AI TRANSPARENCY STATEMENT

for the LEX AI Service

Last updated: March 14, 2026

1. Purpose of This Statement

This AI Transparency Statement describes the safety mechanisms built into the LEX AI service to ensure the quality, accuracy, and reliability of AI-powered legal document analysis.

We believe that the legal domain demands the highest standards of accuracy, which is why we implement a multi-level system to protect against AI errors.

2. HallucinationGuard — AI Hallucination Protection

2.1. What Is HallucinationGuard

HallucinationGuard is a built-in safety mechanism that automatically verifies every AI response before it is presented to the User. The system prevents a common problem with AI models — generation of non-existent information (so-called "hallucinations").

2.2. How HallucinationGuard Works

The verification process includes several stages:

  1. Generated text analysis — the system analyzes every AI response for specific legal claims: references to legislation articles, court decision numbers, dates, judge names, etc.

  2. Cross-referencing with sources — every legal claim is verified against real databases:

    • Ukrainian legislation database (Verkhovna Rada, Zakon Online);
    • court decisions database (EDRS);
    • state registries (EDR, EDRPOU).
  3. Confidence assessment — the system assigns a confidence level to each claim:

    • High — the claim is confirmed by a real source;
    • Medium — the claim is partially confirmed or the source requires additional verification;
    • Low — the claim is unconfirmed, marked with a special indicator.
  4. Filtering and labeling — responses with unconfirmed claims are either blocked or labeled with warnings for the User.

2.3. What HallucinationGuard Verifies

  • Existence of cited articles, parts, and clauses of legislation;
  • Correspondence of court decision numbers to real EDRS records;
  • Correctness of legislation effective dates;
  • Correspondence of judge names to actual court compositions;
  • Logical consistency of legal conclusions.

3. CitationValidator — Citation Validation

3.1. What Is CitationValidator

CitationValidator is a specialized mechanism that verifies every citation in AI responses. When AI references a specific article of law or court decision, CitationValidator confirms that the citation matches the original.

3.2. How CitationValidator Works

  1. Citation detection — the system automatically identifies all references in the AI response to:

    • articles, parts, and clauses of laws and codes;
    • specific court decisions (by case number or registration number);
    • subordinate legislation (Cabinet of Ministers resolutions, ministerial orders, etc.).
  2. Original source retrieval — for each citation, the system retrieves the original text from official sources:

    • Verkhovna Rada of Ukraine portal;
    • Unified State Register of Court Decisions;
    • other official databases.
  3. Text comparison — the system compares the citation text in the AI response with the original and identifies:

    • full match — the citation is accurate;
    • partial match — minor differences exist (e.g., abbreviations);
    • mismatch — the text significantly differs from the original.
  4. User notification — the verification result is displayed in the response:

    • confirmed citations include links to the original source;
    • detected discrepancies are accompanied by warnings.

3.3. CitationValidator Verification Types

Verification TypeDescription
Norm existenceWhether the cited article/part/clause exists in the specified legislation
Norm validityWhether the norm is in effect as of the current date
Text accuracyWhether the cited text matches the original
Decision contextWhether the content of the court decision reference matches the actual decision

4. Semantic Ranking

4.1. What Is Semantic Ranking

Semantic Ranking is an intelligent search result ranking system that ensures the most relevant documents appear at the top of results.

4.2. How Semantic Ranking Works

  1. Vector representation — the User's query and documents in the database are converted into numerical vectors (embeddings) using OpenAI's text-embedding-3-small model. Text analysis and response generation are primarily performed by Anthropic Claude models (Opus, Sonnet, Haiku) via AWS Bedrock, with OpenAI models used as an auxiliary route.

  2. Semantic comparison — the system compares the query vector with document vectors by meaning, not just keywords. This enables finding documents that match by content even when they use different terminology.

  3. Multi-factor ranking — results are ranked by a combination of factors:

    • semantic similarity to the query;
    • court practice relevance (case category, court instance);
    • timeliness (decision date or latest amendments to legislation);
    • source authority (Supreme Court, appellate courts, etc.).
  4. Ranking transparency — for each result, the User can see:

    • relevance score;
    • document source;
    • search tools used.

5. Comprehensive Safety System

The three described mechanisms work together, creating a comprehensive protection system:

User Query
     ↓
Semantic Ranking → finds the most relevant sources
     ↓
AI Analysis → generates a response based on real sources
     ↓
HallucinationGuard → verifies facts in the response
     ↓
CitationValidator → validates every citation
     ↓
Result → verified response with confirmed references

5.1. Benefits of the Comprehensive Approach

  • Accuracy — multi-level verification significantly reduces the probability of errors;
  • Transparency — Users see sources and confidence levels for each conclusion;
  • Traceability — every processing step is logged for audit purposes;
  • Feedback — errors identified by Users are used to improve the system.

5.2. Limitations

Despite the comprehensive approach, no AI system can guarantee 100% accuracy. Possible cases include:

  • new or recently amended legislation may not yet be updated in the database;
  • complex legal situations may require interpretation beyond AI capabilities;
  • ambiguous wording in legislation may lead to different interpretations.

Therefore, AI analysis results are an auxiliary tool and do NOT replace professional legal advice.

6. Our Commitments

The Company commits to:

  1. Continuously improve AI safety mechanisms based on User feedback and new technologies;
  2. Ensure transparency — publish information about changes to the AI system and safety mechanism updates;
  3. Respond promptly to error reports and take measures to resolve them;
  4. Maintain standards — comply with EU AI Act, GDPR, and Ukrainian legislation requirements;
  5. Inform Users — clearly label AI content and provide links to original sources.

7. Contact Information

For AI transparency and safety mechanism inquiries:

Individual Entrepreneur Kyrychenko Igor Viktorovych Tax ID (RNOKPP): 2967214056 Address: 04207, Ukraine, Kyiv, 19 Levka Lukianenka St., apt. 110 Email: [email protected] Website: https://legal.org.ua


This Statement takes effect upon publication on the Service website.