Healthcare AI Compliance Comparison

    Section 1557 vs SB 1120 Nondiscrimination Testing: One Test Plan or Two?

    In most cases, build one shared test plan core and attach two regime-specific overlays. Section 1557 and SB 1120 reach different entities and ask different primary questions: whether AI-supported programs discriminate, and whether AI is being used to deny coverage without physician supervision. A single plan can cover shared work such as AI inventory, input review and outcome analysis. It cannot replace separate treatment of entity scope and human-review evidence. Whatever you choose, pre-deployment bias testing does not show what the AI was allowed to do in production. That needs runtime policy enforcement and retained evidence.

    • Which tool, model version and policy version were in effect
    • What the policy decided: allow, redact, block or escalate
    • Whether a human reviewed the output, who that person was, and when
    • What the human approved, changed or rejected
    • Which test plan version and regime overlay the control maps to
    • Retention period and data residency settings for the record

    Two regimes, two different questions

    Section 1557 of the Affordable Care Act (42 U.S.C. 18116) prohibits discrimination on the basis of race, color, national origin, sex, age or disability in a health program or activity that receives federal financial assistance. It also reaches programs administered by an executive agency and entities established under title I of the ACA. HHS issued its implementing final rule on May 6, 2024. As of November 2026, the regulations at 45 CFR Part 92 are shown as effective from July 5, 2024, with a compliance schedule for particular sections. The rule also expects written policies and procedures, with an effective date, that are reasonably designed for compliance given the entity's size, complexity and the type of programs it runs.

    California SB 1120 (2023-2024) approaches AI from a different direction. State advisories describe it as amending the Knox-Keene Act and the California Insurance Code to limit how health care service plans use AI or other automated decision systems to deny coverage. They also describe it as requiring that licensed physicians supervise the use of AI tools.

    The practical difference is the organizing question. Section 1557 asks whether a covered program treats people differently on protected bases. SB 1120 asks who is making a coverage decision and whether an automated system alone is denying care. A test plan has to answer both questions, which is why the unified-versus-separate choice is not trivial. This page treats it as a requirements-mapping exercise, not a legal opinion.

    What a test plan can share, and what it cannot

    Several plan elements are regime-neutral and can be built once. These are an inventory of AI uses and the decisions they influence, a review of the input variables each tool consumes, analysis of outputs across groups relevant to the protected bases, a mitigation process with a named owner, and a change-control step that triggers retesting when a model, prompt or data source changes. Both regimes benefit from this core.

    Other elements need separate treatment. Entity scope differs, so an organization that is both a federally funded provider and a health care service plan must map each tool to each entity and each regime. Human review also differs in shape. Under SB 1120 the central artifact is evidence of physician supervision on coverage-related decisions. Under Section 1557 the central artifact is the documented policy framework and the discrimination analysis behind it. Reporting audiences differ too, so the same analysis may need to be packaged in two forms.

    The common failure is assuming that a strong bias evaluation satisfies the supervision concern. It does not. An accurate, well-balanced model can still deny coverage with no physician in the loop.

    Deciding between a unified plan and two plans

    A unified plan with overlays is defensible when the same tools, entity and patient population fall under both regimes, and when one governance owner can maintain the plan. In that case the shared core reduces duplicate testing and keeps findings consistent. Inconsistent findings across two plans for the same tool are themselves an exposure.

    Separate plans are easier to defend when the regimes touch different entities, for example a provider organization with no plan business alongside a plan that uses AI only in utilization review. They also make sense when different teams own clinical decision support and coverage operations, since a single document with no single owner tends to go stale.

    Use four criteria to decide. First, tool overlap: does any one tool fall under both regimes? Second, entity overlap: does one legal entity carry both obligations? Third, ownership: can one function maintain the plan through every model change? Fourth, evidence reuse: can the same records satisfy both reviews without being rewritten? If the answer to most is yes, build the shared core and two overlays. If not, keep two plans and align their shared methods. Route the final interpretation to counsel.

    Where Trussed fits, and where it does not

    Trussed is not a pre-deployment testing or bias-evaluation product. It does not run dataset or model bias analysis, and it does not replace the shared core of either test plan.

    Trussed addresses the runtime and documentation portions. It enforces policy in the execution path across models, agents, tools and MCP servers, with actions to allow, redact, block, rate limit or escalate for human approval. Enforcement applies to individual interactions and tool calls, so a policy that requires physician approval before a certain output proceeds is enforced at runtime, not reviewed after the fact. That enforcement generates audit-ready evidence, mapped to frameworks including HIPAA, NAIC, NYDFS, the EU AI Act and NIST. Deployment is self-managed in the customer's cloud, with configurable retention and data residency.

    This supports the record described above under either a unified or a separate plan. It does not determine whether a tool qualifies under Section 1557 or SB 1120, and it does not substitute for legal review.

    One core plan, two overlays

    Shared core

    AI use inventory, input-variable review, outcome analysis, mitigation ownership.

    Section 1557 overlay

    Federally funded health programs and activities; discrimination on protected bases.

    SB 1120 overlay

    Health care service plans and insurers; limits on AI-based coverage denials; physician supervision.

    Runtime layer

    Enforced policy and retained evidence of what the AI was allowed to do and did.

    Define the runtime evidence your test plan needs

    Walk through how policy enforced in the execution path and audit-ready evidence can support the runtime and documentation portions of your plan.

    Talk to an Expert