Domain 3 of 5 · Chapter 3 of 3

Responsible AI for images, video, and generated media

Unlock the complete study guide + 1,040 practice questions across 16 full exams.

Bundled into the existing AI-103 premium course — no separate purchase.

14-day money-back guarantee — no questions asked.

Included in this chapter:

  • Which control sees pixels, and which sees only text
  • Scoring an image with Content Safety
  • Policy rules the harm taxonomy cannot express
  • Configuring the filter that guards generation
  • What a blocked image request looks like in code
  • Injection hidden in an uploaded image
  • Provenance on generated media
  • Exam-pattern recognition

Which control can express which visual rule

Control propertyAnalyze Image (harm categories)Custom categories (rapid)Custom categories (standard)Blocklists
What it inspectsImage pixelsText or image samplesText onlyText only
What it can expressFour fixed harm categoriesAny incident you describe by exampleAny category you train from samplesThe exact terms you list
What it returnsA severity per categoryA match against the incidentA match against the trained categoryA term hit, never a severity
Setup before first useNone, the categories are built inCreate an incident and upload samplesDefine, sample, and train, which can take hoursCreate the list and attach it to a filter
Fits a prohibited-symbol ruleNo, symbols are not a harm categoryYes, image samples define the incidentNo, the trained model reads text onlyNo, it never reads pixels

Decision tree

Is the requirement abouta picture at all?Is it harm in the fourbuilt-in categories?Is it a rule you defineby your own examples?Is it instructions hiddeninside the image?NoYesYesNoYesNoYesNoText-side controlsblocklists, groundedness, protected material, PII filterAnalyze Image, then your own thresholdsfour severities on the 0, 2, 4, 6 scaleCustom categories (rapid)an incident plus image samples, no training stepPrompt Shields, documents analysisafter OCR turns the picture into textContent Credentialsa signed manifest already attached to the file
Cheat sheet

Cheat sheet

  • Analyze Image returns per-category severities, not an allow or block decision
  • Image moderation supports only the four-level severity scale
  • An image is submitted as base64 content or as a blob URL, never both, and blob access needs a data role
  • Omitting the categories array scores all four harm categories in one call
  • Image moderation input limits and animated-format behaviour
  • Text hidden inside an uploaded image is a Document attack, not a User Prompt attack
  • Prompt Shields analyzes text and returns a boolean, so image payloads must be transcribed first
  • The Indirect Attacks filter only works when ingested content is marked with document delimiters
  • Shield and protected-material filters can annotate without blocking
  • Every Azure OpenAI generated image carries C2PA Content Credentials with no setup
  • The manifest attests that the image is AI generated and by which model family, not who prompted it
  • Content Credentials ride in the file rather than being burned into the pixels
  • Credentials are verified with the C2PA verify page or CAI open-source tooling
  • Content Credentials are documented for the image generation models specifically
  • Custom categories (rapid) is the image-capable customization path and needs no training
  • Custom categories (standard) trains a model but works on text content only
  • Blocklists do exact term matching against text and never touch pixels
  • The four harm categories say nothing about brand misuse or off-policy imagery
  • A filtered image prompt fails the call outright with a contentFilter error
  • The agent image tool exposes a moderation level that sits alongside the deployment's filter
  • Groundedness, protected material, and PII filters are text-side and cannot see an image
  • Thresholds are set separately for prompts and completions, and Safe content is never filtered
  • If the filtering system is unavailable the request completes unfiltered and only an annotation says so

Unlock with Premium — includes all practice exams and the complete study guide.

References

  1. Harm categories in Azure AI Content Safety
  2. Quickstart: Analyze image content
  3. Custom categories in Azure AI Content Safety
  4. Use blocklists for text moderation
  5. How to use block lists in Microsoft Foundry models
  6. Content filtering for Microsoft Foundry Models
  7. Use the image generation tool in Foundry Agent Service
  8. How to use image generation models from OpenAI in Microsoft Foundry
  9. Prompt Shields in Microsoft Foundry
  10. Quickstart: Detect prompt attacks with Prompt Shields
  11. Document embedding in prompts for Azure OpenAI
  12. Content Credentials in Azure OpenAI
  13. Content credentials in Azure text to speech avatar