Three Tests for Native AI: Removal, Prior-Impossibility, and Explanation
Updated
Knowledge on this page was mainly distilled from You're Building a Stone Cathedral Out of Concrete.
Most AI products today are incremental improvements on existing software rather than genuinely new forms. Three tests help distinguish a native AI product from a horseless carriage.
The Three Tests
- The removal test. If you removed the AI from the product, would it still basically work? If yes, the AI is a feature, not a new form.
- The prior-impossibility test. Could the product have existed five years ago with enough engineering effort? If yes, AI is making it cheaper, not making it new.
- The explanation test. When you describe the product, do you need an analogy to something that already exists? "It's like X but with AI" signals the carriage phase. Products that matter will need their own vocabulary.
Failing all three tests is not a condemnation. It is a diagnosis of where a product sits on the adoption curve. The concrete is still wet, and most teams are still learning what the material can do.
Q&A
What is the removal test for AI products?
If you strip out the AI layer and the product still fundamentally works, the AI is a feature bolted onto an existing form rather than the structural basis of a new one. A document editor with AI autocomplete passes this test negatively because the editor still functions without it. A system where the interface is a conversation and the output reshapes itself per user would collapse without AI.
What is the prior-impossibility test?
Ask whether the product could have been built five years ago given unlimited engineering time. If the answer is yes, AI is reducing cost or development time but not enabling a genuinely new capability. The test filters for products that exploit properties unique to AI, such as real-time adaptation, natural-language interfaces, or autonomous goal pursuit.
What is the explanation test?
If you describe your product as "like X but with AI," you are still anchored to the old form. Products that represent a truly new architecture tend to need their own vocabulary because no existing category captures what they do. Early automobiles needed the word "car" because "horseless carriage" stopped making sense once the form diverged from carriage design.
Do most AI products today pass these three tests?
No. Most current AI products fail all three tests, which reflects where we are in the adoption cycle rather than a flaw in any particular team. Historically, every new material goes through a phase where it replicates the forms of the material it replaced before native forms emerge. AI is in this transitional period.
How do these tests relate to the horseless carriage pattern?
The horseless carriage pattern describes how new technologies first replicate old forms before native forms emerge. The three tests are a practical diagnostic for identifying which phase a specific product occupies. They translate the historical observation into a decision-making tool for builders evaluating their own work.