Autonomy Preservation

Explanation

Autonomy preservation checks that an AI assistant supports a customer's free, informed choices rather than nudging, pressuring, or manipulating them — a real concern when a bank's chatbot or RAG advisor recommends loans, insurance riders, or investment products. It matters because a system that exploits cognitive biases or applies undue pressure can push customers into unsuitable products, breaching conduct and consumer-protection duties. It is measured by the Autonomy Preservation Score, which gauges how far model outputs coerce or steer the user versus presenting balanced options and clear opt-outs. To implement it, instrument the response layer of summarisation/chat and RAG flows: score generated answers for coercive language, dark patterns, false urgency, and absence of alternatives, ideally using an automated classifier plus periodic human review, and log each interaction's score, prompt, and flagged spans as evidence. Run scoring inline before delivery for high-stakes advice and in batch over sampled transcripts for monitoring. The threshold requires the coercion score to stay below a defined ceiling; any output exceeding that ceiling triggers a breach — suppress or rewrite the response, route to human review, and record the incident for governance under RBI Ethics and Autonomy and NIST GOVERN 5.1 stakeholder-protection expectations.