There is a crisis of legibility in AI-powered products.
The crisis is not about capability. The models are capable. The outputs are often remarkable. The crisis is about trust — specifically, about the gap between what a system can do and what a person can verify.
The verification problem
When a person makes a decision, you can ask them how they made it. They will give you a reason. The reason may be incomplete, or partially wrong, but it is a starting point. You can probe it. You can disagree with it. You can learn from it.
When a system makes a decision, you get an output. The reasoning is not legible in the same way. You cannot probe it with the same tools. You can accept it or reject it, but you cannot really engage with it.
This is why trust is so hard to build in AI-powered products. Trust requires verifiability. Verifiability requires legibility.
What legibility looks like in practice
Legibility is not explainability in the academic sense. It is not a technical requirement for model interpretability. It is a design requirement.
A legible AI product is one where the person using it can form a working model of what the system is doing and why. They do not need to understand the mechanism — they need to understand the intention.
Vessel does this by showing its work. When it surfaces a document as relevant, it shows the passage that triggered the relevance. When it synthesises from multiple sources, it lists them. The person can follow the thread.
They cannot verify the model's internals. But they can verify the connection between the input and the output. That is enough to build a working theory. A working theory is enough to build trust.
The cost of opacity
Opaque AI products are fast to build and slow to trust. The first demo is impressive. The adoption curve is flat.
Legible AI products take longer to design. The question of what to show — and how to show it — is genuinely hard. But the adoption curve is different. People use them because they understand them enough to rely on them.
We believe that AI that earns trust will outlast AI that demands it.