AI Transparency Statement

Last updated: April 2026

The TLG Claims Validation Platform uses AI as a tool within an evidence-based pipeline. This statement explains what AI does in our system, what it does not do, and how we ensure that every output traces to verifiable, published evidence.

1. Our Philosophy: AI as a Tool, Not an Oracle

The foundational principle of the TLG Claims Validation Platform is that AI is never the source of truth. AI is a powerful analytical tool — it can read documents, identify claims, search databases, and evaluate evidence relevance. But it can also hallucinate, confabulate, and produce plausible-sounding output that has no basis in fact.

We designed the Platform around this reality. Every AI operation in our pipeline is constrained by a hard grounding gate: no score, finding, or verdict is produced unless it traces back to a specific, published piece of evidence from a recognized authoritative source. If the evidence does not exist, the Platform says so — it does not fill the gap with AI-generated conjecture.

2. What AI Does in the Pipeline

AI performs specific, bounded tasks within a structured pipeline. Each task has defined inputs, expected outputs, and validation checks:

• Claim extraction. AI reads the uploaded document and identifies factual claims — efficacy statements, safety disclosures, statistical assertions — that can be verified against external evidence. Claims are extracted as structured data with specific text spans.
• Evidence selection. After evidence is retrieved from authoritative databases, AI evaluates which retrieved evidence items are most relevant to each specific claim. This is a relevance judgment, not evidence generation — the evidence already exists in the database.
• Evidence quality scoring. AI evaluates how well published evidence supports each claim across four quality dimensions: Population match, Endpoint match, Magnitude accuracy, and Context completeness. Each dimension is scored 0-2 with explicit justification.
• Fair balance assessment. AI compares the document's safety disclosures against the drug's FDA label and post-market adverse event data to evaluate whether safety information is presented fairly.
• Presentation analysis. AI evaluates the document for misleading language patterns and potential misinterpretation risk, calibrated to the target audience (healthcare professional vs. consumer).

3. What AI Does NOT Do

• AI does not generate medical facts. The Platform never asks AI "Is this claim true?" or "What are the side effects of this drug?" All factual information comes from published evidence databases.
• AI does not make final determinations. Scorecards are analytical tools, not verdicts. The output is designed to inform human reviewers, not replace them.
• AI does not replace professional review. The Platform is a starting point for medical, legal, and regulatory professionals — not a substitute for their judgment and expertise.
• AI does not have memory across runs. Each AI call is stateless — a fresh, independent operation with no conversation history or cross-run context. The model does not learn from or remember previous documents.

4. Evidence Sources

Every score on a TLG scorecard traces to evidence retrieved from one or more of these six authoritative databases:

• DailyMed — FDA-approved drug labels (prescribing information, package inserts)
• Europe PMC — peer-reviewed clinical and biomedical literature
• PubMed — review articles providing drug background and context
• ClinicalTrials.gov — registered clinical trial data and results
• FDA FAERS — post-market adverse event reports from the FDA's surveillance system
• Drugs@FDA — FDA drug approval history and regulatory actions

These are recognized, publicly accessible databases maintained by government agencies and academic institutions. The Platform queries them programmatically via their official APIs.

5. Quality Monitoring

We actively monitor AI performance within the Platform:

• Failure logging. When an AI call fails — due to malformed output, timeout, or other errors — the failure is logged with full diagnostic context. These logs contain the extracted claim text and evidence data that was being processed (not personal information) and are used to identify patterns and improve reliability.
• Structured output validation. Every AI response is validated against expected schemas. Responses that do not match the expected format are rejected and retried, not silently accepted.
• Claim ID integrity. Each claim carries a unique identifier through the pipeline. AI responses must echo the exact claim ID they were given — mismatches are rejected.
• Prompt injection defense. The Platform scans uploaded documents for prompt injection patterns before AI processing, uses per-call isolation boundaries, and instructs models to ignore embedded instructions.

6. Model Information

The Platform uses open-weight, biomedical-domain AI models specifically selected for their performance on medical and scientific text. Key characteristics:

• Open-weight models. We use openly available model weights, not proprietary third-party AI APIs. This gives us full control over the model's deployment and behavior.
• Dedicated hardware. Models run on dedicated GPU hardware rented for our exclusive use, not on shared multi-tenant AI services. Your document content is not processed by OpenAI, Google, Anthropic, or similar services.
• Biomedical specialization. We select models with training that emphasizes biomedical and clinical text understanding, improving accuracy on the type of content our users submit.
• No fine-tuning on user data. We do not fine-tune or train models on user-uploaded documents. The models are used as-is from their published weights.

7. Human Oversight

The scorecard produced by the Platform is designed as a starting point for human review, not a final verdict. The Platform provides tools for human oversight:

• Per-claim review. Reviewers can examine each claim individually, see the evidence it was scored against, and read the full source text.
• Accept / dispute / annotate. Reviewer tools allow qualified professionals to accept findings, dispute them with written rationale, or add contextual annotations.
• Cross-run notes. Reviewer annotations are shared within the organization, building institutional knowledge across multiple document reviews.
• Full evidence links. Every evidence item includes a link to the original source — the published paper, FDA label, clinical trial record, or adverse event report — so reviewers can verify the evidence themselves.

8. Our Commitment

We are committed to using AI responsibly and transparently. As AI technology evolves, we will continue to:

• Maintain the hard grounding gate — AI output without traceable evidence will never be presented as a finding.
• Be transparent about what AI does and does not do in our pipeline.
• Prioritize evidence provenance over AI confidence.
• Keep humans in the loop as the final decision-makers.

9. Questions

If you have questions about how we use AI, want to understand our methodology in more detail, or have feedback on our transparency practices, please contact us at:

The Liggett Group
Email: info@theliggettgroup.com
Website: www.theliggettgroup.com