US Calls Australia's Algorithm Opt-Out 'Censorship.' The Fight Is Really About Who Sets Harm.

The US embassy in Canberra has filed a submission opposing Australia’s draft digital duty of care laws, calling the proposed right to switch off social media algorithms “censorship of protected speech.” The submission, reported by the BBC, argues that letting a government “enforce vague definitions of ‘harm’” would produce “viewpoint-based censorship,” and warns that an opt-out could “allow regulators to impose rigid, one-size-fits-all platform design requirements.”

Strip the free-speech framing and the more interesting claim sits underneath it. The US is not objecting to a toggle. It is objecting to a government writing requirements into the ranking function itself.

What the draft actually does

Under the proposal, tech firms risk hefty fines if they fail to give users the option to turn off algorithmic feeds. Prime Minister Anthony Albanese, speaking in New York during his UN General Assembly visit, framed it plainly: “It’s not about giving government control. It’s about giving people back control over what they receive on their devices.”

That is a defensible position, and it is also a narrower one than the US submission treats it as. A user-level opt-out that returns a chronological or subscription-only feed is a product feature. It is not, on its face, a speech mandate.

The US submission reaches further. It argues the laws risk “reducing the reach of independent journalists or other voices whose content touches on sensitive or controversial topics” if firms have to “pre-emptively demote” them, and it asks Australia to “clarify how exactly ‘harm’ and ‘risks’ shall be determined.” It also raises a jurisdictional point that matters more than the speech argument: “Mandated platform design features, especially when applied to algorithms, may affect what users see, say, and hear not just in or from Australia, but globally.”

That last sentence is the real story. Recommendation systems do not respect borders cleanly.

The AI layer nobody is naming

Every major social platform’s feed is a learned ranking model. The inputs are behavioral signals, the objective is engagement-weighted, and the output is an ordering. When a regulator requires an opt-out, it is not editing a policy document. It is requiring a second inference path, a second serving stack, and a second set of ranking guarantees that must hold for every user in the jurisdiction.

Australia’s under-16 social media ban, introduced last December, already forced platforms into age-assurance plumbing. The duty of care draft adds a ranking-layer obligation on top. Both are compliance engineering, and both land on the same teams that build and evaluate the models.

The US submission says the opt-out would “impose rigid, one-size-fits-all platform design requirements.” That is the industry’s standard objection, and it has a real technical core. Ranking models are not modular. You cannot cleanly excise “the algorithm” and leave a neutral feed behind, because the fallback ordering is itself a design choice with its own biases. Chronological is a ranking. Subscription-only is a ranking. The question is never whether a feed ranks. It is who sets the objective.

Two speeches, one week

The timing is not subtle. Albanese used his UN appearance to defend the tech policies. President Trump used the same stage to play down AI risk, rejecting guardrails and arguing the technology should be renamed “super intelligence” because “the use of the word artificial makes it sound fake.”

Read those two positions together and the shape of the coming fight is clear. One government wants to write harm definitions into platform design. Another wants to write them out. Both are making claims about how AI systems that mediate speech should be governed, and neither is offering a mechanism for measuring whether its preferred arrangement produces better outcomes.

The question is never whether a feed ranks. It is who sets the objective, and how that objective gets audited.

What is actually being contested

The US submission’s procedural asks are the tell. It wants Australia to clarify harm definitions, to address industry concerns that the “biases and preferences” of government officials do not “disadvantage independent or minority-viewpoint speech,” and to consider the effect on Australia’s “reputation as a jurisdiction that enables innovation.”

None of those are free-speech arguments in the strict sense. They are requests for the regulator to publish its evaluation criteria before the obligation takes effect. That is a reasonable ask, and it is one Australia has not yet answered.

The harder question is comparative. If a platform demotes a post under a harm policy, the user has no visibility into the decision. If a platform demotes a post under an engagement objective, the user has no visibility either. The US submission treats the first as censorship and the second as normal operation. That distinction does not survive contact with how these systems work.

Australia’s draft would at least create a user-facing control and a compliance surface. Whether the harm definitions are tight enough to avoid over-removal is a legitimate concern, and the US is right to press on it. But “vague definitions of harm” is a critique that applies equally to the engagement objectives platforms already optimize against without disclosure.

What this means for AI builders

Recommendation and ranking teams should read this as a preview of the compliance layer coming to their stack. The pattern is now familiar: a jurisdiction mandates a user control, the control requires a parallel inference path, the parallel path requires its own evaluation harness, and the evaluation harness requires documented definitions of harm and risk.

That work is expensive and it is not optional. The US submission’s global-spillover argument cuts both ways: if Australian requirements propagate to users elsewhere, so do the engineering obligations.

The open question is whether any of this produces measurable improvement in what users actually see. Neither the US submission nor the Australian proposal includes a framework for answering that. Until one of them does, the debate stays where it is, two governments arguing about who gets to set the objective while the ranking models keep running on objectives nobody voted for.