How are recommendation algorithms being regulated?

Through transparency, risk assessment and user choice obligations rather than through rules about what an algorithm may recommend — regulators have largely avoided dictating ranking, and have instead required platforms to explain, assess and offer alternatives.

The main obligations now appearing in law:

Explaining the main parameters. Platforms must set out, in terms users can understand, what the principal factors in recommendations are and their relative importance.

Offering a non-profiling option. Very large platforms must provide at least one recommendation option not based on profiling — which is why chronological or following-only feeds have reappeared as a setting.

Risk assessment. Designated platforms must assess systemic risks arising from their systems — including to fundamental rights, civic discourse, public health and minors — and take mitigation measures, with independent audit.

Researcher data access, on vetted terms, which addresses the long-standing problem that platform effects could not be studied independently.

Restrictions on profiling minors, including bans on targeted advertising to children based on profiling.

Prohibitions on using sensitive data categories for recommendation targeting.

Dark pattern prohibitions, covering interface design that distorts choice.

Transparency reporting on moderation and recommendation.

Where the difficulties are, and they are substantial:

Explaining a machine learning system meaningfully is genuinely hard — a truthful description may be uninformative, and a comprehensible one may be inaccurate.

Non-profiling options are offered and rarely used, since they are frequently worse and are not the default. Default design decides behaviour, and mandating availability without mandating defaults changes little.

Enforcement capacity against very large firms.

Defining harm in a way that does not become content regulation by another route.

Jurisdictional scope, where obligations apply differently by market, producing divergent experiences.

What individuals can actually do now: switch to a following or chronological feed where offered; reset or clear recommendation history; use the "not interested" controls, which do train the system; and review advertising and profiling settings, which are usually several layers deep.