AI Transparency Notice

• Reads the original post and the pasted discussion as a whole.
• Classifies each comment (for example supportive, genuine question, civil disagreement, provocation, abuse).
• Estimates good-faith likelihood, relevance, reputational risk and escalation risk.
• Suggests an action such as thank, reply, clarify, ignore, hide, report, block or escalate.
• Notes where a point appears to have been answered already in the thread.

• It can be wrong, and can misunderstand context, humour, culture, dialect or sarcasm.
• It does not publish, hide, report or block anything automatically.
• It does not make legal findings, and it does not determine guilt or criminality.
• It does not identify ordinary commenters.

Every recommendation must be reviewed by a person. Mandatory human escalation applies to threats, safeguarding concerns, criminal allegations, doxxing and live legal matters — for those items no suggested reply is produced and approval controls are disabled.

Users record whether they approve, edit and approve, ignore or escalate a recommendation. Both the original suggested reply and any edited version are preserved.

The model name and prompt version used for an analysis may be stored so that past advice can be understood later. Analyses and decisions are append-only records.

AI advises. Humans decide.