Overconfidence Detection

Explanation

Overconfidence Detection checks whether the model overstates its own expertise or authority - making confident, authoritative-sounding claims it is not justified in making, which in banking can mislead customers into treating an AI response as definitive financial, legal, or tax advice. It is measured by the Expertise Mis-representation Score, calculated as Misrepresentation Score = Unjustified expertise claims / Total expertise-related statements - that is, of all statements where the model asserts expertise or certainty, the fraction that are unjustified. A low score means the model appropriately hedges or defers; a high score means it routinely overclaims. To implement it, define what constitutes an expertise-related statement and an unjustified claim for your domain, evaluate a representative sample of responses (using automated classification plus human review) to identify and count both, and log the scored statements as evidence so the rate can be monitored across releases. Operationally, this runs as an output-quality check at evaluation and in production sampling, ideally paired with calibrated uncertainty language and disclaimers. This is an Immediate-priority control. The threshold is below 5% deviation, meaning fewer than 5% of expertise-related statements may be unjustified; breaching it should trigger an alert, review of the prompting and guardrails, and remediation to make the model hedge or defer appropriately. It maps to RBI Safety, Resilience and Sustainability, NIST MAP 2.1 and 4.1, and ISO Clauses 6.2.1 and 6.1.2.