AI Transparency

TrustLens AI exists because AI systems — including ours — make mistakes. Our philosophy is simple: don't trust AI, verify it. That applies to the answers you paste into TrustLens, and it applies to TrustLens itself. This page explains exactly how our analysis works, how scores are calculated, and where the limits are, so you can judge our output the same way we help you judge everyone else's.

TrustLens AI is powered by Anthropic's Claude, a large language model, combined with live web search and our own scoring logic. We do not train our own models, and no AI models are trained on your content. Every analysis follows the same principle: show your work. For AI answers, every claim links to the web sources used to check it. For contracts, every finding quotes the exact document text it is based on — no quote, no finding. You can always inspect the evidence yourself.

When you paste an AI-generated answer, four steps run: 1. Extract — Claude breaks the answer into individual factual claims and flags which ones need checking. 2. Search — for each important claim, we query a web search API (Tavily) for supporting or contradicting evidence. 3. Compare — Claude compares each claim against the evidence found and classifies it as Supported, Partially Supported, Unsupported, Contradicted, or Unverifiable. 4. Report — we compute a trust score and hallucination risk rating and assemble the report with all sources linked. "Supported" means supported by the evidence we found — not proven true. "Unsupported" means we didn't find evidence — not proven false.

When you upload a document, we extract the text (the file itself is discarded), classify the document type with a confidence level, and then Claude analyzes it clause by clause: what each clause says in plain English, why it matters, its risk level, and questions to ask before signing. Every finding must quote the document. If the AI can't point to the text a finding is based on, the finding is dropped. The executive summary shows what the document is, who it favors, where you could lose money, and your key obligations and deadlines. You can also ask the document questions in chat — answers quote the document or tell you the topic isn't addressed.

The trust score (0–100, higher is better) reflects how well the answer's claims held up against evidence. Supported claims and reliable sources raise the score; unsupported claims lower it; contradicted claims lower it heavily. Bands: 90–100 High trust — claims well supported by evidence 70–89 Mostly reliable — minor gaps 50–69 Needs review — several claims lack support 30–49 High risk — significant unsupported or contradicted claims 0–29 Do not rely on this answer The hallucination risk rating (Low / Medium / High) is driven by the share of unsupported and contradicted claims.

The contract risk score (0–100, lower is better) is deterministic: the same findings always produce the same score. Each finding's risk level contributes a weighted amount, and any high-risk finding prevents a document from scoring in the Low Risk band. Bands: 0–24 Low risk — largely standard, balanced terms 25–49 Moderate risk — some terms deserve attention 50–74 High risk — significant one-sided or costly terms 75–100 Critical risk — review carefully with a professional before signing Because clause detection is done by AI, two runs of the same document can detect slightly different findings, which can shift the score a little. Dramatic differences between runs are a bug — email us the report links.

Document classification comes with a confidence percentage. When confidence is low, or the document doesn't look like a contract, we say so on screen and you should weight the analysis accordingly. In answer verification, claims the system couldn't check are labeled Unverifiable rather than silently guessed at. When TrustLens doesn't know, it is designed to tell you — not to fill the gap.

Be aware of what TrustLens cannot do: • It can misread claims, miss clauses, or mislabel risk. Large language models make mistakes, including confident-sounding ones. • Web evidence reflects what was published and findable at query time. The web itself can be wrong, outdated, or incomplete — especially for very recent events and niche topics. • A high trust score is not proof of truth, and a low risk score is not a guarantee a contract is safe. • Scanned or image-only PDFs are not supported yet (no OCR). • It has no knowledge of your specific situation, jurisdiction, or negotiation context.

TrustLens reduces risk — it doesn't eliminate it. Use it to spot what deserves attention and to check claims faster than you could by hand. Then, for decisions that matter — signing a contract, publishing a claim, making a financial or medical choice — check the quoted sources yourself and involve a qualified professional. That's not a disclaimer we hide in the footer; it's the product philosophy: Verify Before You Rely.