Moderation Transparency

How automated moderation works

Every submitted review and every broker response is screened by OpenAI’s Moderation API. The model classifies the text against several content categories — hate, harassment, violence, sexual content, and more — and returns scores we use to route the submission. We also apply our own checks for personally identifiable information (PII), short-length reviews, and an admin-maintained blocklist.

What triggers a hold

A review goes into the human-review queue if any of the following are true:

• Moderation classifier scores above our hold thresholds for hate, harassment, or violence.
• PII is detected — phone numbers, email addresses, SSNs, or specific street addresses.
• The review is unusually short, where the model can’t make a confident judgement.
• Our internal blocklist matches a known abuse term.
• Velocity check trips: three or more reviews submitted on the same broker within 48 hours.
• Three or more independent flags are filed on a published review.

What triggers an auto-rejection

We auto-reject only the hardest cases: graphic violence, sexual content involving minors, hate speech that includes threats, and content the moderation classifier is highly confident violates safety policy. Auto-rejected reviews aren’t published. The reviewer is emailed and can contact us if they believe it was a mistake.

What human moderators do

Every held review is reviewed by a person, usually within 24 hours. Three actions are available: approve, approve with PII redaction, or reject. We log every decision in an internal audit trail so we can review patterns and correct mistakes.

What we will edit

We only edit reviews to remove personally identifiable information: phone numbers, email addresses, SSNs, full street addresses, and other people’s full names. When we do, the redaction is clearly marked with “[redacted]” in the published review. The original text is kept in our private redaction log for audit purposes.

What we won’t edit

We do not edit the substance, opinion, or tone of a review. We do not change ratings. We do not soften a negative review or sharpen a positive one. If a review violates the guidelines beyond what redaction can fix, we reject it instead of rewriting it.

Appeals and disputes

Reviewers can contest a rejection by replying to the rejection email — a human responds within a few business days.

Brokers can file a formal dispute on any review on their profile from their dashboard. Disputes include the broker’s explanation, optional supporting evidence, and a desired outcome (remove, redact a specific claim, or request the reviewer update). A human admin reviews every dispute and resolves it explicitly.

Why we exist on this

Our goal is real reviews from real people. Transparency about moderation is how we earn that trust. If we made a decision you think was wrong, tell us.