Social media platforms should be liable for algorithmic amplification of harmful content

The Frances Haugen documents are the most important evidence here. Internal Facebook research showed that algorithmic engagement optimisation increased the share of 'divisive' and 'misinformation' content in user feeds because that content generates more reactions. The company knew this. They deliberately chose engagement metrics over health metrics because the engagement metrics drove ad revenue. This is the definition of negligence — known risk, conscious disregard. The documents make the liability case almost in themselves.

The key distinction: Section 230 immunises platforms from liability for hosting third-party content. That's defensible. Algorithmic amplification is different — the platform is making an active editorial choice to show content to people who didn't ask for it. When Facebook's algorithm decides to show outrage-generating vaccine misinformation to parents who searched for 'baby rash,' that's a recommendation the platform made. The resulting harm is partially the platform's action, not just a neutral hosting decision. Tort liability applies to negligent recommendations in almost every other context.

The vagueness objection is a design problem, not a reason to abandon liability entirely. Products liability in physical goods also faced definitional problems at first: what's an 'unreasonably dangerous defect'? Courts and regulators defined it iteratively. Algorithmic liability can work the same way: demonstrably false health information, content that platform research has specifically identified as causing harm, incitement meeting existing legal standards. None of these require screening all content — they require not actively promoting content that the platform's own internal research has flagged.

Liability for algorithmic amplification would end algorithmic recommendation as a product feature, which would mean chronological feeds or manual curation. The social media products people actually want and use depend on algorithmic personalisation. Imposing liability creates an incentive to show users only content that could never possibly be considered harmful — i.e., vapid content, or no recommendations at all. Platforms facing uncertainty about what's 'harmful' enough to generate liability will over-remove. The collateral damage to legitimate speech would be enormous.

Products liability analogy fails: physical defects exist at manufacture, can be identified ex ante, and harm a defined class of users in a defined way. Algorithmic 'harm' is contextual, probabilistic, and disputed. The same post about ivermectin is harm to some, essential information to others. Courts and juries are terrible at adjudicating this. The actual mechanism for change is transparency requirements (disclose the algorithm, disclose engagement metrics, disclose internal research) and FTC enforcement for deceptive practices — not tort liability that would inevitably be weaponised by whoever can afford the most lawyers.

The Haugen documents also show something liability advocates don't like: the internal research found that algorithmic amplification of misinformation was a small percentage of overall feed content, not a dominant driver of user experience. The Integrity team's findings were exaggerated in press coverage. More importantly, what regulatory standard would define 'harmful' in a way that's (a) constitutionally defensible, (b) not trivially gameable, and (c) doesn't require platforms to pre-screen every amplified post? No one has specified this, and until they do, the liability concept is too vague to implement.