Colorado’s SB 26-189 and Executive Order 14365

I · Statute

Colorado’s SB 26-189

SB 26-189 rewrites Colorado’s approach to automated decision-making, replacing the more stringent requirements of SB 24-205 — the first major state-level AI legislation in the country. It governs covered automated decision-making technology (covered ADMT), defined as technology that processes personal data and uses computation to generate output used to “materially influence” a “consequential decision.”

In doing so it strips out much of what SB 24-205 required: the duty of care to avoid algorithmic discrimination, mandatory risk-management programs, and reporting to the attorney general, among other things. In their place it imposes three obligations on decision-makers.

1. Notice before use. Decision-makers must notify consumers before covered ADMT is used to make a consequential decision.
2. Post-adverse-outcome notice. If covered ADMT materially influences an adverse decision, decision-makers must give notice within 30 days. That outcome notice must include a plain-language description of the decision, the role the ADMT played in it, instructions for requesting further information about the ADMT and the inputs used, and an explanation of the consumer’s rights and how to exercise them.
3. Consumer-rights process. After an adverse decision, consumers retain the right to access the personal data used and to request meaningful human review and reconsideration — “to the extent commercially reasonable.”

Thoughts & objections

I appreciate that the bill protects consumers who are adversely affected by decisions made by opaque models, and that it is among the first of its kind in state AI legislation. Still, I have two worries.

Objection 1. The input data disclosed to consumers may not actually be useful, because not all of it causally contributes to the decision. To tell whether some input — race, say — is playing a causal role it should not, you need an AI expert on hand to analyze the case, or interpretability techniques such as building a causal abstraction of the model to identify which variables are genuinely causally relevant.

The most consequential part of the process applies only “to the extent commercially reasonable.”

Objection 2. The most impactful part of the consumer-rights process — human review and reconsideration — applies only to the extent that it is commercially reasonable, which may almost never be the case in, say, hiring: the decision-maker will often have already found another candidate by the time the process ends.

II · Executive Order

Executive Order 14365

EO 14365 was signed by President Trump to override a more aggressive earlier order on AI policy. It calls for a “minimally burdensome national policy framework for AI,” on the reasoning that America leads in AI development and that a thicket of robust state laws makes compliance — and therefore deployment — intolerably difficult across states.

It establishes an AI Litigation Task Force to challenge state laws that fail the “minimally burdensome” standard, on the grounds that they:

1. unconstitutionally regulate interstate commerce;
2. are preempted by existing federal regulation; or
3. are otherwise unlawful in the attorney general’s judgment.

The Secretary of Commerce must publish an evaluation of current state AI laws, identify those that fail the standard, and refer them to the task force.

Thoughts & objections

This dovetails with what is sometimes called the socio-technical problem of alignment: even if we solve the technical and normative problems of alignment, there remains the problem of coordinating resources and political capital to ensure that future AI is aligned with human values. We should be wary of orders like this one, which threaten to undermine the risk-management policy that may be necessary to raise the probability that the systems we deploy are well aligned.

Objection 1. The AI Litigation Task Force threatens to stifle any attempt to build protections — like SB 26-189 — for consumers facing adverse decisions influenced by opaque models. I am for removing barriers to AI development and deployment in the U.S. to bolster our national and economic security. But we need to strike a balance between a minimally invasive framework and legislation that protects American citizens, and mobilizing an entire task force to suppress state legislation, where little already exists, lands too far on one side of that balance — far enough that we risk infringing the rights of American citizens and consumers.