The trial that opened in Oakland this week puts a question in front of a court that has mostly been argued in conference papers: what exactly is a recommendation algorithm optimising for, and who decided that? The legal argument depends on the technical answer, so it is worth setting out plainly.

A recommender is a prediction machine

A recommender system is a filter. Faced with far more content than any person could see, it predicts which items a particular user is most likely to engage with, and orders the feed accordingly. That is all it does, and it is genuinely useful: without one, a feed is an unusable firehose.

The consequential part is what the system is told counts as success. A model has to be trained against a measurable target, and the targets available are behavioural: time spent, items viewed, likes, shares, returns to the app. Nobody has a sensor for whether a user is better off. So the objective becomes engagement, not because anyone decided harm was acceptable, but because engagement is what can be counted.

Everything downstream follows from that substitution. The system does not know what a video is about. It knows that users like this one watched to the end.

Why researchers say the substitution matters

Academic work on recommender systems has argued for some years that optimising purely for engagement produces predictable side effects: reinforcement of whatever a user already responds to, narrowing of what they are shown, and disproportionate promotion of content that provokes a reaction. The literature on algorithmic radicalisation makes a stronger version of the claim for political content, arguing that feeds tuned to sustain attention tend to move users toward more extreme material.

There is also a body of work proposing alternatives — systems optimised against measures of user wellbeing rather than platform engagement — which exists precisely because researchers regard the current objective as a design choice rather than a technical necessity.

These are contested arguments, not settled findings. Effect sizes are disputed, causation is hard to establish outside a laboratory, and platforms hold the data that would resolve much of it. Treating the critique as proven would be as unwarranted as dismissing it.

Why lawyers care about the objective function

Here is where the technical detail becomes a legal one. The protections that shield platforms from liability were written around content — what users post. An objective function is not content. It is an engineering decision, documented in internal specifications, approved by named people, and measured against targets.

That is why the states leading the case against Meta have framed their claims around design rather than posts, as we set out in our report on the trial. If a company chose a target, measured the consequences, and kept the target, that is a corporate decision of the kind courts routinely assess. Whether the law will accept the distinction is the open question.

What regulators are doing instead

Legislatures have generally not tried to specify what a recommender may optimise for, which would be difficult to write and harder to audit. They have gone at it sideways.

One route is disclosure: the EU’s transparency rules, which came into force this month, require systems to identify themselves and synthetic content to be labelled. Another is age-based restriction: several US states have moved to limit algorithmic feeds for minors specifically, rather than regulating the algorithm itself. A third is access for outside researchers, on the theory that the argument cannot be settled while only one party holds the data.

Each approach has obvious weaknesses. Disclosure does not change behaviour on its own. Age limits depend on knowing who is a child, which platforms are poor at. Researcher access depends on the platform granting it.

The question worth holding on to

Whatever the Oakland court decides, the underlying point outlives the case. Any system that ranks what people see is optimising for something. The choice of that something is made by people, is written down, and could be made differently.

Asking what a feed is measuring is not a technical question for engineers. It is the whole argument, in a form a non-specialist can follow.

Sources

• Overview of recommender systems and how they are trained
• Overview of the algorithmic radicalisation literature and its critics
• Ognibene et al., Frontiers in Artificial Intelligence — recommender systems optimised for collective wellbeing