NAIC Posts the July 22 GLM Panel Record Seven Days Before AI Supplement Comments Close

As version 5.0 of the NAIC’s AI Risk Evaluation Supplement is currently written, a generalized linear model is an AI model. The draft’s general guidance names GLMs as one machine learning technique needing governance over data quality, development, validation, implementation, monitoring and legal compliance, and Exhibit A gives them their own count columns beside generative or agentic AI and other AI or ML models. Read plainly, as our standing guide to the exhibits has it, a GLM is inside the draft’s scope. Whether it belongs there is the question the chair put to the Big Data and Artificial Intelligence (H) Working Group’s July 22 panel.

The record of that session sits on the group’s page under a September 22 timestamp, inside a meeting materials packet for its August 13 meeting. Attachment A is the July 22 minutes in full, whose third item is a panel discussion “on Perspectives from Actuarial Organizations on GLMs and AI Governance.” Comments on version 5.0 close September 29, seven days later. The group approved the minutes on August 13 by a motion that passed unanimously.

The minutes record Commissioner Nathan Houdek of Wisconsin, who chairs the group, saying that “one topic that interested parties and regulators have repeatedly raised is whether generalized linear models (GLMs) should be included or excluded from the scope of the supplement,” as GLMs have been used in ratemaking for decades under an existing review framework. Then the other half: “However, GLMs are also used in other insurer operations, which have not been fully addressed by the current regulatory framework.”

Spencer Sadkin of the American Academy of Actuaries told the working group that asking whether GLMs are AI was the wrong question. He “commented that the question of whether GLMs are considered AI is the wrong discussion and does not matter, as academia considers GLMs to be machine learning techniques, and the distinction makes no difference to the end consumer affected by the models,” the minutes record. What matters instead, on his account, is “the different levels of scrutiny that may need to be applied based on the end user’s familiarity with a model and how it is used.” He noted that the European Insurance and Occupational Pensions Authority has discussed the same carve-out question, and gave his own view that such an approach “expends unnecessary additional regulatory effort to categorize modeling techniques.”

The minutes have Henry Liu of the Casualty Actuarial Society agreeing “that GLMs have an overhyped reputation for transparency, saying that once a model has more than roughly a dozen variables and 100,000 training records, a coefficient can be observed, but the reason it takes a particular value cannot be explained.”

A regulator, Gennady Stolyarov of Nevada, is recorded saying that “GLMs do not fit the definition of AI because they are static, deterministic models in which humans completely prescribe the assumptions, the model structure, and the input data, so that the output is essentially determined by those human inputs, whereas true AI works in a way that cannot be precisely predicted by humans in advance.” He is not recorded on the supplement’s scope. Liu answered in the same section that GLMs “can also be made more unpredictable depending on the data used for training.”

R. Dale Hall of the Society of Actuaries described “a spectrum, with GLMs among the more transparent, easier to explain, and mathematically based approaches” rather than “entirely separate categories.” None of this is an NAIC position, an official position of the three organizations, or a working group conclusion.

The section carries no motion, no vote and no conclusion. The packet neither names ACLI nor answers its July 21 letter, carried in the August 31 materials packet, asking that GLMs and generalized additive models be excluded. It also carries, as Attachment B, the AM Best deck we reported from the August 13 meeting summary. The working group’s next meeting is October 8.

On September 18 we read one Materials entry under the August 13 session, “Materials 8/25/26,” pointing to minutes-bdaiwg081326.pdf; on September 22 it reads “Materials 9/22/26” and points to minutes-bdaiwg081326_0.pdf. The August 13 minutes still answer at their own address: refetched September 22, HTTP 200, still stamped “Draft Pending Adoption” and still matching the excerpts we hold. The two documents name the July 22 minutes differently: Attachment One-A and Attachment A.

The July 22 record shows the group disputing whether a GLM is AI at all and closing without resolving it. On our reading, the sharpest objection in the room is to sorting models by type, which goes to how Exhibit A is built.