Platform Suppression
Decades-old internet liability law was written for message boards, not machine learning — and the gap is starting to matter.
NewsOnScale Staff
August 4, 2026
For thirty years, Section 230 of the Communications Decency Act has functioned as the load-bearing wall of the commercial internet. The law, passed in 1996 when most Americans still connected via dial-up modems, gave online platforms broad immunity from liability for content generated by their users. It enabled the modern web. It also, critics now argue, enabled modern harms.
The debate has returned with new urgency — not because the platforms have changed their behavior, but because the technology underlying them has changed entirely.
## What Section 230 Actually Says
The relevant text is deceptively simple: no provider of an interactive computer service shall be treated as the publisher or speaker of information provided by another content provider. Congress meant to encourage early platforms to moderate content without becoming legally responsible for every post they touched. The logic was sensible for 1996.
The problem is that AI systems don't receive and redistribute content from third parties. They generate it. A large language model responding to a user prompt is not a bulletin board passively hosting someone else's words — it is an active producer of new text, recommendations, and decisions. The 'another content provider' that Section 230 imagined simply does not exist in that transaction.
That distinction is more than semantic. It determines whether companies building AI products can be held accountable when those products cause harm.
## The Design Defect Opening
Courts are beginning to identify a related crack in the immunity framework, one that predates AI but has gained new relevance. Product liability law has long held that manufacturers can be sued not just for what a product does, but for how it was designed. Several recent legal challenges have argued that platform algorithms — recommendation engines, engagement optimization systems, content ranking mechanisms — constitute design choices, not neutral hosting decisions.
If a company makes a deliberate architectural decision to surface inflammatory content because it drives engagement, that is not the passive transmission of someone else's speech. It is an engineering choice with foreseeable consequences. Some courts are now agreeing.
The Massachusetts Supreme Judicial Court recently moved in this direction, finding that design-based claims against social media platforms are not automatically foreclosed by federal immunity law. That is a meaningful shift. It means plaintiffs can potentially argue that the harm came not from what a user posted, but from how the platform was built to respond to it.
## Why AI Sharpens the Stakes
Apply that framework to AI systems and the implications expand considerably. When a generative AI product provides harmful medical advice, facilitates manipulation, or produces discriminatory outputs, the liability question is immediate: who built the system that produced this, and what choices did they make in building it?
Those are not questions about third-party content. They are questions about product design, training data curation, output filtering, and deployment decisions. Section 230 was not written to answer them, and there is a credible argument that it cannot.
The commercial pressure to resolve this ambiguity in favor of immunity is enormous. AI development is expensive, and legal exposure is a cost companies work to minimize. But the public interest calculation runs the other direction. A liability framework that made sense for passive hosting is not automatically appropriate for systems that actively generate outputs affecting health, employment, financial decisions, and political belief.
## Accountability Without Overcorrection
None of this requires abandoning the original insight behind Section 230. Platforms genuinely should not be held responsible for every harmful thing a user posts, and a rule that made them liable for all third-party content would devastate open communication online. That concern remains valid.
But the question in front of courts, Congress, and regulators right now is narrower and more specific: does immunity designed for passive hosts logically extend to active AI systems that produce their own outputs? The honest answer is that it probably does not — and that designing a workable replacement will require the kind of careful, evidence-based rulemaking that the last thirty years largely skipped.
The liability vacuum around AI is not an accident. It is the predictable result of law failing to keep pace with technology. Closing that gap is accountability work, not censorship work. The difference matters.