Portrait of a Graduate Artifact Submission: How AWS Moderation Works

Overview

This article explains how Wayfinder uses two AWS services—Amazon Comprehend and Amazon Rekognition—to automatically screen POG artifact submissions before they are saved.

Students can view the related Community Agreements by clicking the link within the Artifact set up screen. A link to the Community Agreements will also be included in any warning pop-ups, should an artifact be flagged.  

The article covers:

  • Text moderation — how Amazon Comprehend scans title, description, and reflection fields for toxic or harmful language, including the confidence thresholds Wayfinder applies and known edge cases
  • Image moderation — how Amazon Rekognition evaluates uploaded image attachments against a hierarchical taxonomy of inappropriate content categories
  • What's not covered — non-English text, PDF and ZIP attachments, and content types outside the scope of these services

Text Moderation — Amazon Comprehend

What it is Amazon Comprehend is a natural language processing (NLP) service that provides pre-trained APIs to derive insights from text. Its Toxicity Detection feature identifies toxic content by classifying text across seven categories: sexual harassment, hate speech, threat, abuse, profanity, insult, and graphic content. Wayfinder uses it to automatically scan text fields on artifact records before they're saved.

How it works When a student or educator submits an artifact record, the title, description, and reflection fields are sent to Comprehend before the record is saved. The API assigns an overall toxicity score from 0 to 1 (likelihood of being toxic) and also categorizes the text into individual labels with a confidence score for each. If any field is flagged above the confidence threshold, the save is blocked and the user receives an error message identifying which field triggered it.

Taxonomy — what Comprehend detects Comprehend returns a flat list of seven categories with individual confidence scores for each.

Category What it covers
HATE_SPEECH Speech that insults or dehumanizes a person based on race, ethnicity, gender, religion, sexual orientation, or other identity
GRAPHIC Visually descriptive, unpleasantly vivid language intended to amplify harm or discomfort
HARASSMENT_OR_ABUSE Language that seeks to affect psychological well-being or objectifies a person
PROFANITY Obscene or offensive language
INSULT Demeaning or belittling language directed at a person
SEXUAL Sexually explicit or suggestive language
VIOLENCE_OR_THREAT Language describing or threatening physical harm

Scoring — two layers Each response includes both an array of individual label scores and an overall toxicity score for the full text. All values range from 0 to 1, where 1 is the highest confidence. Wayfinder checks both—a field can be blocked either by an individual label score or by the overall toxicity score exceeding the threshold.

What level Wayfinder operates at Comprehend doesn't use a multi-level taxonomy like Rekognition; it operates on a single flat level. Wayfinder blocks on all seven categories equally, meaning any category that exceeds the confidence threshold will reject the submission. The flagged field name and category are included in the error message returned to the user.

Confidence threshold Wayfinder text moderation is set to 0.75 out of 1.0, consistent with the image moderation threshold of 75%. Labels are returned for all categories regardless of score—it's up to the application to decide which scores to act on. Setting the threshold at 0.75 means Comprehend must be at least 75% confident before a field is blocked, which reduces false positives on legitimate student writing.

Important note: there are known cases where the SEXUAL category returns unexpectedly high confidence scores for completely neutral phrases. Because of this, Wayfinder applies a higher threshold of 0.90 specifically for the SEXUAL category to avoid blocking innocent student submissions.

What fields are scanned

Field Max length
title 255 characters
description 2,000 characters
reflection 5,000 characters

What it doesn't scan Comprehend currently supports English-language input only. Text submitted in other languages won't be scanned and will pass through without moderation. This is a known limitation of the service.

 

Image Moderation — Amazon Rekognition

What it is Amazon Rekognition is an AWS service that detects inappropriate, unwanted, or offensive content in images. Wayfinder uses it to automatically scan images uploaded as artifact attachments before they're saved.

How it works When a student or educator uploads an image attachment to an artifact record, the image is sent to Rekognition before the record is saved. Rekognition uses a three-level hierarchical taxonomy to label categories of inappropriate content. If the image is flagged, the upload is blocked and the user receives an error message. Clean images save normally.

Taxonomy levels Rekognition organizes detections into three levels of specificity.

Level Description Example
L1 Broad top-level category Explicit Nudity
L2 Specific subcategory Graphic Male Nudity
L3 Fine-grained detail Fully Nude Male

Most platforms focus on L1 and L2 for policy decisions, using L3 only when extra precision is needed.

What level Wayfinder operates at Wayfinder blocks at L1—the broadest, most restrictive level. If Rekognition detects anything under a blocked parent category above the confidence threshold, the upload is rejected regardless of which subcategory triggered it. Blocking at L1 is intentional for a K-12 platform—it's safer than trying to allow some subcategories while blocking others.

What Wayfinder blocks

Category What it covers
Explicit Nudity Nudity, graphic sexual content
Suggestive Revealing clothing, suggestive poses
Violence Graphic violence, weapons
Visually Disturbing Corpses, emaciated bodies
Hate Symbols Nazi imagery, KKK symbols
Drugs & Tobacco Drug use, paraphernalia
Gambling Casino imagery
Rude Gestures Obscene hand gestures

Confidence threshold Wayfinder is set to 75 out of 100. Rekognition must be at least 75% confident before an image is blocked. AWS's own default is 50—our setting of 75 is stricter, reducing false positives while still providing strong protection appropriate for a K-12 environment.

What it doesn't scan Rekognition only scans image files (.jpg, .png, .gif, .webp). PDF and ZIP attachments aren't scanned. Rekognition doesn't claim to be an exhaustive filter of all inappropriate content and doesn't detect illegal content such as CSAM.