CyberWell asks AI moderators to read context and preserve human review

Recommendations, not an adopted platform rule

CyberWell, a nonprofit focused on online antisemitism, publicized five recommendations on October 21 for platforms using large language models to moderate content. The proposals concern expert input, assessment across text and other media, language-specific testing, escalation to human reviewers and explanations of enforcement decisions. They form part of a submission to the Oversight Board, rather than an announced change to Meta’s rules. Source: cyberwell.org

The organization’s report index dates the underlying submission October 15. October 21 is therefore the date of the public release being reported here, not a verified first-publication date for all its recommendations. Source: cyberwell.org

The submission argues that moderation must consider context beyond isolated words: a comment’s relationship to its parent post, for example, or information conveyed through an image or audio. It also recommends testing dialects, irony and coded language, and giving users explanations that identify the content element and policy behind a removal. These are proposed design and accountability measures; the submission does not provide a comparative trial establishing how much a particular language model improves moderation. Source: cyberwell.org

Two kinds of error to examine

The Oversight Board’s October 1 research announcement supplies a wider frame. It identifies potential benefits including fewer wrongful removals and better explanations, alongside risks from bias, fabricated outputs and missed context. The announcement requests evidence about human escalation, languages with fewer resources, testing, appeals and audits. That research agenda is distinct from CyberWell’s contribution to the work. Source: oversightboard.com

Read together, the documents suggest a concrete question for religious-community media teams, including church communicators: can a moderation system account for both harmful content left available and legitimate expression removed? A single removal total cannot answer both questions. Nor does a more fluent explanation by itself show that the underlying decision is correct.

Our practical suggestion is to preserve the context needed for an appeal: the original post, the relevant comment or media element, the stated policy and the platform’s reason. Keep this record privately where appropriate rather than reposting abusive material to demonstrate the problem. For a hypothetical educational post discussing hateful rhetoric, a useful review would ask whether the system distinguished reporting that rhetoric from endorsing it. This is an editorial example, not a tested outcome from the submission.

Christian application: hear before judging

For Christian communicators, Proverbs 18:13 offers a reason to slow down before judging. The proverb rebukes answering before listening; it concerns human conduct, not a test of model performance. Our application is to hear the surrounding context and allow correction, whether reporting abuse or challenging the removal of legitimate expression. The recommendations raise questions about that process; they do not prove it works. Proverbs 18:13 (NIV)