Quick Summary:
|
|---|
The Problem: Confident and Wrong
AI-generated content has a specific failure mode that's more dangerous than an obviously bad answer: it's wrong in the same tone it's right. A hallucinated statistic reads exactly like a correct one. A fabricated citation looks exactly like a real one until someone checks. For a business shipping AI-written support responses, reports, or customer-facing content, that's not a quality issue — it's a liability sitting quietly in the pipeline until a customer, auditor, or reporter finds it first.
The instinctive fix — ask an AI model to check its own output, or ask a similar model to review it — has a well-documented weakness. Multiple independent analyses of the AI verification space describe this as structurally circular: the same blind spots that produced the hallucination are likely to confirm it, because the reviewing model shares the training data, the tone calibration, and often the specific gaps that created the error in the first place. It's the AI equivalent of grading your own homework.
That's the gap Lenz is built to sit in.
How Lenz Actually Works
Lenz is a fact-checking API — available at lenz.io — built around a five-stage pipeline that runs the same way whether it's called through an API in a production workflow or used to check a single pasted claim on the website.
Stage | What happens |
|---|---|
1. Framing | The claim is stripped of emotional language and distilled into a neutral, testable statement |
2. Research | Parallel web search across multiple independent search engines; quotes are extracted directly from source text and scored for authority, relevance, and recency |
3. Debate | Two separate AI models argue opposing sides across two rounds, including targeted rebuttals to each other's arguments |
4. Panel Review | Three separate AI models each score a different axis — source reliability, whether the evidence logically supports the claim, and whether the claim's specific numbers and wording match what the evidence says |
5. Conclusion | All debate and panel output is synthesized into one verdict (True, Mostly True, Mixed, Mostly False, or False) with a 1–10 score and a summary of where reviewers agreed or disagreed |
Lenz describes this as using eight models across the five stages. The company's own documentation explicitly accounts for five of those eight — two in the debate stage, three in the panel review stage — without a full public breakdown of where the remaining model calls happen in framing, research, or conclusion synthesis.
Every verification ships with the underlying evidence attached, so — per Lenz's own framing — a team using the output can check the reasoning rather than trust a bare score.
Where It Fits: API, MCP, CLI, or Zapier
Lenz is built to be embedded into a workflow, not to be a destination. The primitives are the same everywhere: /extract pulls verifiable claims out of any text, /assess returns a fast multi-model verdict for real-time use, /verify runs the full five-stage pipeline with citations, and /ask answers follow-up questions grounded in a completed verification.
Which access point makes sense depends on who's using it. A development team building fact-checking into a product's content pipeline will likely reach for the REST API directly, or the typed Python and TypeScript SDKs, or the CLI (pipx install lenz-io[cli]) for scripting and CI use. A non-technical operations team screening AI-generated content before it ships can add Lenz as a step in an existing n8n or Zapier workflow without writing any code — the native apps on both platforms expose the same four primitives as workflow steps. And a developer already working inside Claude Desktop, Claude Code, Cursor, or ChatGPT can point that assistant at Lenz's remote MCP server and get a fact-check without leaving the tool they're already using. There's also a WhatsApp number for a quick manual check on a single claim, sharing quota with the web account.
How Lenz Is Different From Other Verification Approaches
The current AI verification market splits roughly into three categories.
AI-authorship detectors — tools like Originality.ai or GPTZero — answer a different question entirely: whether text was written by a human or an AI. That's useful for policy enforcement, but it says nothing about accuracy. A document flagged as 100% AI-written can be completely correct; one flagged as human-written can be full of errors.
Single-model or same-vendor fact-checkers ask one model, or a closely related model from the same provider, to review the output. As covered above, this carries a structural weakness: shared blind spots between the generator and the reviewer.
Enterprise-only verification platforms, such as Norway-based Factiverse, offer genuine multilingual claim-detection capability but require a sales conversation — there's no self-serve tier to test the product against a real use case before committing.
Lenz's stated approach is multi-vendor, not just multi-model — the debate and panel stages draw on separate AI providers rather than variations of the same underlying model — combined with a free self-serve tier that lets a team test the pipeline against their own content before paying for anything.
Pricing, Honestly
Plan | Price | Deep verifications ( | Fast checks ( |
|---|---|---|---|
Free | $0 | 10/month | 100/month |
Plus | $7.99/month | 50/month | 500/month |
Developer | $99/month | 500/month | 5,000/month |
Scale | $399/month | 2,000/month | 20,000/month |
Enterprise | Custom | Custom, plus SLAs and white-label | Custom |
For most SaaS teams evaluating this seriously, the free tier is enough to test the pipeline against a handful of real claims from an actual product, but not enough to run in production. The Developer tier at $99/month is the realistic starting point for a small live integration — 500 deep verifications a month covers a modest but real content pipeline. Scale and Enterprise exist for teams running fact-checking as a standing part of a high-volume content or support pipeline.
AIWerse Verdict
The mechanism behind Lenz is a genuinely different answer to a real problem, not just marketing language wrapped around "multi-model" — the adversarial debate plus independent panel review structure directly addresses the documented weakness of same-model or same-vendor verification. The self-serve pricing, starting at free, is also a real point of difference against enterprise-only competitors in this space. Worth noting: this is Lenz's second Product Hunt launch, not a brand-new company, so treat it as an established product gaining fresh momentum rather than a fresh discovery. For any team shipping AI-generated content, reports, or support responses where a wrong fact carries real cost, it's worth testing against your own content on the free tier before deciding whether it earns a place in production.
FAQs
What is Lenz?
Lenz is a fact-checking API for AI products and teams. It takes a factual claim — generated by an AI system or pasted in manually — and returns a sourced verdict backed by independent research, AI debate, and panel review.
How is Lenz different from asking ChatGPT to fact-check itself?
Lenz's pipeline uses models from multiple separate AI vendors in an adversarial debate structure, plus independent panel review, rather than relying on one model (or a closely related model from the same provider) to check its own output — an approach several independent analyses describe as structurally prone to repeating the same blind spots.
What does Lenz cost?
Lenz has a free tier (10 deep verifications and 100 fast checks per month), a Plus tier at $7.99/month for individuals, a Developer tier at $99/month aimed at production integrations, a Scale tier at $399/month for higher volume, and custom Enterprise pricing.
Can Lenz be used inside Claude or Cursor?
Yes. Lenz runs a remote MCP server that connects to Claude Desktop, Claude Code, Cursor, and ChatGPT, letting a fact-check run without leaving the assistant.
Is Lenz built for developers or for anyone with a claim to check?
Both. Development teams can integrate the REST API, Python/TypeScript SDKs, or CLI directly into a content pipeline, while anyone can paste a single claim into Lenz's website or WhatsApp number and get a sourced verdict without writing any code.
Related News & Updates
Junaid Nawaz is the founder of AIwerse and a developer focused on AI tools, agentic workflows, and builder-focused tech. He covers AI model releases, coding tools, and platform updates for developers and teams building with AI. You can follow AIwerse on X (@AIwerse).