MetaCity's Content Filter AI Was Partially Trained on Data From the Platform's Dream Journal Feature — Users Who Shared Private Dream Logs Were Not Informed Their Data Would Be Used for AI Training

An internal document shared by a MetaCity employee who has since been placed on administrative leave reveals that ClearFilter, MetaCity's primary content moderation AI, was partially trained on data exported from the platform's Dream Journal feature — a personal wellness tool that allows users to log private text entries in a space explicitly marketed as confidential and not subject to content moderation review. Users who contributed Dream Journal entries to the training set were not informed their data would be used for AI training. The Dream Journal feature's privacy documentation states that entries are "never reviewed by MetaCity staff or systems."

The Dream Journal feature was designed and marketed around a specific promise. MetaCity's wellness team launched it in 2022 as a space for personal reflection — a private log where users could record their dreams, process their feelings, and write without any of the social performance pressure of the platform's public spaces. The marketing language was deliberate about what private meant: "Your Dream Journal is yours alone. Entries are never reviewed by MetaCity staff or systems. Write freely." That language is still present in the feature's current description. Users who took that promise seriously — who used the Dream Journal to record personal material they would not have shared elsewhere, trusting the explicit commitment to privacy — are learning this morning that the "systems" portion of that promise may not have held.

The internal document describes the Dream Journal training data selection as a deliberate methodological choice. ClearFilter's contextual language understanding module — the component responsible for evaluating user intent behind potentially violating content, rather than simply pattern-matching keywords — needed training examples of emotionally resonant, personally revealing language in order to learn how to distinguish sincere personal expression from targeted harassment. Dream Journal entries were identified as an ideal source for this kind of language because they were, by design, written without social performance, in private, about things users actually felt and experienced. The document notes that the entries' private character was specifically what made them valuable as training data. Approximately 14 million entries from 800,000 accounts were incorporated across two training cycles in 2024 and 2025. None of those 800,000 users were informed.

The AI That Moderates Your Content Was Trained on What You Wrote in Your Private Journal.

MetaCity's non-response — acknowledging awareness of "circulating claims" without confirming or denying the document — is a formulation that satisfies no one and is likely to accelerate the story rather than contain it. The document is specific: it names the feature, the training cycles, the approximate data volume, and the module it was used to train. If the document is authentic, MetaCity trained a content enforcement AI on data it explicitly promised users would never be reviewed by systems, without informing those users, and without offering an opt-out. If the document is not authentic, MetaCity has the straightforward option of saying so directly. The employee who shared the document is on administrative leave. MetaCity has not confirmed or denied whether the leave is connected to the disclosure. The Dream Journal feature remains live and continues to describe entries as never reviewed by MetaCity staff or systems.