How KLAR verifies results, which sources are used, and why you can trust our verdicts.
KLAR does not use a single AI model that guesses. Instead, every text goes through a multi-stage pipeline with independent verification steps:
Google Gemini 2.5 Flash identifies every factual statement in the text as a single, verifiable claim. Opinions, questions, and subjective statements are deliberately skipped.
For each claim, 4 source systems are searched in parallel: Wikipedia, Wikidata, Google Search Grounding (live web search), and our curated knowledge base. 3-10 sources per claim.
A second AI pass compares each claim against found sources and renders a verdict: Supported, Contradicted, or Unverifiable. Every verdict includes reasoning and a specific actionable recommendation.
NLP algorithms check source consensus, detect hallucination risks, and adjust confidence scores based on source agreement.
KLAR exclusively uses external, publicly accessible sources. The AI does not invent information — it can only confirm or contradict what exists in real sources.
Encyclopedic facts, dates, definitions
Real-time web search via Google's infrastructure
Nature, PubMed, arXiv, JSTOR, Springer
WHO, EU, Destatis, Federal Government, CDC
Correctiv, Snopes, Politifact, AFP Fact Check
Reuters, AP, BBC, Tagesschau, Spiegel
Every source receives a credibility score (0-1.0) based on: domain category (academic, government, news, social media, blog), historical reliability, and source diversity. Social media (Twitter, Reddit) and blogs receive low scores (0.15-0.45). Tabloid media like Bild are scored significantly lower than quality journalism.
KLAR uses Google Gemini 2.5 Flash — an existing Large Language Model by Google. We have NOT trained a custom model. Instead, we use advanced prompt engineering with structured JSON schemas to achieve precise and consistent results.
Can an AI detect hallucinations if it can hallucinate itself? Our solution: The AI is NOT asked to know facts. It is asked to compare claims against EXTERNAL sources. Gemini Search Grounding accesses live web data — the AI doesn't need to recall anything from memory. Additionally, NLP algorithms (not AI) check for hallucination risks in the AI output.
Bias detection, AI detection, and plagiarism checking use statistical NLP algorithms — no AI calls. These engines analyze language patterns, sentence structure, and text similarity using deterministic methods. They cannot hallucinate.
Every analysis displays the estimated and actual token consumption. You can see exactly how many tokens were used for your analysis — complete cost transparency.
The plagiarism check compares n-grams (4-word and 6-word sequences) of the input text against sources found during verification. It is NOT a database of all academic papers (like Turnitin). Instead, it checks whether the text has substantial overlap with the web sources used. The originality percentage shows how much of the text is unique.
All data in Frankfurt, Germany (Supabase EU)
Each user only sees their own data
Art. 17 (deletion) & Art. 20 (export) implemented
No cookies, no analytics, no data selling