Who Keeps the Remedy?

The September 20 RealClearPolitics listing “No Blanket Immunity for AI Companies. Ever.” prompted this examination. Its linked page returned HTTP 403 during reporting, so its full argument could not be checked. This article does not treat the headline as evidence of a pending immunity bill, a company’s lobbying position or the present legal treatment of AI.

There is, however, a concrete precedent for examining how a technology’s liability protections are drawn. It is more useful than arguing over a blanket whose dimensions nobody has supplied.

The Boundary Is the Policy

Congress added Section 230 to the Communications Act in 1996. Its stated policies include promoting the development of interactive computer services and removing disincentives to blocking objectionable material. The statute pairs those ambitions with specific protections, definitions and exceptions.

Section 230(c)(1) says a provider or user of an interactive computer service shall not be treated as the publisher or speaker of information supplied by another information content provider. That last distinction matters. The statute defines an information content provider as someone responsible, wholly or partly, for creating or developing the information.

It also contains a separate protection for good-faith restriction of certain material. And it expressly preserves, among other things, enforcement of federal criminal law, while saying it neither limits nor expands intellectual-property law.

This is not a finding that Section 230 does—or does not—protect any particular AI output. Applying those provisions to a service, its conduct and a particular claim requires more than reading a definition. It is a narrower point: the statutory text does not say that belonging to a technology industry makes every legal claim disappear.

The historical parallel is therefore a design problem, not a verdict. Public policy can encourage a technology while specifying which conduct receives protection. “Innovation” is a rationale. It is not a complete set of legal boundaries, however convenient that would be for the drafting department.

Follow the Remaining Route

Any proposal for AI immunity should face three practical questions.

First, what is protected? Hosting another person’s material, developing a system, deploying it in a consequential decision and making promises about its performance are different activities. A serious proposal should make clear which claims its protection would cover rather than letting the label “AI company” do the work.

Second, what must be proved? Rejecting blanket immunity need not mean making developers liable whenever something goes wrong. A workable regime must distinguish an allegation from an established breach of a duty and connect responsibility to the relevant conduct. Otherwise, accountability becomes another word for assigning every loss to the largest available defendant.

Third, what remedy remains? If a proposal blocks a civil claim, its defenders should identify the alternative, if any: who can investigate, who can require correction, and who can secure compensation. An internal complaint channel should not be mistaken for independent adjudication.

These are proposed tests, not a description of an enacted AI liability regime.

The coherenceist stake is practical: people affected by a system need a way to make their evidence count against it. That does not require every complaint to succeed. It requires more than permission to submit feedback to the same organization whose conduct is disputed.

Predictability is a legitimate policy aim. But predictability for developers alone is an incomplete accounting. If a shield removes a route to remedy without replacing it, the uncertainty has not been abolished. It has been transferred to the person who must live with the loss.