What actually governs your feed in 2026

For a decade, the case against recommendation feeds was the filter bubble: algorithms wall each of us into a comfortable echo chamber, and polarization follows. It’s a tidy story. The best evidence we have says it’s mostly wrong, and the reasons it’s wrong are more unsettling than the original worry.

The chronological feed is not the fix

In July 2023, a set of studies run jointly by Meta and independent academics landed in Science and Nature. The headline experiment, Guess et al., took tens of thousands of Facebook and Instagram users off the algorithmic feed and gave them a reverse-chronological one during the 2020 US election. The chronological feed increased exposure to untrustworthy content, by 67% on Facebook, and produced no detectable change in polarization, political knowledge, or participation over three months. A companion Nature paper found like-minded content was common but not the cause of polarization.

Two caveats keep this from being a platform exoneration. Meta made dozens of “break-glass” changes to its algorithm after the 2020 election, which muddies the comparison, and three months is too short to catch effects that built over a decade. But the core finding holds up and it’s the one that matters for policy: mandating a chronological feed, the reform everyone reached for, makes the information problem worse, not better.

So if the bubble was the wrong thing to worry about, what’s the right thing?

Feeds move what you believe, on the specifics

In February 2026, a field experiment on X published in Nature gave one answer. Over seven weeks, users on the standard algorithmic feed shifted their policy attitudes toward more conservative positions relative to a chronological control. It didn’t change their partisan identity or how much they disliked the other side, the deep constructs the filter-bubble story fixated on. It changed what they thought about specific issues. A 2025 Science study went further and showed that deliberately reranking a feed to cut partisan animosity can move affective polarization too.

Put those together and the picture inverts. The feed isn’t a passive mirror that traps you with your own opinions. It’s an active instrument that can nudge what you believe about a tax policy or an immigration rule, and the operator controls the dial. That’s a governance question about who holds the dial and what they optimize it for, not a psychology question about echo chambers.

The pipes are full of slop

The 2019 version of this essay could assume the content in the feed was made by people. That assumption is gone. By early 2025, most social content was AI-generated, and analysts estimate more than one in five algorithmically recommended YouTube videos are now AI-generated “slop.” The mechanism is the problem: platforms optimize for engagement, slop is cheap to produce at volume, and the algorithm amplifies it whether or not anyone chose to follow its source. It’s a feedback loop that runs on the platform’s incentives, not the user’s preferences.

Worse, it’s getting autonomous. A USC study accepted for The Web Conference 2026 showed that swarms of AI agents can coordinate influence campaigns across X, Reddit, and Facebook with no human direction, manufacturing the look of grassroots consensus. Unlike the bot networks of the last decade, each post is unique and the coordination is implicit, which makes it far harder to catch. The old detection playbook, matching duplicate text, doesn’t apply.

There’s a slower danger underneath: models increasingly train on the slop other models produced, compressing the range of what shows up into a homogenized average. The feed stops reflecting what people actually think and starts reflecting what the models think people think.

The only rules with teeth are European

If feeds shape belief and carry mostly synthetic content, the accountability question is overdue. The answers differ sharply by jurisdiction.

The EU’s Digital Services Act is the one binding regime built for this. Article 27 makes any platform disclose the main parameters of its recommender in plain language. Article 38 requires the largest platforms to offer a feed not based on profiling. Articles 34 and 35 require annual systemic-risk assessments of algorithmic amplification. Article 40, via a July 2025 delegated act, forces platforms to open data to vetted researchers, the single most useful provision, because it lets outsiders test the claims platforms make about their own systems. And it has teeth: in October 2025 the Commission issued preliminary findings that TikTok and Meta were failing the researcher-access obligation, with fines that can reach 6% of global turnover.

The US is going the other way. In Moody v. NetChoice (2024), the Supreme Court held that a platform’s editorial curation is protected speech, while signaling doubt that pure engagement optimization, ranking with no editorial policy behind it, earns the same protection. The Section 230 question, whether a recommendation is the platform’s own speech, is live: the Third Circuit’s Anderson v. TikTok (2024) held that pushing videos to a child’s For You page is first-party speech, so Section 230 doesn’t shield it. That conflicts with other circuits and hasn’t reached the Supreme Court. Meanwhile, the real US action is at the state level and aimed at minors: New York’s SAFE for Kids Act makes a non-personalized feed the default for under-18s, and California’s SB 976 requires parental consent for addictive feeds. Several of these laws are tied up in First Amendment challenges.

The pattern rhymes with AI regulation generally: one comprehensive European framework with extraterritorial reach, and a contested US patchwork moving in the opposite direction. I mapped that split for AI broadly in the regulatory landscape brief.

What this means if you buy attention

For marketers, brand safety used to mean keeping ads away from bad content. It’s now bidirectional. Your ads run next to AI-generated propaganda and impersonation that scale through the same recommender pipes that deliver your campaign, and your brand gets impersonated by synthetic content those pipes distribute. The ANA’s Q3 2025 benchmark reports 99.1% of programmatic spend running in “low-risk” environments, but that low-risk label doesn’t distinguish AI-generated content from human, so the number is measuring the wrong thing. Keyword blocklists, the legacy tool, were quietly deprecated across major platforms in 2024 and 2025. The IAB shipped its first AI transparency framework in January 2026, and it’s voluntary.

What responsible design looks like now

The evidence points away from the reforms that poll well and toward less obvious ones. The Knight-Georgetown Institute’s “Better Feeds” work (May 2025) frames it as optimizing for long-term user value instead of short-term engagement, and the specifics matter more than the slogan:

• Disclose the objective, not just the parameters. What the recommender teams are graded on, time-on-platform versus reported satisfaction, tells you more than a list of ranking signals.
• Make the non-profiling feed easy, or default. The DSA mandates the option; New York makes it the default for minors. Defaults decide behavior.
• Open the system to independent audit. DSA Article 40 is the model. The claims platforms make about their algorithms are untestable without longitudinal, non-aggregated data.
• Label and detect synthetic content in-feed. No binding rule requires this yet, which is the largest open gap in feed governance.

The filter bubble was a comfortable thing to worry about because the fix seemed simple: turn off the algorithm. The actual problems, feeds that move specific beliefs, synthetic content that outnumbers the real kind, and an accountability regime that stops at the EU border, don’t have a settings-toggle fix. They need someone to decide what the feed is for. Right now, in most of the world, that decision belongs entirely to the company selling the ads against it.