Real-time Multimodal Misinformation Shield
In light of growing concerns over misinformation on social media platforms and the potential for malicious deepfakes, this challenge focuses on developing a cutting-edge, real-time AI assistant for content moderation and misinformation detection. The system will monitor a simulated social media feed, identify multimodal content (text, image, video frames) that violates platform policies or spreads misinformation, and provide immediate alerts or suggested counter-narratives. This challenge emphasizes the integration of advanced multimodal LLMs like Claude Opus 4.5, a voice interface for rapid human intervention, and robust memory management for policy enforcement. The assistant must operate in near real-time, making accurate assessments of complex, often ambiguous, content and providing explainable justifications for its decisions.
What you are building
The core problem, expected build, and operating context for this challenge.
Build a real-time multimodal misinformation detection assistant using Claude Opus 4.5, Hamming for voice interaction, and Semantic Kernel for orchestration.
How work is evaluated
The assistant will be evaluated on its accuracy in identifying policy violations and misinformation across multimodal inputs, the relevance and speed of its alerts, and the quality of its generative responses (e.g., counter-narratives, explanations).
Challenge datasets
Public datasets, plus the private datasets you have access to.
What you should walk away with
Master multimodal input processing with Claude Opus 4.5, analyzing text, images, and simulated video frames for contextual understanding and policy compliance.
Integrate Hamming for a natural voice interface, allowing moderators to receive audio alerts and issue voice commands for real-time inquiry or action.
Design and orchestrate skills within Semantic Kernel to handle various moderation tasks: content analysis, policy lookup, alert generation, and counter-narrative drafting.
Implement a long-term memory system using Milvus to store and retrieve comprehensive policy documents, past moderation decisions, and detected misinformation campaigns, enhancing contextual understanding.
Utilize serverless functions (e.g., AWS Lambda, Azure Functions) to process incoming simulated social media streams in near real-time, feeding data to the Semantic Kernel agent.
Develop explainability features for the assistant's decisions, outlining why certain content was flagged and referencing specific policy clauses or evidence.
Reference links and supporting material
Sample data for 2 tasks
Challenge setup
The assistant will be evaluated on its accuracy in identifying policy violations and misinformation across multimodal inputs, the relevance and speed of its alerts, and the quality of its generative responses (e.g., c...
Challenge input
{ 'post_text': '...', 'image_url': '...', 'video_frame_url': '...' }
Agent execution
No agent tool is listed for this challenge.
Evaluated output
{ 'flagged': true/false, 'reason': '...', 'policy_violation_type': '...', 'confidence': 0.0-1.0 }
- Correctly flag at least 90% of known misinformation/policy violations in a test set.
- Generated counter-narrative must be rated as neutral and informative by a human expert or sentiment analysi...
- Confidence Score Average target: 0.95
- Python execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- Correctly flag at least 90% of known misinformation/policy violations in a test set.
- Generated counter-narrative must be rated as neutral and informative by a human expert or sentiment analysis tool (e.g., 80% positive/neutral sentiment).
Metrics
- Confidence Score Average · Target 0.95
- Response Generation Latency · Target 500
Workflow tasks
- Task 1
Input: { 'post_text': '...', 'image_url': '...', 'video_frame_url': '...' }
Output: { 'flagged': true/false, 'reason': '...', 'policy_violation_type': '...', 'confidence': 0.0-1.0 }
Execution harness: Python
- Task 2
Input: { 'flagged_post': { ... }, 'policy_violation_type': '...' }
Output: { 'explanation': '...', 'counter_narrative_suggestion': '...' }
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.