Nobody Told Me The Algorithm Had Buried My Posts

In August 2026 Bloomberg published a confidential TikTok report showing that the company built a safer version of its recommender in 2021 and then withheld it from part of the user base. Roughly 10% of American users, about 15 million people at the time, stayed on the older build so that managers could measure what the safety change cost in daily active users. Chase Nasca, 16, was among them. He died in 2022 after months of suicide and self-harm videos in his feed.

Earlier reporting had far less behind it. The Tech Edvocate described a leak dated May 5, 2026, putting outrage and fear at 3 times the weight of factual correctness in the ranking, and a protest hashtag at 2.3 billion views in a day. Neither figure has surfaced outside that outlet and the sites reprinting it, while the underlying documents remain unpublished. Bloomberg drew on a dated internal file and stated the size of the control group.

Sensor Tower measured the response and gave the numbers to CNBC. During the 5 days after control of the American business changed hands on January 22, 2026, uninstalls of TikTok in the US climbed by about 150% against almost no change in daily active use. Downloads of UpScrolled rose more than tenfold that week and Skylight Social by 919%, a burst of curiosity with almost no migration at all.

Once the major platforms replaced chronological feeds with ranked ones, publication and audience came apart. Officials in Apia and Suva can put a cyclone warning into Gagana Sāmoa and na vosa vaka-Viti, then watch the recommender rank it against a dance clip on the identical scale. Although the warning exists, the families needing it never see one. The officials end up with an impressions counter and no explanation for it.

The Economist released 90 minutes with Elon Musk on 23 July 2026, recorded three days earlier at the Texas Gigafactory in Austin, TX. Zanny Minton Beddoes spent much of it on his political influence in Europe. Musk has argued since 2022 for throttling distribution instead of deleting posts, a doctrine of free speech without free amplification. Publishers in Tongan or Marshallese will find the difference academic, because throttling a post to zero and deleting it just means six of one and half a dozen of the other, right? Throttling generates no takedown notice for a transparency report to count.

On December 31, 2021 the Cyberspace Administration of China promulgated the Provisions on Algorithmic Recommendation together with 3 other ministries. Since March 1, 2022 providers must publish their ranking rules and file services with public-opinion attributes. Weibo duly filed Hot Search, its trending list. The filed description names 3 popularity inputs and multiplies the total by an interaction-rate coefficient.

Weibo added a public index to Hot Search in January 2025, after an official campaign against algorithm problems across the sector. In fact, Beijing had already suspended the list twice. The 2018 suspension lasted a full week, following an accusation in state media of failure to censor material against Party values. Publication of a formula coexists comfortably with editorial control. Pacific governments negotiating Chinese loans face the mirror image, since the terms come drafted in Mandarin and the choice between Mandarin varieties in island negotiations changes whether a borrower grasps a condition.

Recommenders score video partly on transcript signals. Tongan and Fijian clips shipped without a caption track may become part of the classification pool but with no transcript to score, and the ranking treats them accordingly. Buyers wanting graded evidence on machine output across 30 Pacific pairs can consult the AI comparison tables that we published. Dubbing choices for Pacific audiences answer to the constraint too. Sensor Tower publishes weekly download series for apps in dozens of markets. No comparable series exists for caption coverage in the Pacific.

For writers with an audience, the answer has been to buy it outright. Substack took $100 million from Bond and The Chernin Group in July 2025 at a valuation of $1.1 billion, on more than 5 million paid subscriptions, up from 4 million the previous November. Pacific publishers gain the plainest form of the deal, since a newsletter in Vagahau Niuē or Reo Māori Kūki ʻĀirani goes out whole to its subscribers.

A select committee in Wellington reported in December 2025 that most of its members judged the risk of algorithmic harm greater than the advantage. The committee recommended 3 things: A national online safety regulator, transparency duties on ranking systems and an account restriction for children under 16. Submissions counted roughly a dozen public bodies sharing online oversight, a split the committee proposed to end. Ministers had 60 days to respond, until June 3, 2026. The Prime Minister has promised age legislation by end of year. Worth noting that publication in te reo Māori appears in no recommendation.

Australia closed listed services to children under 16 on December 10, 2025 and has been testing the proposition ever since. Platforms deleted or froze about 4.7 million accounts by mid-December and blocked a further 300,000 by March 2026. By the 3-month mark the eSafety Commissioner had found 81% of the children concerned still using the barred apps. A further rule made in March 2026 goes at recommender systems and continuous feeds directly. The ceiling on penalties is AUD 99 million, a small number against the revenue of the companies involved.

After Fiji's Parliament adopted the Online Safety Act on May 16, 2018 by 27 votes to 14, the country had a commission and a fresh offence under section 24. Dialogue Fiji asked for reform of that section in September 2025, once investigations and charges came in quick succession. By February 2026 the courts in Suva were handing down deterrent sentences under the Act. Island media institutions therefore live under a developed body of law about the person posting and hardly a line about the system ranking the post.

Meanwhile, Brussels has gone at the design itself. The European Commission issued preliminary findings on addictive design at 3 services: Facebook and Instagram from Meta together with TikTok, naming infinite scroll and the personalised recommender among the features at issue. Meta and TikTok have appealed. France voted an under-15 restriction into law on July 21, 2026, then stripped out the age-verification duty and left enforcement to CNIL under the Digital Services Act. In February 2026 the Council of Europe Commissioner for Human Rights told governments to prefer binding duties on platform design to bans on children.

In Auckland, reviewers on a diaspora service handled a miniature of the problem: Consent under privacy law must be specific and freely given, which makes an over-broad consent line defective on its face. They narrowed such a line back to the operator's declared purpose, then checked the data-subject rights summary phrase by phrase across Samoan and Tongan. Of course the name of a support service stayed in English, since the string must match a user's search, but all the rest was not in English. Variable merge fields kept their source form, two rounds closed the file, a legal-correctness check and a plain-wording edit.

Consent scope and rights wording stay auditable without access to the model. Count the versions of the transparency report, then name the reviewer of the rights summary. Wellington and Brussels will finish drafting their digital platform obligations during 2026. So far the drafts name the systems needing disclosure while staying silent on the tongue of the disclosure itself. Regulators will get to that question eventually.