Insurer Trade Groups Ask the NAIC to Take GLMs Out of Its AI Risk Evaluation Supplement

The NAIC’s Big Data and Artificial Intelligence (H) Working Group now links a 206-page file from the Exposure Drafts area of its page, with the text “Comments Received on version 5.0 exposure” and no posting date. The file name carries 10.02.2026. Its table of contents lists 26 numbered entries, three of which hold two files each. Several of the insurer trade associations’ letters make the same request, which is to take generalized linear models out of the AI Risk Evaluation Supplement.

The American Council of Life Insurers presents it as a repeat request: “ACLI therefore reiterates its request that traditional rule-based systems, deterministic algorithms, fixed mathematical formulas, and transparent predictive models such as GLMs and GAMs be excluded from the Supplement.” GAMs are generalized additive models. The Insured Retirement Institute draws the same line: “At minimum, we recommend that GLMs and GAMs be excluded from the scope of the Supplement.” The National Association of Mutual Insurance Companies “recommends that GLMs be explicitly removed from scope of the Supplement.”

The Committee of Annuity Insurers aims at the drafting. It “is strongly opposed to the new emphasis placed on including Generalized Linear Models” within the supplement’s scope, while granting a point to the other side: “While GLMs can present risks and impact consumer outcomes, this technology is tried, trusted and transparent.”

The American Property Casualty Insurance Association says it is “contemporaneously submitting separate comments from Pinnacle Actuarial Resources” on GLMs used for pricing. Its own letter says: “We believe further consideration should be given to whether conventional GLMs should be distinguished from AI systems for purposes of the Supplement.” The Pinnacle paper confines itself to pricing: “we believe GLMs developed specifically for the purpose of supporting rating factors should be excluded from the NAIC AI Risk Evaluation.” The Blue Cross Blue Shield Association’s GLM recommendation addresses a single table: “Remove the separate quantification of generalized linear models (GLMs) from Exhibit A.” Monitaur’s letter leans the opposite way: “We appreciate the inclusion of GLMs within the Supplement’s model taxonomy.” The working group’s July 22 panel on whether a GLM is AI, covered here in September, recorded no conclusion.

A second objection, from ACLI, IRI and NAMIC, concerns Exhibits C and D, two of the four exhibits in our guide to the tool, and each letter invokes the pilot. ACLI, saying “many pilot participants tested only Exhibit A” in that exercise, requests that the two “not be used broadly” until their burden, clarity and consistency have been tested. IRI’s letter says: “IRI recommends that Exhibits C and D be removed or deferred until further piloting can be conducted. These exhibits are duplicative in some cases and are not relevant to financial or solvency risk.” NAMIC’s letter says: “From a substantive perspective, NAMIC continues to suggest the removal of Exhibits C and D.” Separately, citing member feedback, it writes that “it does not appear that the majority of the piloting states utilized Exhibits C and D” and that it “does not believe that these Exhibits are ripe for adoption.”

Not every letter asks for less. Five NAIC consumer representatives, Lucy Culp, Adam Fox, Amy Killelea, Peter Kochenburger and Silvia Yee, write on Exhibit C: “We continue to be concerned that AI model risk criteria are set by the insurance company. We believe that this poses an inherent conflict of interest and that consumers would benefit from a uniform definition of high-risk models.” Their letter also says “There are reports that human review of model outcomes may be less than 10 seconds” and treats a review that brief as little assurance of meaningful human review. A one-page entry headed “From Wanchin Chou (CT DOI)” and listed in the contents as “CT Comments” proposes, among other points, an addition to Exhibit C, numbered 11a. It reads: “Where applicable, discuss whether the company evaluates input variables, derived variables, or model outputs for potential proxy relationships or effects that may contribute to Adverse Consumer Outcomes or unfair discrimination, including the methodology used and any resulting action taken.” The page’s text gives no date and ends without a sign-off.

AHIP’s letter is dated October 2, 2026. Written comments were due September 29. As we reported from the August 31 meeting packet, the working group’s timeline slide placed a 14-day exposure of version 6.0 in October, with version 7.0 to be considered at the Fall National Meeting. The group’s page lists a meeting for October 8 and gives its purpose as “continue public discussions on the AI Risk Evaluation Supplement.” The compiled file stops at the comments. It holds no working group response to any of them and no version 6.0 text.