ASOP 12's Unintended Bias Section Rewrites Pricing Model Rules

From tracking ASOP revisions across multiple standards-setting cycles, the proposed rewrite of ASOP No. 12 stands apart. The existing standard, last revised in December 2005, predates the widespread adoption of generalized linear models in personal lines pricing, the explosion of third-party data sources, and the regulatory wave that followed Colorado's SB 21-169. Remarkably, the 2005 standard does not contain the word "model" anywhere in its body (ASB, ASOP No. 12, 2005). The September 2023 exposure draft addresses that gap directly, and its most consequential addition is an entirely new section, 3.4, on Potential for Unintended Bias, that creates professional guidance where none previously existed (ASB, Exposure Draft, September 2023).

The comment deadline closed May 1, 2024, with 57 letters received from organizations including the CAS, NAIC, American Academy of Actuaries, Allstate, Travelers, USAA, and Verisk (ASB, Comment Letters). Given the volume and complexity of comments, a second exposure draft or final adoption is expected in mid-to-late 2026. But the regulatory environment is not waiting for the standard to be finalized. The NAIC's Model Bulletin on Use of Artificial Intelligence Systems by Insurers (adopted December 2023, now adopted in at least 24 states) and Colorado's SB 21-169 already require insurers to demonstrate that algorithmic systems do not produce unfairly discriminatory outcomes (NAIC, December 2023; Colorado DOI). For pricing actuaries filing GLM-based rating plans, the practical question is not whether to build a bias-testing framework but how to build one that satisfies both the emerging professional standard and active regulatory mandates.

Why the 2005 Standard Could Not Hold

ASOP No. 12 was originally adopted in 1989, when risk classification relied on manually defined rating tiers and limited univariate analysis. The 2005 revision updated the standard's language and expanded its applicability across all practice areas, but its analytical framework assumed that actuaries could evaluate each rating variable's effect in isolation. Section 3.2 of the 2005 standard discusses statistical criteria, operational criteria, and social criteria for risk classification without acknowledging that variables interact in complex, non-linear ways when combined in a multivariate model.

That assumption was reasonable in 2005. Personal auto rating plans typically used five to ten multiplicative factors, and most commercial lines pricing was performed using judgment-based class plans with loss cost relativities derived from limited fluctuation credibility. Today, a typical personal auto GLM includes 15 to 30 rating variables with interactions, and ML-augmented models in homeowners and commercial auto incorporate hundreds of features. The 2005 standard's univariate framing left a structural gap: an actuary could evaluate credit score, territory, and vehicle age individually against the standard's criteria and conclude each was justified, while the combined model produced systematic pricing disparities correlated with protected-class status that no single-variable review would detect.

The CAS's 2022 Research Paper Series on Race and Insurance Pricing documented this gap with empirical evidence. Their research found that individually defensible rating variables, when combined in multivariate models, can produce aggregate pricing disparities of 11 to 17 percent for applicants in predominantly minority zip codes compared to applicants with identical risk profiles in predominantly non-minority zip codes (CAS, 2022). The interaction effects that generate these disparities are precisely what the 2005 standard's univariate framework could not capture.

What the Exposure Draft Changes

The September 2023 exposure draft restructures the standard around four material changes that collectively redefine the pricing actuary's professional obligations.

New definitions expand the analytical scope. The draft introduces two new defined terms. "Risk measure" replaces the existing standard's repeated references to "expected outcomes," which could lead an actuary to focus exclusively on the mean of the loss distribution. The new definition reinforces that actuaries should consider aspects of the loss distribution beyond expected value, including variance, tail behavior, and distributional effects across subgroups (ASB, Exposure Draft, Section 2). "Unintended bias" is defined as an emerging issue requiring actuaries to evaluate whether their risk classification systems produce classification effects that were not intended by the system's design. The ASB characterizes this as requiring actuaries to monitor the evolution of this issue and its potential impact on their work (Academy of Actuaries, Professionalism Counts, February 2024).

Section 3.2.4 on Multivariate Effects closes the interaction gap. This new section acknowledges that variable interactions in GLMs and ML models can produce classification effects invisible in univariate review. For the pricing actuary, this means that evaluating each rating variable's loss ratio relativity in isolation no longer satisfies the standard. The actuary must consider how variables interact and whether those interactions produce classification outcomes that would not be apparent from examining each variable independently. This is particularly relevant for territory-by-credit-score interactions, age-by-vehicle-type interactions, and other combinations known to correlate with protected-class demographics.

Section 3.4 on Potential for Unintended Bias creates new professional guidance. This entirely new section requires actuaries to consider the potential for unintended bias in risk classification systems. The guidance stops short of mandating specific statistical tests or threshold values, but it establishes that an actuary who designs, evaluates, or opines on a risk classification framework must consider whether that framework produces unintended disparate impacts. The section operates as a professional obligation: the actuary cannot simply assume that facially neutral variables produce facially neutral outcomes.

Scope expansion pulls in peer reviewers and appointed actuaries. The 2005 standard applied primarily to actuaries who designed risk classification systems. The revised standard broadens the list of activities to include "developing, selecting, evaluating, or opining on any elements of a risk classification framework." This language brings peer reviewers, appointed actuaries who sign Statements of Actuarial Opinion on rate filings, and consulting actuaries who review third-party models into the compliance chain. If you review or opine on a rating plan that uses predictive models, the revised ASOP No. 12 applies to your work.

A Three-Stage Bias Audit for GLM Rate Filings

Section 3.4 does not prescribe a specific testing methodology. The guidance is principles-based, leaving implementation to the actuary's professional judgment. From working through the intersection of the exposure draft's requirements, ASOP No. 41's communication standards, and the regulatory frameworks already in force, a three-stage audit framework emerges as a practical template for meeting the standard's new obligations in the context of a GLM-based rate filing.

Stage 1: Variable-Level Proxy Screening

Before evaluating the model's aggregate output, the actuary screens each rating variable for its correlation with protected-class indicators. The goal is to identify which variables in the model are most likely to serve as proxies for race, ethnicity, gender, or income status.

For categorical variables (territory groupings, credit tiers, occupation codes), the standard association metric is Cramer's V, which measures the strength of association between two categorical variables on a 0-to-1 scale. For continuous variables (credit score, years of driving experience, property age), Pearson correlation coefficients quantify the linear relationship with protected-class indicators or their proxies.

The screening step requires demographic data. Where protected-class data is unavailable at the policyholder level (the norm in most P&C applications), the CAS research recommends Bayesian Improved Surname Geocoding (BISG) to impute race and ethnicity probabilities from surname and census-tract demographics (CAS, 2022). Colorado's Regulation 10-1-1, effective for life insurers since November 2023 and in active rulemaking for auto and health, explicitly endorses BISG as an acceptable demographic estimation method.

A practical materiality threshold for the proxy screen depends on the regulatory context. Colorado's four-part bias testing methodology uses the four-fifths rule as one benchmark, but the proxy screening stage operates upstream of that test. Variables with Cramer's V above 0.15 or Pearson correlation above 0.20 with imputed protected-class indicators warrant closer scrutiny in Stage 2. These thresholds are not regulatory requirements; they represent a reasonable starting point for the pricing actuary's professional judgment.

Stage 2: Model-Level Disparate Impact Testing

Stage 2 shifts the analysis from individual variables to the model's aggregate output. Using the fitted GLM, the actuary computes predicted loss ratios, pure premiums, or premium-to-expected-loss ratios for protected-class subgroups and compares the distributions.

The core metric is the adverse impact ratio: the ratio of the outcome rate for the least-favored group to the outcome rate for the most-favored group. The EEOC's four-fifths rule, originally developed for employment selection procedures under the 1978 Uniform Guidelines, provides one possible benchmark. Under this framework, an adverse impact ratio below 0.80 (meaning the least-favored group's rate is less than 80 percent of the most-favored group's rate) is generally regarded as evidence of disparate impact.

In an insurance pricing context, the "outcome rate" is typically the ratio of the subgroup's average predicted premium to the subgroup's average predicted loss cost. If the model charges one demographic subgroup a systematically higher premium relative to its expected losses than it charges another subgroup, that differential is the disparate impact to be evaluated.

This is where Section 3.2.4's multivariate effects guidance becomes operationally critical. A territory variable might pass the Stage 1 proxy screen with a Cramer's V of 0.12 against imputed race, below the 0.15 threshold. But when that territory variable interacts with credit score in the GLM, the combined effect on predicted premiums could produce an adverse impact ratio of 0.74 for a specific racial subgroup, failing the four-fifths benchmark. The interaction effect was invisible at the variable level; it only surfaces in the model-level test. The 2005 standard's univariate framework would have missed it entirely.

The CAS Research Paper Series provides practitioners with tested methodologies for this stage, including propensity score matching, standardized mean difference calculations, and permutation-based fairness tests that account for legitimate risk differences across subgroups (CAS, 2022). The key distinction from employment law applications is that insurance pricing has a statutory basis for treating different risks differently: actuarially justified rate differentials are lawful. The question Section 3.4 poses is whether the observed differentials are fully explained by legitimate risk factors or whether residual disparities remain after controlling for actuarial risk.

Stage 3: Actuarial Justification and Documentation

Where Stages 1 and 2 identify potential disparities, Stage 3 requires the actuary to determine whether those disparities are actuarially justified and to document the analysis under ASOP No. 41's communication standards.

Actuarial justification means demonstrating that the rating variable's predictive power for expected losses is sufficient to support the observed disparate impact. For a territory variable that produces a 0.74 adverse impact ratio after interaction with credit score, the actuary must show that the territory variable's inclusion in the model is supported by loss experience data, that the variable's coefficients are statistically significant and stable across validation samples, and that removing or modifying the variable would materially degrade the model's predictive accuracy for the affected subgroups.

This is not a new analytical concept for pricing actuaries. Rate filings have always required demonstrating that rating variables are supported by loss data. What Section 3.4 adds is the requirement to frame that justification explicitly in terms of its relationship to unintended bias: the actuary must document not only that the variable is predictive but that the actuary considered whether the variable's predictive power adequately justifies its disparate impact on protected-class subgroups.

The documentation requirements flow through ASOP No. 41, which is itself undergoing a concurrent revision with a second exposure draft released in 2025. Under the current ASOP No. 41, the actuary must prepare documentation sufficient for another qualified actuary to assess the reasonableness of the work. Under the revised ASOP No. 12, this means the file must include: the proxy screening methodology and results, the model-level disparate impact test methodology, thresholds, and results, the actuarial justification for any identified disparities, and disclosure of limitations in the analysis (data quality, BISG imputation uncertainty, threshold selection rationale).

Alignment With ASOP 53 and the CAS Comment Letter

The CAS's comment letter on the exposure draft raises a structural question that pricing actuaries should track closely. ASOP No. 12 governs the grouping of risk subjects into risk classes. ASOP No. 53 governs the estimation of future costs for each risk class. In practice, these two activities are performed simultaneously in a GLM: the model both defines risk segments (through variable selection and interaction specification) and estimates the cost for each segment (through coefficient estimation).

The CAS recommends that the ASB clarify the boundary between ASOP No. 12 and ASOP No. 53 to eliminate overlap and gaps. Specifically, ASOP No. 53 section 1.2 states that it applies to "developing or reviewing the future cost estimates by class within a risk classification system," which overlaps with the classification design work covered by ASOP No. 12. When both standards require the actuary to consider unintended bias, but through different analytical lenses, the documentation burden could expand significantly unless the final standards are coordinated (CAS Comment Letter, April 2024).

For the pricing actuary building a bias-testing framework today, the practical implication is to design the three-stage audit so that it can satisfy both standards simultaneously. The variable-level proxy screening and model-level disparate impact testing address ASOP No. 12's classification concerns. The actuarial justification analysis, which evaluates whether the model's cost estimates by class are supported by experience data, addresses ASOP No. 53's cost estimation concerns. A unified framework avoids duplicating effort if the final standards draw the boundary differently than the exposure draft.

The Regulatory Context: NAIC and Colorado Are Not Waiting

The regulatory environment has outpaced the standard-setting process. Two frameworks already impose substantive bias-testing obligations on rate-filing actuaries, making ASOP No. 12's new requirements immediately practical even before formal adoption.

NAIC Model Bulletin on AI (December 2023). Adopted by at least 24 states and the District of Columbia as of mid-2026, the bulletin requires insurers to develop, implement, and maintain written programs for responsible AI use. Insurers must adopt governance frameworks, risk management protocols, and testing methodologies designed to ensure that AI systems do not produce unfairly discriminatory outcomes. The bulletin's guiding principles require AI use in insurance to be fair, ethical, accountable, compliant, transparent, secure, safe, and robust (NAIC, December 2023). For actuaries, the bulletin creates an external expectation that any predictive model used in pricing has been evaluated for discriminatory effects.

Colorado SB 21-169 and Regulation 10-1-1. Colorado requires insurers to test their algorithms, predictive models, and external consumer data sources for unfair discrimination against nine protected classes. Life insurers faced the first compliance milestones: risk management framework due November 2023, first progress report June 2024, first full compliance attestation December 2024. Auto and health insurance rulemaking is active, with Regulation 10-1-1 amendments effective October 2025. The regulation prescribes a four-part bias testing methodology that includes the four-fifths rule, proxy variable audits, intersectional testing, and counterfactual analysis (Colorado DOI, SB 21-169). BISG is accepted for demographic estimation.

The three-stage audit framework described above maps directly to these regulatory requirements. Stage 1's proxy screening satisfies Colorado's proxy variable audit. Stage 2's disparate impact testing satisfies the four-fifths rule and intersectional testing components. Stage 3's documentation satisfies both the NAIC bulletin's governance and transparency requirements and Colorado's annual compliance attestation. Building the framework now, aligned with the ASOP No. 12 exposure draft's Section 3.4 principles, positions the pricing actuary for compliance across all three regimes simultaneously.

What Pricing Actuaries Should Do Now

The standard is not yet final. But the regulatory obligations are current, the exposure draft signals clear professional direction, and the analytical infrastructure required for compliance takes time to build. Patterns from previous ASOP adoption cycles suggest that once a standard reaches the comment-response stage, the final version typically preserves the core analytical requirements even as specific language is refined.

First, inventory every rating variable in current GLM-based rating plans and assess each for protected-class correlation using BISG-imputed demographics. Where BISG is not feasible (commercial lines, specialty programs), document the data limitation and the alternative assessment methodology employed.

Second, build the model-level disparate impact test as a post-modeling diagnostic in the existing ratemaking workflow. The test should run automatically on each model iteration, producing adverse impact ratios by subgroup that the actuary reviews before finalizing rate relativities. Automating this step reduces the marginal cost of compliance and creates an audit trail.

Third, update rate filing documentation templates to include a dedicated bias-analysis section. Under ASOP No. 41's documentation standards, the analysis should be structured so another qualified actuary can assess the reasonableness of the methodology, thresholds, and conclusions. Include a summary of which variables were flagged at Stage 1, the Stage 2 adverse impact ratios, and the actuarial justification for any disparities that remain in the final model.

Fourth, coordinate with appointed actuaries and peer reviewers. The revised ASOP No. 12's expanded scope means that actuaries who review or opine on rating plans must also consider unintended bias. If the pricing actuary has not documented the bias analysis, the reviewing actuary has no basis to evaluate it, creating a compliance gap for both parties.

The ASB is processing 57 comment letters and transitioning the task force leadership from Patricia Matson to Lisa Slotznick (Academy of Actuaries). The timeline for a second exposure draft or final adoption remains uncertain but is expected in mid-to-late 2026. Regardless of when the standard is finalized, the pricing actuary who builds the bias-testing framework now will be ahead of both the professional standard and the regulatory curve.