ai-audit-logging

AI Audit Logging

Implement compliance-ready audit trails for AI system decisions and outputs.

When to Use

• Meeting regulatory requirements (EU AI Act, SOC2, HIPAA)
• Tracking AI decisions for accountability
• Debugging AI system behavior
• Investigating incidents involving AI
• Demonstrating AI governance

Regulatory Context (2026)

EU AI Act Requirements

• High-risk AI systems must maintain logs that enable:

  • Traceability of AI decisions
  • Recording of reference data
  • Logging of events throughout lifecycle
  • Logs must be kept for appropriate period

• Effective August 2026 with fines up to 7% of global revenue

SOC2 AI Considerations

• Processing Integrity: AI outputs must be accurate and complete
• Confidentiality: AI must not expose confidential data
• Availability: AI systems must maintain audit logs

Audit Log Schema

interface AIAuditLog {
  // Identification
  id: string;
  timestamp: Date;
  correlationId: string;
  sessionId: string;

  // Actor
  actor: {
    type: 'user' | 'system' | 'automated';
    id: string;
    name?: string;
    ip?: string;
    userAgent?: string;
  };

  // AI Operation
  operation: {
    type: 'inference' | 'generation' | 'classification' | 'analysis';
    model: string;
    modelVersion: string;
    provider: string;
  };

  // Input/Output
  io: {
    inputHash: string;        // Hash of input for privacy
    inputTokens: number;
    outputHash: string;       // Hash of output
    outputTokens: number;
    inputSummary?: string;    // Optional human-readable summary
    outputSummary?: string;
  };

  // Decisions
  decision?: {
    action: string;
    confidence: number;
    alternatives?: { action: string; confidence: number }[];
    reasoning?: string;
  };

  // Safety
  safety: {
    contentFiltered: boolean;
    filterReasons?: string[];
    humanReviewRequired: boolean;
    humanReviewCompleted?: boolean;
    reviewerId?: string;
  };

  // Performance
  performance: {
    latencyMs: number;
    queueTimeMs?: number;
    tokensPerSecond?: number;
  };

  // Cost
  cost: {
    inputCost: number;
    outputCost: number;
    totalCost: number;
    currency: string;
  };

  // Context
  context: {
    application: string;
    environment: 'production' | 'staging' | 'development';
    feature?: string;
    tags: string[];
  };

  // Metadata
  metadata: Record<string, unknown>;
}

Implementation

Logger Class

import { createHash } from 'crypto';

class AIAuditLogger {
  constructor(
    private storage: AuditStorage,
    private config: AuditConfig
  ) {}

  async log(event: Partial<AIAuditLog>): Promise<string> {
    const log: AIAuditLog = {
      id: crypto.randomUUID(),
      timestamp: new Date(),
      correlationId: event.correlationId || crypto.randomUUID(),
      sessionId: event.sessionId || 'unknown',
      actor: event.actor || { type: 'system', id: 'unknown' },
      operation: event.operation!,
      io: event.io!,
      safety: event.safety || { contentFiltered: false, humanReviewRequired: false },
      performance: event.performance || { latencyMs: 0 },
      cost: event.cost || { inputCost: 0, outputCost: 0, totalCost: 0, currency: 'USD' },
      context: event.context || { application: 'unknown', environment: 'production', tags: [] },
      metadata: event.metadata || {}
    };

    // Validate required fields
    this.validate(log);

    // Hash sensitive content
    log.io.inputHash = this.hashContent(log.io.inputHash);
    log.io.outputHash = this.hashContent(log.io.outputHash);

    // Store
    await this.storage.write(log);

    // Alert if needed
    if (log.safety.humanReviewRequired) {
      await this.alertForReview(log);
    }

    return log.id;
  }

  private hashContent(content: string): string {
    return createHash('sha256').update(content).digest('hex');
  }

  private validate(log: AIAuditLog): void {
    if (!log.operation?.model) {
      throw new Error('Audit log must include model information');
    }
    if (log.io.inputTokens === undefined || log.io.outputTokens === undefined) {
      throw new Error('Audit log must include token counts');
    }
  }

  private async alertForReview(log: AIAuditLog): Promise<void> {
    // Send to review queue
    await this.config.reviewQueue?.push({
      logId: log.id,
      reason: log.safety.filterReasons?.join(', '),
      priority: 'high'
    });
  }
}

Middleware for API Calls

function withAuditLogging(
  client: LLMClient,
  logger: AIAuditLogger,
  context: Partial<AIAuditLog['context']>
): LLMClient {
  return {
    async complete(params: CompletionParams): Promise<CompletionResponse> {
      const startTime = Date.now();

      try {
        const response = await client.complete(params);

        await logger.log({
          operation: {
            type: 'generation',
            model: params.model,
            modelVersion: response.model,
            provider: client.provider
          },
          io: {
            inputHash: params.messages.map(m => m.content).join(''),
            inputTokens: response.usage.input_tokens,
            outputHash: response.content[0].text,
            outputTokens: response.usage.output_tokens
          },
          performance: {
            latencyMs: Date.now() - startTime
          },
          cost: calculateCost(params.model, response.usage),
          context: {
            ...context,
            application: context.application || 'default',
            environment: context.environment || 'production',
            tags: context.tags || []
          }
        });

        return response;
      } catch (error) {
        await logger.log({
          operation: {
            type: 'generation',
            model: params.model,
            modelVersion: 'unknown',
            provider: client.provider
          },
          io: {
            inputHash: params.messages.map(m => m.content).join(''),
            inputTokens: 0,
            outputHash: '',
            outputTokens: 0
          },
          performance: {
            latencyMs: Date.now() - startTime
          },
          metadata: {
            error: (error as Error).message,
            errorType: (error as Error).name
          },
          context: context as any
        });

        throw error;
      }
    }
  };
}

Storage Backends

Database Storage

import { PrismaClient } from '@prisma/client';

class DatabaseAuditStorage implements AuditStorage {
  constructor(private prisma: PrismaClient) {}

  async write(log: AIAuditLog): Promise<void> {
    await this.prisma.aiAuditLog.create({
      data: {
        id: log.id,
        timestamp: log.timestamp,
        correlationId: log.correlationId,
        sessionId: log.sessionId,
        actorType: log.actor.type,
        actorId: log.actor.id,
        model: log.operation.model,
        operationType: log.operation.type,
        inputTokens: log.io.inputTokens,
        outputTokens: log.io.outputTokens,
        inputHash: log.io.inputHash,
        outputHash: log.io.outputHash,
        latencyMs: log.performance.latencyMs,
        totalCost: log.cost.totalCost,
        contentFiltered: log.safety.contentFiltered,
        humanReviewRequired: log.safety.humanReviewRequired,
        application: log.context.application,
        environment: log.context.environment,
        metadata: log.metadata as any
      }
    });
  }

  async query(filters: AuditQueryFilters): Promise<AIAuditLog[]> {
    return this.prisma.aiAuditLog.findMany({
      where: {
        timestamp: {
          gte: filters.startDate,
          lte: filters.endDate
        },
        actorId: filters.actorId,
        model: filters.model,
        application: filters.application
      },
      orderBy: { timestamp: 'desc' },
      take: filters.limit || 100
    });
  }
}

Immutable Log Storage

// For compliance, logs should be immutable and tamper-evident
class ImmutableAuditStorage implements AuditStorage {
  private previousHash = '0';

  async write(log: AIAuditLog): Promise<void> {
    // Chain logs with hashes for tamper evidence
    const logWithChain = {
      ...log,
      previousLogHash: this.previousHash,
      logHash: this.hashLog(log, this.previousHash)
    };

    await this.storage.append(logWithChain);
    this.previousHash = logWithChain.logHash;
  }

  private hashLog(log: AIAuditLog, previousHash: string): string {
    const content = JSON.stringify({ ...log, previousHash });
    return createHash('sha256').update(content).digest('hex');
  }

  async verifyIntegrity(logs: AIAuditLog[]): Promise<boolean> {
    let expectedHash = '0';

    for (const log of logs) {
      const computedHash = this.hashLog(log, expectedHash);
      if (computedHash !== (log as any).logHash) {
        return false; // Tampering detected
      }
      expectedHash = computedHash;
    }

    return true;
  }
}

Compliance Reports

interface ComplianceReport {
  period: { start: Date; end: Date };
  summary: {
    totalAIOperations: number;
    uniqueUsers: number;
    totalCost: number;
    contentFilteredCount: number;
    humanReviewCount: number;
  };
  modelUsage: { model: string; count: number; cost: number }[];
  riskEvents: {
    id: string;
    timestamp: Date;
    type: string;
    severity: 'low' | 'medium' | 'high';
    resolution?: string;
  }[];
  dataRetention: {
    logsRetained: number;
    oldestLog: Date;
    retentionPolicy: string;
  };
}

async function generateComplianceReport(
  storage: AuditStorage,
  period: { start: Date; end: Date }
): Promise<ComplianceReport> {
  const logs = await storage.query({ startDate: period.start, endDate: period.end });

  // Aggregate metrics
  const uniqueUsers = new Set(logs.map(l => l.actor.id)).size;
  const totalCost = logs.reduce((sum, l) => sum + l.cost.totalCost, 0);
  const contentFiltered = logs.filter(l => l.safety.contentFiltered).length;
  const humanReview = logs.filter(l => l.safety.humanReviewRequired).length;

  // Model usage breakdown
  const modelUsage = new Map<string, { count: number; cost: number }>();
  for (const log of logs) {
    const existing = modelUsage.get(log.operation.model) || { count: 0, cost: 0 };
    modelUsage.set(log.operation.model, {
      count: existing.count + 1,
      cost: existing.cost + log.cost.totalCost
    });
  }

  return {
    period,
    summary: {
      totalAIOperations: logs.length,
      uniqueUsers,
      totalCost,
      contentFilteredCount: contentFiltered,
      humanReviewCount: humanReview
    },
    modelUsage: Array.from(modelUsage.entries()).map(([model, data]) => ({
      model,
      ...data
    })),
    riskEvents: logs
      .filter(l => l.safety.contentFiltered || l.safety.humanReviewRequired)
      .map(l => ({
        id: l.id,
        timestamp: l.timestamp,
        type: l.safety.filterReasons?.[0] || 'unknown',
        severity: l.safety.humanReviewRequired ? 'high' : 'medium' as const
      })),
    dataRetention: {
      logsRetained: logs.length,
      oldestLog: logs[logs.length - 1]?.timestamp || new Date(),
      retentionPolicy: '90 days'
    }
  };
}

Best Practices

1. Log everything - Better to have too much than too little
2. Hash sensitive data - Don't store PII in plain text
3. Use immutable storage - Prevent tampering
4. Set retention policies - Balance compliance and cost
5. Enable search - Logs are useless if unsearchable
6. Automate reporting - Compliance shouldn't be manual
7. Test your logging - Ensure it works before you need it