Generalized linear models drew more objections than any other provision in version 5.0 of the NAIC's AI Risk Evaluation Supplement. Of the 33 submitters in the comment compilation posted October 2, six insurer trade groups (ACLI, APCIA, BCBSA, CAI, IRI and NAMIC) asked the Big Data and Artificial Intelligence (H) Working Group to exclude GLMs or stop counting them separately (NAIC, Oct. 2, 2026).
The working group took verbal comments on October 8. Its plan, per the August 31 minutes, is a version 6.0, a further 14-day comment period and two more public meetings before a version 7.0 goes up for adoption at the Fall National Meeting (NAIC, Aug. 31, 2026). On the record so far, the trade groups and the actuaries regulators invited to brief them in July disagree about what the GLM column in Exhibit A is supposed to capture.
Key Takeaways
- 33 submitters, 225 pages: the October 2 compilation carries six trade-group requests to remove GLMs or their separate count, while Monitaur, a model governance firm, wrote that it appreciates the inclusion of GLMs in the model taxonomy.
- Columns 4 and 7: Exhibit A now asks for GLM counts separately for models with direct consumer impact and models with material financial impact, across 13 program areas that include Reserves/Valuations, Catastrophe Triage and Reinsurance.
- 1 of 6 governance items: the Pinnacle Actuarial Resources letter filed alongside APCIA's maps data quality, development, validation, implementation and legal compliance to the 2020 CASTF white paper, and leaves ongoing performance with no section.
- A dozen variables, 100,000 records: past roughly that size, CAS panelist Henry Liu told regulators on July 22, a GLM coefficient can be observed but the reason it takes its value cannot be explained.
- $467 billion in premium: NAMIC, whose members write 61% of the homeowners and 53% of the auto market, asks that GLMs be "explicitly removed from scope of the Supplement."
What the Comment Letters Ask the Working Group to Change
NAMIC's letter (dated Sept. 29) recounts the drafting history: in an early version of the supplement, GLMs were a defined term and excluded from scope through the definition of artificial intelligence. Version 5.0, exposed August 31 for 30 days, reverses that. Its General Guidance calls GLMs "one such technique" among machine learning models and adds that, "as with other machine learning techniques, GLMs are not without risk of causing unfair discrimination or other adverse consumer outcomes" (NAIC, v5.0, Aug. 31, 2026).
Wisconsin Commissioner Nathan Houdek, the working group's chair, gave the reasoning when he opened the July 22 actuarial panel. A robust framework already exists for GLMs used in ratemaking, he said, but GLMs "are also used in other insurer operations, which have not been fully addressed by the current regulatory framework" (NAIC, Aug. 13 materials). The letters answer a narrower question than the one he posed.
| Commenter | Request on GLMs | Stated basis |
|---|---|---|
| ACLI | Exclude GLMs, GAMs and other transparent predictive models | The deployed model is "a fixed scoring formula or set of rules that does not learn" |
| APCIA | Distinguish conventional GLMs from AI systems | Decades of review; EIOPA's April 13, 2026 letter on GLMs and the AI Act |
| Pinnacle (filed with APCIA) | Exclude GLMs used for ratemaking | GLMs do not learn; the CASTF white paper covers their governance |
| BCBSA | Drop the separate GLM count in columns 4 and 7 | Counting a low-risk technique diffuses regulators' review capacity |
| CAI | Remove the emphasis on GLMs | GLM governance is misaligned with what agentic AI needs |
| IRI | Exclude GLMs and GAMs | Risks arise from governance and business practice, per EIOPA |
| NAMIC | Remove GLMs "explicitly" from scope | Regulated since the 1990s; GLM counts bury novel AI uses |
Source: NAIC comment compilation, Oct. 2, 2026.
Where Rate-Filing Review of GLMs Stops
The industry case rests on rate-filing review, and Pinnacle's letter, signed by Roosevelt C. Mosley Jr., FCAS, states it most precisely. GLMs "have been used for pricing for over 30 years," and the governance the supplement asks for is already covered by the CASTF Predictive Model White Paper released September 9, 2020, with state checklists in Arizona, Michigan, Florida, Connecticut, California, Washington and New York (Pinnacle, Sept. 29, 2026).
Pinnacle's own mapping table shows the seam. Data quality maps to white paper sections A.1, A.3 and A.4, and implementation to C.1 through C.8. Ongoing performance maps to nothing. Pinnacle answers that most companies review GLMs and pricing plans at least annually, which is company practice. An examiner working from the supplement asks for documents, and a filed rating plan is reviewed once, at filing, while the model keeps scoring renewals until the next filing.
The bigger seam sits in the other twelve rows of Exhibit A. A rate filing reaches the Ratemaking row only. A claims severity GLM that sets initial case reserves, a fraud score, a prior authorization model or a catastrophe triage model has no filed exhibit and no white paper section. Those are the "other insurer operations" Houdek named. The supplement's materiality definition, "the dollar amount above which the examiner's perspective of an insurer's financial position will be influenced," also points a reserving GLM toward column 7, the material financial impact count, where a personal auto rating plan rarely lands (NAIC, v5.0).
The actuaries on the July panel were less attached to the line than the trade groups. Liu said GLMs have an "overhyped reputation for transparency." The minutes record the Academy's Spencer Sadkin calling the question of whether GLMs are AI "the wrong discussion," since academia treats GLMs as machine learning and the label changes nothing for the consumer. Nevada's Gennady Stolyarov, a working group member, took the industry's side, calling GLMs static models whose output humans fully determine (NAIC, Aug. 13 materials).
Two Exclusion Tests With Different Reach
The letters offer two tests, and they would take different models out of Exhibit A. Pinnacle's is narrow: exclude GLMs "developed specifically for the purpose of supporting rating factors." A carve-out written that way removes the one row a filing already covers and leaves claims, fraud and reserving GLMs in columns 4 and 7, which concedes Houdek's point about other operations.
ACLI's test is broad. The deployed model, it argues, "is a fixed scoring formula or set of rules that does not learn, self-update, or change during use," and a model refit on newer data is "a separate, intentional development process rather than autonomous learning by the deployed model" (ACLI, Sept. 29, 2026). That description also fits a gradient-boosted tree frozen at deployment and refit on a schedule, which is how most supervised models reach production in pricing and claims.
Written into the definition of an AI model, the deployed-state test would empty column 6, "Other AI/ML Models," along with columns 4 and 7. Liu's July 22 point cuts the same way from the other side: a GLM fitted on a sampled subset of a very large dataset need not reproduce its coefficients either, so determinism is a property of how a model is trained and frozen, shared by most of the models in column 6. Whatever line version 6.0 draws for GLMs, it will be drawing it for those models as well.
Further Reading
- NAIC AI Risk Evaluation Supplement v5.0 Makes the Model Inventory an Examination Exhibit – the September analysis of what v5.0 changed, including the materiality threshold and the model inventory request.
- NAIC's July 22 Actuarial Panel Maps AI Governance Duties – the preview of the SOA, CAS and Academy panel whose minutes are quoted here.
- NAIC's AI Evaluation Tool Turns Bulletins Into a Scored Exam – how the supplement's predecessor ran in the state pilot.
- How Actuaries Validate AI Models for State Rate Filings – the filing-side validation work the trade groups say already covers GLMs.
- EU AI Act Omnibus Pushes Insurance AI Deadline to December 2027 – the European timeline behind the EIOPA letters the commenters cite.
Sources
- NAIC, AI Risk Evaluation Supplement, Version 5.0 (clean), exposed Aug. 31, 2026
- NAIC, AI Risk Evaluation Supplement Version 5.0: Comments Received, compiled Oct. 2, 2026
- NAIC Big Data and Artificial Intelligence (H) Working Group, Oct. 8, 2026 agenda with Aug. 31 minutes
- NAIC Big Data and Artificial Intelligence (H) Working Group, Summer National Meeting materials, Aug. 13, 2026 (July 22 panel minutes)