Who decides what's divisive?

Photo credit: Drew of The Come Up Show (Flickr), CC BY-SA 2.0 https://creativecommons.org/licenses/by-sa/2.0, via Wikimedia Commons

On September 4 and 5, Macklemore opened for Ed Sheeran at MetLife Stadium and said "Free Palestine" on stage. On September 14, the promoter announced he was off the rest of Sheeran's US dates. [1]

The promoter, Messina Touring Group, said venues on the upcoming dates "will not allow a concert to take place with Macklemore on the lineup." [2] Sheeran posted: "Macklemore coming off tour was the promoter's decision, it was not mine." [3] Robert Kraft, whose Kraft Group owns Gillette Stadium, cited Macklemore's comments at the New Jersey shows and "a broader history of antisemitic rhetoric and imagery." [2] Macklemore says Sheeran told him Kraft rallied other stadium owners into an ultimatum. [4] Sheeran has confirmed that venues threatened to pull the shows and that Kraft was one of the people he spoke with, but not that Kraft organized the others. [3] Four acts then quit the tour in protest: Finneas, Lukas Graham, Aaron Rowe and Beoga. [5]

The mechanism isn't in dispute. Whoever owns the room decides who gets heard in it.

That's an old story in concert venues. It's a newer one online, because online you often can't see the room, the owner, or the decision.

And it's one case. Censorship is bigger than one tour, and algorithmic bias shows up in both kinds described below.

Algorithmic bias is when an automated system produces unfair outcomes for some groups of people, whether or not anyone meant it to. The system might rank a feed, filter posts for moderation, or sort job candidates.

It usually gets in one of three ways. The data it learned from already carries the unevenness of the world that produced it. The goal it was built to hit, like engagement or "quality," favors some voices over others. Or it picks up on something that stands in for a trait it was never given, so topic or network size ends up tracking gender or politics.

When that system decides what people get to see, bias turns into suppression. Sometimes a moderation filter over-flags posts in one language. [15] Sometimes a ranking shows some political posts to more people than others. [14] Those are two different kinds of suppression, and it helps to tell them apart.

Active suppression is the kind you can see. A post comes down. An account gets restricted. You get a notice, even if it's a bad one.

Passive suppression is quieter. Your post stays up. Fewer people get shown it. Often nobody tells you, so you can't appeal something you were never told happened.

Most arguments about "shadowbanning" collapse these two together, and then collapse again into "the algorithm hates me." So here's what's actually on the record, graded by how solid it is.

The strongest evidence for passive suppression isn't leaks or conspiracy theories. It's the platforms describing it themselves.

LinkedIn told EU regulators in its 2024 risk assessment under the Digital Services Act that "dismissive or divisive content" in the feed "will not be eligible for broad distribution," that "the LinkedIn platform favors knowledge sharing rather than virality," and that it tends not to spread content that doesn't meet "a minimum quality bar." [6] After counting its own safeguards, it rated the remaining risk to freedom of expression as low. [6] In 2018, LinkedIn told BuzzFeed News it "does not limit the distribution of posts" and removes content that violates its policies. [7]

YouTube said in 2019 that after more than 30 changes, US watch time from recommending borderline videos and harmful misinformation to non-subscribers fell 70% on average. [8] YouTube defines borderline content as content that "brushes up against our policies, but doesn't quite cross the line." [8] YouTube decides what counts as borderline.

TikTok publishes standards for what can appear in the For You feed, and tells creators when a video "isn't eligible for the For You feed" with a warning bar on the post's Performance tab. [9] [26]

Instagram and Threads began limiting recommended political content by default in 2024, covering posts "likely to mention governments, elections, or social topics that affect a group of people and/or society at large." [10] In January 2025 Meta said it would phase political content back in. [11] Instagram also lets professional accounts see whether their posts are blocked from recommendations, which is Meta confirming that being blocked from recommendations is a real state. [12]

X calls its version "Freedom of Speech, not Freedom of Reach." [13] Posts it judges to violate its rules can stay up while being excluded from search results, trends and recommended notifications and removed from the For You and Following timelines, with a label both the author and other users can see. [13] [31]

Ranking isn't neutral even when nobody is trying to suppress anything. Twitter's own researchers used a control group of nearly 2 million daily active accounts and found its algorithm amplified the mainstream political right more than the left in six of seven countries. Germany was the exception. [14]

Meta commissioned its own review of how it handled Israel and Palestine content in May 2021. The report, by BSR, a business and human rights consultancy, found Meta over-enforced against Arabic content, and that this harmed Palestinian users' freedom of expression. BSR said it found no intentional bias, only unintentional bias, and found Hebrew content had greater under-enforcement, largely because Meta lacked a Hebrew classifier and had lost Hebrew-speaking staff. [15]

Human Rights Watch documented 1,050 cases of removed or suppressed content on Instagram and Facebook in October and November 2023, and 1,049 of them were peaceful pro-Palestine content. [16] HRW publicly asked for cases, so these aren't a random sample. HRW itself says they don't necessarily reflect the overall distribution of censorship.

The Markup, a nonprofit tech newsroom, ran 70 test accounts on Instagram in February 2024 and found nongraphic photos of the war were 8.5 times more likely than other posts to be hidden from a hashtag it created for the test. In a separate test, some comments stayed visible to the person who wrote them and to nobody else, with no notification. Meta said "any implication that we deliberately and systemically suppress a particular voice is false." [17]

On LinkedIn, 7amleh, a Palestinian digital rights group, published a report in October 2025 based on 15 user testimonies and interviews with LinkedIn and Microsoft employees. It describes pro-Palestinian posts removed and accounts restricted while "hate speech against Palestinians remained largely unchecked." [18] It's testimony, not an audit.

Leaked TikTok moderation documents, reported in 2020, told moderators to suppress posts from users deemed too "ugly," poor or disabled. TikTok said the rules were "an early blunt attempt at preventing bullying" and were no longer in place. [19]

In late 2025, women on LinkedIn started testing whether the feed treats them differently. In one test, Cindy Gallop's post reached 0.6% of her followers, while a man posting the same content got impressions equal to 143% of his follower count. [30] Another woman switched her profile to male and said her impressions jumped 238% within a day. [20] These are individual experiments, not a study.

LinkedIn says its systems don't use gender as a signal. [21] That can be true and still leave the outcome gendered, because network size and topic can stand in for gender without anyone designing it that way. LinkedIn hasn't ruled that out.

There's precedent for LinkedIn measuring this in its own systems. In 2019 its researchers published a method that measures gender skew in search results from LinkedIn Recruiter, its paid hiring search tool, and re-ranks them to match the qualified candidate pool, and LinkedIn rolled it out to all Recruiter users. [22]

You'll see a number going around that women get 8x less reach than men on LinkedIn. I couldn't find a study behind it. The experiments above are real. That number isn't.

Nobody has shown platforms deliberately target causes. The admissions above are about "divisive," "borderline," "political" and "low quality" content. Posts that take a stand tend to land in those buckets. That's a design choice with real consequences, but it isn't the same as proof of intent.

You can't reliably tell suppression from a slow month. Researcher Kelley Cotter calls it "black box gaslighting": platforms lean on their authority over their own algorithms to make users doubt what they know about them. [23] There's no validated test from the outside. Watch out for anyone selling a "shadowban checker."

The feed follows what you linger on. If your For You page turns strange after you engage with something heavy, the documented explanation is how fast these systems adapt: a Wall Street Journal test found TikTok could work out a new account's interests in under two hours. [24] The buttons meant to steer them barely work. In a Mozilla study with more than 22,000 YouTube volunteers, "Not interested" prevented only 11% of unwanted recommendations. [25] I found no evidence that TikTok punishes users for engaging with controversial topics.

The companies that decide what's divisive in your feed are, in many cases, the same companies building the AI tools teams now run their work on. They write the usage policies. They decide which accounts stay open.

We talk about AI democratizing access. Access that a company can narrow, label or revoke based on its own read of what's acceptable isn't democratic. It's rented. That's my opinion, not a finding, and it's why I think mission-driven teams should plan for it the way they'd plan for any single point of failure.

A few people worth following, each through their own lens:

Macklemore knew the people who own the room could shut him out. He said it anyway. Four acts walked with him. That's the part worth copying.

Online, the room is harder to see. So look where the platforms have admitted the dials are. Keep your own records. Build at least one room nobody else can turn down. And when someone in tech takes a stand and it costs them reach, share their work, because the feed won't do it for you.

Today's card in my daily tarot series is the Eight of Swords: bound and blindfolded, unable to see what's holding you. The ropes are loose in that card. She could walk out. Most of us can too, once we see what's tying us down.