The trade group asked for GLMs to come out. Version 5.0 gave the GLM a definition.

The packet for the working group’s August 31 meeting carries both a July 21 letter from ACLI asking that predictive models with transparent structures, among them generalized linear and generalized additive models, be excluded from the scope of the exhibits, and a draft of version 5.0 that files the GLM under new machine learning guidance. Staff’s summary of changes sits between those two documents, and its machine learning entry is about when a regulator may stop short.

This week’s story: the letter is about version 4.0, and version 5.0 defines the GLM

The letter’s heading names version 4.0 of the AI Systems Evaluation Tool rather than the draft posted for comment after that meeting. The ask itself is one sentence: “At minimum, ACLI recommends excluding predictive models with transparent structures, such as Generalized Linear Models (GLMs) and Generalized Additive Models (GAMs), from the scope of the exhibits.”

New General Guidance on machine learning models names GLMs and puts risk and governance in one sentence: “As with other machine learning techniques, GLMs are not without risk of causing unfair discrimination or other adverse consumer outcomes and require effective governance over data quality, model development, validation, implementation, ongoing performance monitoring, and compliance with applicable laws and regulations.” The definitions section gives the GLM its own entry.

The summary of changes records that addition as offering “regulators the opportunity to limit inquiry after Exhibit A responses are received if the inquiring regulator deems it appropriate.” The model inventory Exhibit A names in this version sits on the AI Systems Request List in Part Two, after the counting exercise in Part One. Our standing guide to what each exhibit asks covers the version this draft revises. Our reading is that the new discretion changes nothing about which models Exhibit A counts: a rating GLM in production is either on that inventory or off it, and that discretion arrives after the answer has already been written down. The exclusion the letter asks for would have changed the count itself.

Two of the letter’s four asks line up with that itemized summary: the removal of “Degree of Potential Harm to Consumers” and a definition of AI Model “leveraging the NIST/WH Executive Order language.” The other two have no counterpart there, a definition of “Insurance Regulatory Standards” and a clarification that traditional rule-based models and fixed mathematical formulas fall outside the AI System definition, though an earlier staff slide lists AI System under “Clarification of Definitions.” Which of the four the working group weighed, the record does not say: the draft attributes nothing to any commenter, and the forty pages hold no answer to the letter.

What to do this week

• Ask the people who build your rating models whether a generalized linear model now in production would show up on an Exhibit A response, and keep their reply in a form you can send onward. The scope question in this draft is that question, and the reply sets how long your Exhibit A list runs.
• Write the comment as a one-model letter this week and send it to ssobel@naic.org or maromero@naic.org, the only route the posted notice names. The notice sets the close of business on September 29 as the cutoff, and describing one model and what answering for it would take is shorter work than a position paper.
• Set the states you write in against the twelve pilot states we listed, and against our record of which jurisdictions have adopted the model bulletin and which built their own framework. The supplement is one instrument, and the department that decides to send it to you is a separate fact about your book.

On the Docket

• September 29, 2026: The 30-day comment period on version 5.0 ends. The working group’s posting sets the close of business that day as the cutoff and names two NAIC staff to receive comments, without saying when the 30 days began. Anyone whose model inventory would change if GLMs stay inside the exhibits. NAIC working group page

• October 2026: The packet’s timeline slide gives the month a working group meeting to take comment on version 5.0, and version 6.0 exposed for 14 days. No date for that meeting appeared on the working group’s page or the NAIC’s AI page as of our check on September 1, so the month is all the calendar supports. Anyone who means to comment twice, and anyone tracking the two rounds we reported on August 27. Packet timeline slide

• November 2026: The Fall National Meeting, which the NAIC’s events page puts on November 14 to 17 in Dallas at the Gaylord Texan. The timeline slide files version 7.0 there for adoption consideration, two versions past the one now open for comment. Whoever maintains an examination file written against the wording now in force. Packet timeline slide

This week in brief

Governance

The American Medical Association’s deck reached the task force’s prior authorization page later on August 31, after the morning count in our report on that session, and the page listed five documents rather than four when we checked it on September 1. Its twelve pages carry no instance of AI, artificial intelligence, machine learning, or algorithm on our own read of the text, and the four appearances of automation include a slide headed “Evidence-based criteria must power automation.” AMA slides

Version 6.0 gets a shorter window than this one. The packet’s timeline slide puts a 14-day exposure of the next version in October, against the 30 days named for the version open now.

One data point

A GLM, defined for the first time. The definitions section in version 5.0 calls it “a statistical modeling technique that extends linear regression to by relating predictive variables to an outcome variable using an assumed additive or multiplicative relationship through a link function and an appropriate probability distribution.” That sentence is quoted as it stands in the draft. Our reading is that a term with its own entry in an instrument’s definitions is harder to argue out of an exhibit’s scope than a term the instrument never names, which is a separate question from whether the technique carries risk. The governance the new guidance attaches to these models covers ground an AIS Program already answers for under the model bulletin.

The letter and the draft came in one packet, and only the draft is open to comment until September 29. If naming your own GLMs took more than one conversation this week, that difficulty is the comment. Say where it stalled in a reply to the email this issue arrived in.

– The Editor, InsureAI Wire