In September 2021, Frances Haugen, a former product manager at Facebook, delivered to the United States Senate a set of internal company documents she had copied before her resignation.[1] Among them was research conducted by Facebook’s own integrity team, showing that the platform’s recommendation algorithm systematically amplified content that provoked strong emotional reactions — anger, outrage, moral indignation — because such content generated more engagement. The team had identified the problem. They had proposed mitigations. The mitigations were not implemented. The decision, as Haugen testified, was commercial.
This is what editorial policy looks like when it is made by an optimisation function. No editor chose to amplify outrage. No editorial meeting weighed the public interest against the engagement metrics. The outcome was produced by a system designed to maximise a proxy variable — time on platform, clicks, shares — that correlates with revenue but not with any conception of informational value. The system worked as designed. The design was the problem.
The legal and regulatory frameworks that govern media in most democracies were built around a different model: a publisher who selects, edits, and takes responsibility for content. Platforms have spent a decade arguing that they are not publishers — that they are neutral conduits, mere infrastructure, no more responsible for what flows through them than a telephone company is for what is said on a call.[2] This argument has served them well in court and in legislatures. It has also become progressively harder to sustain.
The argument fails on its own terms. A telephone company does not sort calls by emotional valence and promote the ones most likely to provoke a reaction. A recommendation engine does exactly this. The selection is active, continuous, and consequential. When YouTube’s recommendation system consistently guided users from mainstream political content towards more extreme material — a pattern documented by researchers at Google and by independent academics — it was making an editorial choice.[3] The fact that no human reviewed each step in that journey does not mean no choice was made.
The consequences are not abstract. Research published in the American Journal of Political Science found that social media platforms, specifically Twitter and Facebook, contributed measurably to political polarisation in the United States between 2012 and 2018 — not primarily through the content users chose to seek out, but through what the platforms selected to show them.[4] A 2023 study by Meta’s own research team, published in Science and Nature, found that reducing algorithmic amplification on Facebook during the 2020 US election period decreased the reach of low-quality content without reducing user satisfaction.[5] The intervention was temporary. It was not made permanent.
Europe has moved further than the United States in naming this as a regulatory problem. The Digital Services Act, which came into full effect in February 2024, imposes specific obligations on very large online platforms: transparency about recommendation systems, the right for users to access a version of the feed not based on profiling, and risk assessments for systemic effects.[6] The first enforcement actions under the DSA are underway. Their outcome will determine whether the law has teeth or merely paperwork.
The question the DSA does not resolve is the one that sits beneath the regulatory argument: what would a recommendation system designed for purposes other than engagement maximisation actually look like? Researchers have proposed alternatives — systems optimised for informed satisfaction rather than immediate reaction, for source diversity rather than echo chamber reinforcement, for long-term wellbeing signals rather than session length.[7] None of these have been adopted at scale by a major platform. The incentive structure does not reward them.
The arrival of generative AI into the content layer of these platforms adds a further dimension. When the content being recommended is itself produced by language models — summaries, synthetic news, AI-generated video — the recommendation engine is no longer selecting among human expression. It is selecting among outputs that were themselves optimised for engagement. The feedback loop closes. What circulates is not what people said, but what a machine predicted would perform.
Calling this an editorial system is not a metaphor. It is a description. The question is not whether these platforms exercise editorial power. It is whether that power should be subject to any of the accountability structures we have historically applied to editors.