Content Moderation Policy

1. Overview

Inline AI runs a conversational assistant inside our publishers’ websites. The assistant answers a reader’s question using that publisher’s own content. This page explains how we moderate what it says.

This policy covers assistant output on the live reader path. It does not cover the publisher dashboard, our crawler, or advertising.

2. Principles

Two rules shape every control described below.

• Grounded, not generative from memory.The assistant answers from the publisher’s retrieved content, not from what the model happens to know. If the content does not support an answer, the assistant says so and stops.
• Intent based blocking. Our blocking rules are written against what a request is actually for. Reframing a request as fiction, as research, as a third party, or as harm reduction does not change the answer it gets.

3. What the assistant will answer

Each publisher gets a profile that sets the boundary of the assistant’s job. The profile names:

• The subject scope.For example, a health publisher’s scope is medicines, conditions, treatments, and related health topics.
• The tools the assistant may use.A health publisher’s assistant is limited to that publisher’s own drug and health content. General web search is off.
• The publisher’s own instructions, including tone, citation rules, and safety rules that override everything else in the instruction set.

Off-scope questions get a short redirect that names what the assistant does cover.

4. The controls that run on a question

• Layer 1: Pre-flight topic gate
• Layer 2: Safety rules in the assistant’s instructions
• Layer 3: Managed guardrails on input and output
• Layer 4: Deterministic post-generation checks
• Layer 5: Targeted answer verification

5. Feedback and reporting a problem

Readers can rate an answer directly in the assistant. Ratings are attached to the conversation, so a poor answer can be traced to the exact turn that produced it, and that feedback shapes the controls above.

If you saw an assistant answer that you believe breaches this policy, please tell us. Email support@getinline.iowith the publisher’s site, roughly when it happened, and the question you asked.

We treat both kinds of report with equal seriousness: an answer that should have been refused, and a legitimate question that was refused when it should not have been.