We asked five frontier AI models to fact-check the same 1,000 real-world claims. The headline finding from our LLM disagreement research: - On 63% of real-world fact-check claims, at least one of five frontier models disagreed with the others. - On 23%, the disagreement was substantive - verdicts were at least two categories apart. The dataset behind the research is now on @huggingface ↓ lnkd.in/dDmUW55h 🛠️ Fork it, benchmark with it, build on it, and let us know - we'd be happy to chat!

Lenz is now on @n8n_io AI-generated reports, newsletters, summaries - get them fact-checked before they publish. The node returns a verdict, score, and citations for every claim. Route anything that fails for human review before it ships → n8n.io/integrations/lenz #n8n #automation #AIfactchecking

We measured how often five new frontier models — Fable 5, GPT-5.5, Gemini 3.1 Pro, Sonar Deep Research, and Grok 4.3 — disagree when checking the same real-world claims. Version 1.1 of our LLM disagreement study arrives soon: a five-model frontier panel with retrieval, a five-point verdict scale, reasoning traces, and a fresh dataset. The methodology was iterated directly from the Hacker News discussion that put Version 1.0 on the front page. The finding that shaped the Lenz pipeline: even well-calibrated frontier models regularly diverge on the same inputs. That’s why Lenz runs models from multiple vendors and preserves disagreement rather than averaging it away. The trace behind every verdict includes what each reviewer found, the evidence it used, its reasoning summary, and where the models differed. If you’re shipping AI output to users, the question isn’t whether your model hallucinates. It’s whether your process catches it before it reaches them. Version 1.1 coming soon. #AI #LLMs #GenAI #MachineLearning #AIResearch

Lenz is now a remote MCP server. Add it to Claude Desktop, Claude Code, Cursor, VS Code, or any MCP client, and your assistant can fact-check a claim against independent sources — a sourced verdict with citations, without leaving the chat. • assess: 3-model panel verdict, ~5–10s • verify: full 8-model pipeline with citations, ~90s • ask: grounded follow-ups on a completed verification Lenz checks claims against retrieved sources, with models from rival vendors, outside the model that wrote the content. An assistant can now cite its own outside check. Start free, no card required: lenz.io/mcp-server --- Lenz is a fact-checking API — an independent layer that verifies claims in AI-generated content before it ships, with sources and a full audit trail. #MCP #AI #LLM #Claude #Cursor #VSCode #AIFactChecking #GenAI #DeveloperTools

Lenz is a fact-checking API for LLM output: claim-by-claim verification against independent sources, before it ships. Your product writes documents, summaries, answers - and increasingly acts on them. Lenz catches AI hallucinations before a wrong summary becomes a wrong action. → 8 models from rival vendors, 5 stages: Framing → Research → Debate → Panel Review → Conclusion → /assess — 3-model panel verdict in ~5–10s → /verify — full 8-model pipeline with citations, ~90s → /extract — pulls verifiable claims from any text, free, 1,000/day → structured JSON: 5-bucket verdict, Lenz Score, citations, full trace — block, publish, or escalate lenz.io/verify-ai-output #LLM #AIEngineering

The popular framing — "bots run social media" — relies on a real number applied to the wrong metric. Bots are ~50% of internet traffic. Different from social-media engagement (5–15%). Verified: lenz.io/c/bot-engagement-soc… #AI #SocialMedia

Most source-checking asks: is this a credible source? The harder question: is the specific claim inside it accurate? A peer-reviewed paper can include findings that don’t replicate. A government site can rely on outdated data. A respected author can misread their own sources. Credibility ≠ correctness. Read: How to verify sources for a research paper → lenz.io/blog/how-to-verify-s… #Research #Citations #AcademicIntegrity

We asked Lenz: "Do large language models hallucinate less than 5% of the time?" The actual rate across independent audits: 28–45% depending on task type. If you ask ChatGPT about its own error rate, it'll understate it. That's not a bug. It's the blind spot adversarial systems exist to catch. lenz.io/c/ai-language-models… #AI #Hallucination #FactCheck

Journalists spend hours per story on source verification. Lenz maps the evidence on any topic — 40+ primary sources, tiered by credibility, in ~30 seconds. Not a replacement for judgment. An accelerator for it. lenz.io/verify-before-publis… #Journalism #Research #AI #FactCheck

We've verified 400+ claims people actively share, cite, and debate. The numbers: — 47% simply False — 25% Misleading — 10% fully True A confident, widely-repeated claim is a reason to check. Not trust. lenz.io/blog/things-we-learn… #Research #Misinformation #FactCheck

The day we launched on @ProductHunt, someone in the community (@milkoslavov ) shipped a Lenz CLI and a custom skill. We didn't ask. They just built it. Huge thank you to Milko for turning curiosity into code so fast ⚡️ That's the ecosystem we're building. lenz.io #BuildInPublic #OpenSource #AI

AI doesn't hallucinate randomly. It reproduces whatever was repeated most — regardless of whether it's true. We fact-checked 3 AI answers that sound completely right. All three are False. lenz.io/blog/ai-hallucinatio… #AIHallucination #FactCheck #Research

"Why use Lenz when I can just ask ChatGPT / Perplexity / Grok?" A single model inherits its own training biases. Ask it about something it believes and it'll confirm it. The adversarial structure exists to catch what one model misses. It's not a smarter model — it's a different architecture. lenz.io #AI #FactCheck #Epistemics

Not all sources are equal. We built a 4-tier credibility system so you can see exactly *what* a claim is based on. Tier 1: Peer-reviewed studies, government statistics, systematic reviews Tier 2: Reputable journalism, expert institutions Tier 3: Opinion, commentary, secondary sources Tier 4: Unverified, flagged, or retracted The hard problem: a retracted PNAS paper cited by a credible blog. The blog is Tier 2. The paper is Tier 4. Which tier does the claim get? We surface both. You see the chain. That's by design.

The 3-Judge Panel at Lenz.io After two AIs debate your claim, 3 independent models evaluate the arguments — each focused on a different axis: logical fallacies, ideological bias, source quality. Majority verdict wins. No single model can swing the outcome. Jury-style fact-checking. Because one opinion isn't a verdict. #AI #FactCheck

At Lenz, two AIs Debate Each Other One AI will confidently tell you something wrong. Two AIs arguing? Much harder to fool. Lenz runs a structured adversarial debate — one model argues Pro, one argues Con. Then 3 independent judges evaluate the arguments. It's how peer review works. We automated it. lenz.io #AI #FactCheck #Epistemics