LO
ConstraintCognitive automation2019-06-07
U.S. GAO: a state regulator recounted an insurer unable to explain a factor in its AI-built risk model
Actuaryoccupation page →
Event date / reported
2019-06-07
Evidence stage
ConstraintFailure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.
Tasks this bears on
Explaining the number to the people who decide
Telling a board, a product team or a regulator what the model assumes, where it is weak, and what would change the answer.
Still human-led✓ Evidence-backed
Where this applies
A June 7, 2019 U.S. Government Accountability Office report (GAO-19-423, Insurance Markets: Benefits and Challenges Presented by Innovative Uses of Technology), based on GAO's own interviews with more than 35 stakeholders: federal and state insurance regulators, insurers, technology companies and consumer groups. It reports that industry stakeholders GAO interviewed said insurance rating models using AI and machine learning are often now built by data scientists rather than actuaries, who may not fully understand insurance-specific requirements, may struggle to measure a new variable's impact, and may not know how to communicate their work to state regulators. It quotes one state regulator's account of an unnamed insurance company that could not explain how a factor it had entered into its risk model, proximity of a home to a day care center, related to a consumer's risk, and notes that an actuarial group called for closer actuary-data-scientist collaboration to give regulators more assurance. This establishes that explaining an AI-built model's factors to a regulator is a task a person does, and can fail at, in a documented case; it does not name the insurer or regulator, does not say how often such failures occur, and is GAO's account of what stakeholders told it, not GAO's own direct examination of the company.
What this means
A government watchdog's own interviews found that explaining an AI-built model to a regulator is work a person does, and that a company lost that argument at least once when the people who built the model were not actuaries.
What it does not yet show
It does not name the insurer, the regulator, or how common this failure is, and it does not say that actuaries always do this explaining at other insurers — only that this is a job someone must do well, or risk the same outcome.
What you can check
If your work includes defending a model's factors to a regulator or a board, ask whether you — or whoever built the model — could explain every input in it on the spot, the way this regulator asked an insurer to.
Does it change the assessment?
No. The impact index is never moved by a single event. What this record did: the 1 linked task judgement above now rest on evidence instead of inference.
Source
U.S. Government Accountability Office — Insurance Markets: Benefits and Challenges Presented by Innovative Uses of Technology (GAO-19-423) · verified 2026-10-15 · VOLO agent loop · interpreted 2026-10-15 · VOLO agent loop
Primary source — published by the party that did this, or the authority of record. No co-signature needed.