In credit assessment, a generative answer does not merely summarize information; it silently constructs a decision surface.
What the phenomenon looks like
An answer can infer risk, solvency, or desirability without ever stating that a score-like judgment has been made. The user sees guidance or classification, while the system has already hardened access through implicit ranking.
Why it happens
The model fills gaps by borrowing the nearest stable pattern from public discourse, documentation, and training priors. The result is often coherent, but coherence here comes from inference, not from authorized interpretation.
Why it matters
The problem is not only opacity. Once scoring logic is hidden inside ordinary language, it becomes difficult to contest, audit, or even detect as a form of decision support affecting access.
What must be governed
- Treat ranking language as potentially decision-bearing whenever access or eligibility is affected.
- Expose factors, exclusions, and limits instead of letting the system imply a score through style.
- Separate descriptive financial information from any inference that changes practical access.
