists of two main components: defining quality and producing conformance. Quality is defined as conforming to requirements, meaning that writing requirements is part of defining quality. Development work either involves defining quality or producing something that meets that definition. This perspective differs from the traditional Cost of Quality model, where quality is considered a slice of the total cost.

Defining quality can be a complex process, particularly when requirements are uncertain. Processes like prototypes, beta releases, and staged rollouts help discover requirements. These steps are crucial in managing risk and ensuring that the quality definition is appropriate. Verification, which checks if the product matches the definition, and validation, which assesses whether the definition itself is correct, are essential. Both activities are necessary for ensuring a product's correctness.

The relationship between definition and product is bidirectional: the product's correctness depends on the definition, and the definition's correctness depends on the product. Value emerges from this interplay, not from either side alone. This relationship explains why prototyping remains important: it allows for testing and validating the definition without committing fully to the product.

In the AI era, a significant shift is occurring: the abundance of generated code makes it harder to determine which code is correct. The real value now lies in proof of correctness, which is more challenging to establish when vast amounts of code are produced. The underlying assumption that building a successor to a product is cheaply possible is flawed. It overlooks the fact that building the successor requires both the definition and the product, neither of which can be assumed to be cheaply reproducible.