How an open call for evidence and one email address produced 1,050 documented cases of content suppression

What happened

Between October and November 2023, Human Rights Watch ran an open call for evidence in English, Arabic and Hebrew from its own accounts on Instagram, X and TikTok, asking people whose posts had been removed or hidden to email in the details. It reviewed 1,285 submissions, published 1,050, and coded them into six recurring patterns, each appearing at least a hundred times. In more than 300 cases the appeal mechanism itself malfunctioned, so the person could not contest anything.

Its conclusion on automation:

Meta’s reliance on automation for content moderation is a significant factor in the erroneous enforcement of its policies, which has resulted in the removal of non-violative content in support of Palestine on Instagram and Facebook.

Separately, Meta’s own Oversight Board reviewed how the company treated the Arabic word shaheed (شهيد), which in ordinary Arabic covers memorialisation, journalism, religious usage and the reporting of violence. Meta treated it as praise for designated individuals regardless of context. In March 2024 the Board ruled that this caused widespread and unnecessary censorship of Arabic speakers and Muslim communities, and told Meta to remove such posts only where there were accompanying signals of violence. In September 2022 BSR, commissioned by Meta itself, had concluded that Meta’s 2021 conduct had an adverse human rights impact on Palestinians’ freedom of expression.

What the automation part shows, and what it does not

A classifier trained mostly on English, applied to Arabic without equivalent investment, over-removes legitimate Arabic speech. When the appeal path is automated too, the error has no correction route. That is what the 300-plus broken appeals mean.

Automation is one of four causes Human Rights Watch names, not the cause, and nobody outside Meta knows which classifiers run on this content.

The method a five-person office can copy

One email address and a spreadsheet.

1. Pick one narrow harm you can describe in a single sentence, so submissions are comparable. “Your post was removed or hidden by Instagram or Facebook” is the right size.
2. Publish an open call from your own accounts, in every language your community actually uses. Human Rights Watch posted in English, Arabic and Hebrew. For Baghdad that is Arabic, Kurdish Sorani and English.
3. Set up one dedicated intake email address. No form builder, no database, no engineer.
4. Publish the required fields up front. Copy this list almost verbatim, it is the highest-value part of the method: a screenshot of the original content; the platform; the date and country of posting; the form of suppression experienced; the platform’s notification, if any; prior engagement figures, if you suspect the post was quietly demoted; the account address; and the status of any appeal.
5. Set a close date and say so publicly.
6. Verify before you count. Human Rights Watch judged each claim against the screenshots, the background in the email, and publicly available information. Discard what you cannot substantiate, and publish how many you discarded. Reviewing 1,285 and publishing 1,050 is what makes the 1,050 credible.
7. Code the verified cases into a few named recurring patterns and report how often each appeared. Six patterns, each recurring at least a hundred times, is what turned an inbox into a finding.
8. State your sample’s limits in the report itself, before a critic does. Human Rights Watch wrote that the distribution of cases does not necessarily reflect the overall distribution of censorship. That sentence costs you nothing and removes the easiest attack on your work.
9. Write to the company before publishing, and give a deadline. HRW wrote to Meta on 15 November 2023 and published on 21 December.
10. Convert the pattern counts into specific, checkable demands. Vague demands are ignorable. HRW asked for disclosure of which classifiers are used and their error rates, and for a human being in the loop with meaningful oversight of decisions made by automated systems.

The version built to last

Human Rights Watch ran a one-off campaign. 7amleh’s Palestinian Observatory for Digital Rights Violations is a standing intake database, which is how it accumulated 3,520 cases across 2021 to 2025 and tracked 2,800 appeals. MENA-run, permanent, small: the realistic model for Baghdad. Run the one-off campaign first to learn the fields and the volume. Build the standing intake only if the submissions keep arriving.