-
-
Configuring policy and guidelines for the community
-
Stats page explaining how many actions are taken by the moderator
-
Shows users previous history
-
Shows the previous posts for which any action was initiated
-
Shows the recommended action for the post based on policy and guidelines
-
Detailed explanation of why the action is recommended
ModPilot — The Investigation Layer Between Reports and Moderator Actions
Inspiration
While exploring Reddit moderation workflows, we noticed that moderators rarely struggle with taking actions—Reddit already provides buttons for approving, removing, locking, and escalating content. The real challenge lies in everything that happens before that decision.
When a report arrives, moderators often need to:
- Read the reported content
- Review subreddit rules
- Check the author's history
- Examine previous moderator actions
- Understand surrounding context
- Explain their decision to users
This investigative process can take significantly longer than the moderation action itself.
Existing moderation tools are excellent at automation, filtering, banning, and queue management. We wanted to build something different:
Most moderation tools classify content. ModPilot investigates context.
Our goal was to create an AI-powered moderation copilot that helps moderators gather evidence, understand context, and make informed decisions—while ensuring the moderator always remains in control.
What ModPilot Does
ModPilot is a Devvit application that transforms moderation reports into structured investigation case files.
When a moderator is unsure how to handle a post or comment, they can launch an investigation directly from Reddit.
ModPilot then:
- Analyzes the reported content
- Reviews subreddit-specific moderation rules
- Examines relevant user history and prior moderation actions
- Identifies potential rule matches
- Generates an evidence-backed recommendation
- Explains its reasoning with citations
- Calibrates confidence based on available evidence
- Presents everything in a clear investigation report
The moderator can then:
- Approve
- Remove
- Lock
- Escalate
with a single click.
Importantly, ModPilot never performs moderation actions autonomously. Every action requires explicit moderator approval.
Key Features
🔍 Investigation Reports
Instead of a simple AI verdict, ModPilot creates a complete moderation case file containing:
- Current Content Assessment
- Potential Rule Matches
- Moderator Memory
- Evidence & Context
- Confidence Analysis
- Investigation Timeline
- Reasoning Report
This allows moderators to understand why a recommendation was made rather than blindly trusting an AI score.
🧠 Context-Aware Decision Making
ModPilot evaluates content first.
User history and previous violations are treated as contextual information—not evidence of guilt.
This helps avoid a common moderation AI failure mode where users with poor histories are unfairly penalized for otherwise acceptable content.
✍️ AI-Assisted Moderator Responses
Moderators often spend time writing repetitive explanations.
After choosing an action, moderators can provide simple instructions such as:
"Approve this post but warn the user that stronger language may be removed in the future."
ModPilot generates a professional draft response aligned with community rules.
The moderator can edit, approve, or discard the draft before sending it.
📊 Moderation Analytics
Each subreddit receives a statistics dashboard showing:
- Investigations performed
- Average confidence
- Average latency
- Action distribution
- Recommendation trends
- Recent moderation activity
This gives moderators visibility into how moderation decisions are evolving over time.
⚙️ Community-Specific Configuration
Every subreddit is unique.
Moderators can configure:
- Community rules
- Geographic preferences
- Moderation strictness
- Workflow preferences
- Investigation behavior
- AI provider settings
This allows ModPilot to adapt to different moderation styles and community expectations.
How We Built It
ModPilot was built entirely on Reddit's Developer Platform using Devvit.
Core Stack
- Devvit
- TypeScript
- Devvit Web
- Redis
- Gemini 2.5 Pro
- Reddit Moderation APIs
Investigation Pipeline
Each investigation follows a structured workflow:
Report
↓
Strategy Selection
↓
Evidence Collection
↓
Rule Matching
↓
Reasoning
↓
Citation Validation
↓
Confidence Calibration
↓
Moderator Recommendation
This architecture allows the system to provide explainable recommendations rather than opaque moderation decisions.
Challenges We Faced
Balancing Context and Fairness
One of the biggest challenges was determining how much influence user history should have.
Early versions sometimes over-weighted prior violations, causing otherwise acceptable content to receive removal recommendations.
We redesigned the evaluation flow so that:
- Content is assessed first
- Rule violations are determined
- User history is only used as supporting context
This produced significantly more fair and trustworthy recommendations.
Explainability
Moderators need transparency.
A simple:
"Remove — 92% confidence"
is not enough.
We invested heavily in building investigation reports that clearly explain:
- What evidence was found
- Which rules were considered
- What increased confidence
- What reduced confidence
This became one of the defining aspects of the project.
Confidence Calibration
Another challenge was avoiding overconfident AI outputs.
Rather than displaying raw model confidence, ModPilot calibrates confidence using multiple factors including:
- Evidence convergence
- Rule-match strength
- Community history
- Available moderation signals
This allows the system to honestly communicate uncertainty when evidence is limited.
Building Within Platform Constraints
Our original architecture included an external Python-based reasoning engine.
However, platform networking constraints made external deployment impractical for the hackathon environment.
To solve this, we migrated the investigation engine into the Devvit application itself, resulting in a faster, more integrated solution that better aligned with Reddit's platform requirements.
What We Learned
Building ModPilot taught us that moderation is fundamentally an investigation problem rather than a classification problem.
The most valuable assistance is not automatically deciding what to do—it is helping moderators understand a situation faster.
We also learned the importance of:
- Human-in-the-loop workflows
- Explainable AI systems
- Confidence calibration
- Moderator trust
- Community-specific moderation standards
These lessons shaped nearly every major design decision in the project.
Future Improvements
We see several exciting directions for ModPilot:
- Semantic policy matching
- Enhanced moderator feedback learning
- Similar-case retrieval
- Improved confidence calibration
- Advanced moderation analytics
- Larger-scale community memory systems
Our focus will remain the same:
Helping moderators investigate faster while keeping humans firmly in control of moderation decisions.
Conclusion
ModPilot is not an automated moderator.
It is an investigation copilot for moderators.
By transforming reports into evidence-backed case files, surfacing relevant context, and helping moderators communicate decisions efficiently, ModPilot reduces moderation effort while preserving human judgment.
You decide. ModPilot provides clarity.
