Lenz Review: Fact-Checking for AI Output That Has to Be True

What Is Lenz?

The platform is an audit-grade fact-checking service launched in 2026 and built around a single question that most AI tools sidestep: is what you answer actually true? Instead of asking one model to judge its own output, the platform extracts every claim worth checking, tests each one against independent sources, and runs a structured review before returning a verdict.

It was created in response to a familiar failure mode. Ask a single model whether a statement is accurate and you are trusting its memory and its blind spots at once. The design here spreads that judgment across multiple models and an explicit evidentiary trail, so the conclusion is something you can inspect rather than something you are asked to trust. Where models disagree, the difference is resolved by weighing the evidence each side produced, not by whoever happens to be more persuasive by default.

You can use it through a free web tool to check documents directly, or through the API and MCP server to embed verification inside a product that cannot afford to hallucinate.

How Lenz Verifies Claims

Verification follows a staged pipeline rather than a single model call. First, claims are extracted from the text so nothing gets skipped. Each claim is then framed precisely, because a vague statement is uncheckable and wording shifts the truth value of what you test.

Next come the sources. The platform searches independent references for the evidence relevant to each claim, then runs a debate between models positioned on opposing sides. A final review panel of three independent reviewers weighs the arguments and returns a verdict with citations and every step laid out.

The point of the ceremony is accountability. Most tools give you one model's best guess from memory; this one exposes the arguments for and against, the sources behind each side, and the reasoning trail that led to the score. That is what makes the output usable in an audit. Teams can replay the steps behind any verdict, which means a flagged claim can be investigated rather than argued about.

The company publishes research that shows why a panel matters. In its own study, five frontier models each ruled on the same 1,000 real claims, and the group fully agreed on only 37% of them; on 23%, verdicts differed by two or more points on the scale. A single model's answer can look confident and still be a minority opinion, which is the gap the pipeline is built to close.

• Claim extraction, framing, sourcing, debate, and review stages
• Opposing-side model debate surfaces counterarguments
• Three independent reviewers decide the verdict
• Every source and argument returned with the score

Lenz Scores and Verdicts

Output arrives as a score per claim rather than a blanket pass or fail. Each claim gets a verdict on how supported it is, backed by the sources and the debate steps that produced it. Unsupported claims are flagged so editorial or product teams can catch them before they ship.

The design deliberately handles the messy middle. Claims are rarely simply true or false — some are accurate but incomplete, some rest on weak sources, some are impossible to verify. The scoring reflects that spectrum instead of collapsing everything into a binary, which makes follow-up decisions defensible, and because each verdict carries its paperwork, the same claim can be re-checked later when sources have changed or new evidence has appeared.

For products embedding the service, that means risk becomes visible. A RAG pipeline that returns confident-looking nonsense gets caught at the scoring layer, and the audit trail tells a customer exactly which claim failed and why.

• Per-claim scores instead of a single pass or fail
• Sources and debate steps returned with each verdict
• Flags unsupported and weakly sourced claims
• Audit trail makes follow-up decisions defensible

Lenz API and MCP

For product teams, the service is exposed through a REST API with endpoints for the verification pipeline, from extracting claims out of raw text to asking a follow-up question about a checked claim. That maps naturally onto RAG pipelines and agent loops that need an independent check before output leaves the system.

An MCP server is also available, which means agent frameworks can call the fact-checker as a tool mid-task rather than bolting on verification afterward. An agent can draft, verify, and only then respond, keeping the safety mechanism inside the workflow instead of beside it.

The engineering reality is that verification adds latency. Checking a whole document takes longer than autocorrect — the trade is depth for speed, which is fine when the output is a report or a regulated answer and wrong when the use case tolerates guesses. In practice that means teams batch heavier checks behind the scenes and reserve synchronous checks for the few answers that matter most.

The published numbers make the trade explicit. A fast /assess verdict returns in roughly 5–10 seconds, a full /verify with the eight-model pipeline and citations takes about 90, and /extract is free for up to 1,000 calls per day. That split lets teams screen everything inline and escalate only the claims worth a deep check.

• REST API covering extract, assess, verify, and ask
• MCP server for in-workflow agent checking
• Fits RAG pipelines and regulated answer flows
• Verification depth costs latency by design

Who Lenz Is For

The natural audience is any team shipping machine-generated text that has to be right: enterprise AI products with RAG pipelines, media organizations verifying claims before publication, research platforms summarizing literature, and legal or compliance teams checking generated summaries.

Individual researchers and writers get the most accessible version through the free web tool, which reads a document and reports which statements hold up. For them the value is catching a confidently wrong citation or a fabricated fact before it enters an article or an essay, and the free tool makes that check a habit rather than a ceremony.

The fit weakens where speed is everything and mild inaccuracy is tolerable. A quick chat about general knowledge does not need an audit trail, so the platform is overkill for conversational search that tolerates an occasional wrong detail.

• Enterprise products shipping AI answers that must be true
• Media and research teams verifying claims before publication
• Free web tool suits researchers and writers
• Overkill where speed trumps occasional inaccuracy

Lenz Alternatives

Comparable research tools already live in this directory. Scite evaluates whether scholarly claims are backed by citations and lets you see supporting and contrasting evidence from the literature, which is the closest structural match for academic work.

Consensus and Elicit both answer research questions by synthesizing paper evidence, covering the discovery side of research rather than the verification side. If your problem is finding good evidence, they fit better; if your problem is checking claims that already exist in text, this platform is the more direct answer.

Evidence Core focuses on extracting evidence from documents on demand. That overlaps with the claim-extraction stage, while the full verification loop — debate, review, citations, scores — sets the fact-checker apart from both research search and document extraction tools.

• Scite — citation-backed claim evaluation for scholarly work
• Consensus and Elicit — evidence discovery for research questions
• Evidence Core — on-demand evidence extraction
• Full debate-to-verdict loop is what separates this from the pack

Lenz Pricing

The web-based claim checker is free to use, and API access is self-serve from day one. Published plans run from a Free tier with 100 monthly credits through Plus ($7.99/month), Developer ($99/month), and Scale ($399/month), with custom Enterprise pricing for higher volumes.

Best For

Recommended use cases and scenarios where Lenz shines.

Frequently Asked Questions

Common questions about Lenz, answered.