AI Transparency & Student Data

1. Overview

Ludwitt uses Anthropic's Claude API to power AI features across the platform. This document describes exactly what student data interacts with AI, how it is processed, and what safeguards are in place. It is intended for institutional compliance officers, FERPA reviewers, and university procurement teams evaluating Ludwitt for campus deployment.

2. AI Provider

Ludwitt partners with a single AI provider for all AI-powered features.

3. What Student Data IS Sent to AI

The following educational content is transmitted to Anthropic's API for grading and feedback purposes. None of this data contains personally identifiable information.

4. What Student Data is NEVER Sent to AI

The following categories of student data are never included in any AI API request. These protections are enforced at the application layer before any external API call is made.

5. Data Minimization Implementation

All AI API endpoints in Ludwitt follow a strict data minimization pattern. The AI never receives enough information to identify a specific student. The following sequence is enforced for every AI interaction:

6. AI Features and Their Data Usage

The following table enumerates every AI-powered feature in Ludwitt, the data sent to the AI provider, and the data returned.

7. Data Flow Diagram

The following diagram illustrates how data flows through the Ludwitt platform during an AI-powered interaction. Note that PII is stripped at the API route layer before any data reaches the AI provider.

┌─────────────┐     HTTPS/TLS 1.3      ┌──────────────┐
│   Student    │ ──────────────────────► │   Vercel     │
│   Browser    │                         │  Edge Network│
└─────────────┘                         └──────┬───────┘
                                               │
                                    ┌──────────▼───────────┐
                                    │  Next.js API Route    │
                                    │                       │
                                    │  1. Authenticate user │
                                    │  2. Validate request  │
                                    │  3. Strip all PII     │
                                    │  4. Build prompt with │
                                    │     content only      │
                                    └──────────┬───────────┘
                                               │
                              ┌────────────────┼────────────────┐
                              │                │                │
                    ┌─────────▼──────┐  ┌──────▼───────┐  ┌────▼──────────┐
                    │  Anthropic API  │  │  Firestore   │  │  Credit System│
                    │                 │  │              │  │               │
                    │  Receives ONLY: │  │  Stores:     │  │  Logs:        │
                    │  - Question     │  │  - Progress  │  │  - Token count│
                    │  - Answer       │  │  - XP/Streak │  │  - Cost       │
                    │  - Subject      │  │  - Accuracy  │  │  - Endpoint   │
                    │  - Difficulty   │  │  - Attempts  │  │  - Timestamp  │
                    │                 │  │              │  │               │
                    │  Does NOT       │  │  Protected   │  │  No content   │
                    │  receive:       │  │  by Firestore│  │  stored in    │
                    │  - Names        │  │  security    │  │  transaction  │
                    │  - Emails       │  │  rules       │  │  logs         │
                    │  - User IDs     │  │              │  │               │
                    └─────────────────┘  └──────────────┘  └───────────────┘

8. Data Retention

AI prompt content is never persisted. The question, answer, and AI response exist only in-memory during the API call and in the HTTPS response to the student's browser.

9. AI Training Opt-Out

Ludwitt's commitment: Student data is never used to train AI models.

• Anthropic's API Terms of Service state that API inputs/outputs are not used for model training
• Ludwitt does not fine-tune models on student data
• Ludwitt does not share student data with any third party for AI training purposes
• No student data is included in any dataset used for model improvement

10. AI Decision Transparency

How AI Grading Works

• AI does not assign grades autonomously. AI provides feedback and correctness assessment for individual practice problems. It does not determine course grades, GPAs, or academic standing.
• AI decisions are reviewable. Students see the AI's feedback immediately and can attempt problems again. There is no penalty system based on AI judgment.
• Adaptive difficulty is algorithm-based, not AI-based. Difficulty adjustment uses deterministic algorithms (Item Response Theory, Bayesian Knowledge Tracing, SM-2 spaced repetition) based on accuracy patterns — not AI model predictions.
• AI-generated content is labeled. When AI generates practice problems or explanations, the content is clearly presented as practice material, not as authoritative academic content.

11. Bias Mitigation

Ludwitt takes active measures to prevent AI bias from affecting student outcomes:

• AI grading prompts are standardized and do not include student identity information
• Grading rubrics are explicit in prompts (e.g., “check mathematical correctness”)
• AI does not have access to student demographics, race, gender, or socioeconomic data
• Regular prompt auditing ensures consistent grading across student populations

12. Institutional Controls

Universities deploying Ludwitt retain full control over AI feature usage within their organization.

13. Compliance Summary