What version 6.0 did with the comment letters

A company that answers the AI Risk Evaluation Supplement will still find Exhibits C and D in version 6.0, an option for regulators to leave GLMs out of Exhibits B to D, and a rewritten definition of an Adverse Consumer Outcome. The NAIC’s Big Data and Artificial Intelligence (H) Working Group takes written comments on this exposure draft until October 30. Our October 20 report walks through more of the draft’s edits. Two NAIC staff members, Miguel Romero and Scott Sobel, left margin notes in the tracked-changes copy naming whose input some edits address. Where an edit below carries a commenter’s name, it comes from those notes, which are staff annotations and not working group decisions.

This week’s story: what 6.0 did with Issue 13’s three questions

Issue 13’s story asked whether the next draft would keep Exhibits C and D, keep GLMs in scope and take up either proposed addition. 6.0 answers all three in its text.

ACLI asked that Exhibits C and D not be used broadly until tested by more piloting or a phased rollout; the Insured Retirement Institute asked for removal or deferral, NAMIC for removal. Both exhibits remain in 6.0. A new instruction, with the note “ACLI related input,” states the intention that regulators “will most often use Exhibit A.” C and D inquiries may follow for “a subset” of a company’s AI uses, “where regulators wish to probe further,” so regulators pick the subset.

All three also asked that GLMs be taken out of the supplement, and 6.0 keeps them in. It still says GLMs “are not without risk of causing unfair discrimination,” and a new sentence just below says “regulators may choose to exclude some or all GLMs for Exhibit B through D related inquiries if appropriate.” The choice stops short of Exhibit A, whose table keeps a GLM column.

Of the two proposed additions, Connecticut’s proxy question is in 6.0, as Exhibit C’s item 10a (11a as proposed). It asks, “Where applicable,” whether a company evaluates inputs or outputs for proxy effects. Its note reads “New question proposed by CT.” The consumer representatives asked for a uniform definition of high-risk models; 6.0 has none. Exhibit C still says “AI model risk criteria are set by the insurance company.”

What to do this week

• We would keep GLMs in whatever you prepare for Exhibit A. For Exhibits B to D, the reason the new sentence gives for letting regulators set GLMs aside is “existing regulatory oversight provided via rate making.” A list made now of each GLM and the rate filings that use it puts that oversight on one page.

• Give the most time to an edit that falls outside Issue 13’s three questions and has no margin note beside it. Adverse Consumer Outcome becomes Adverse Consumer Impact, and the definition loses its closing words “in a manner that violates those standards.” It keeps the condition that the decision be one “that is subject to insurance regulatory standards enforced by the Department.” We read the deletion as bringing into the term any of your AI uses whose decisions work against a consumer while complying with those standards. The wider term reaches into Connecticut’s item 10a, whose proxy question covers effects that “may contribute to Adverse Consumer Impacts.” Go through the uses your supplement answers describe and mark the ones that would count only under the new wording.

On the Docket

• October 26, 2026: Colorado’s Attorney General holds the hearing on its proposed ADMT and chatbot rules at 10:00 AM MT. The October 6 interim draft was the newest text on the rulemaking page on October 20. Written comments close the same day at 11:59 PM MT. Rulemaking page

• October 30, 2026: Written comments on 6.0 go to Sobel or Romero and are due by close of business. The working group’s page gives no time zone. NAIC working group page

• November 2, 2026: The working group meets by Webex at 10:00 AM CT. Without naming a version, the listing gives the purpose as “to continue public discussions on the AI Risk Evaluation Supplement.” NAIC working group page

• November 16, 2026: The working group’s session at the NAIC Fall National Meeting runs from noon to 1:00 PM CT. The working group’s page did not mention version 7.0 on October 20, but a timeline in the August 31 meeting packet planned to consider 7.0 for adoption at this session. August 31 meeting packet

This week in brief

Governance

Exhibit B’s narrative version adds four questions a regulator may ask about how a company finds unsanctioned AI use, keeps nonpublic personal information from unapproved models, governs enterprise assistants and decides when a general-purpose tool’s workflow makes it a high-risk use; a note credits them to InsurTech Advisors. Its item on uses directly affecting consumers no longer asks whether staffing was cut when an AI system was introduced (ACLI and BCBSA) and now asks how a company assures itself human review is appropriate (consumer representatives). Version 6.0, tracked changes

From the Guides

For California business. Our new guide works out which of an insurer’s AI systems fall under the California Privacy Protection Agency’s rules on automated decisionmaking technology and which remain with the Insurance Code: Does the CCPA Apply to Insurance Companies? Where California’s AI Rules Stop.

Which item in 6.0 did you have to read twice? Reply with the exhibit and item number.

– The Editor, InsureAI Wire