The fact-checking layer for content you can't afford to get wrong. Stamp AI checks every claim against screened live-web evidence in seconds and hands back a cited verdict you can defend. When the stakes are higher, a research panel of models from different providers goes deeper.
1,000 free credits on signup, no card required
Research Panel models from
Why Stamp AI
Language models fill gaps from memory, confidently. Stamp AI never does: it gathers the evidence first, screens it, and rules only on what the sources say. Then we check its citations.
97%
of high-confidence verdicts right, on real-world claims fact-checkers had settled
~8 s
a claim, with every claim in a document checked in parallel
7-pt
verdict scale. Misleading and Unverifiable are their own answers, never rounded to True or False
100%
of claims anchored to the exact sentence in your document
Worried a model will confidently make something up?
It never rules from memory. Stamp AI searches the live web before any model sees the claim, screens every result for relevance and prompt injection, and rules only on that evidence.
Read the methodologyHow do you know it's right?
When it says high confidence, it's right 97% of the time. Measured on 2,154 high-confidence verdicts across AVeriTeC and LIAR2, real-world claims professional fact-checkers had settled. When it is less sure, it says so, and those are the verdicts to read twice.
See the benchmarkNeed to show your work to an editor or a lawyer?
Sourced and auditable. Live-research verdicts cite the open web for every call. Every vote is on the record, and each report exports as a certificate you can hand over.
See a real reportGlobal EV Outlook 2025
iea.org · supports · audited
Electric car sales, 2012 to 2024
ourworldindata.org · supports · audited
EV Volumes annual review
ev-volumes.com · supports
Moving faster than a research desk can?
A whole document in well under a minute. Every claim is checked in parallel. Watch verdicts stream in, or get a signed webhook when the check completes.
Explore the API// POST → your callbackUrl, HMAC-signed
{
"type": "check.completed",
"check": { "status": "completed", "claims": 12 }
}How it works
Evidence first, ruling second, audit last. Results stream in as each claim lands.
Drop in a document or paste text. You see every claim we found and the exact price before anything runs.
Capabilities
Stamp AI for everything you publish; the Research Panel when a claim has to survive scrutiny. Every check is priced per claim and quoted before it runs, so the price you see is the price you pay.
Every claim checked against screened live-web evidence in seconds, with a cited verdict you can audit.
5credits / claim
See pricingUp to four models from different providers each research the claim, a challenger attacks unanimous calls, and a judge rules.
40credits / claim
See pricingSpelling and grammar across the whole document in one fast pass, with every fix shown as a diff.
5credits / check
See pricingDevelopers
Estimate, submit, poll and fetch checks over a token-authenticated API. Signed webhooks tell your CMS the moment a verdict lands.
FAQ
Everything you need before your first check.
A chatbot answers from memory and fills the gaps confidently. Stamp AI never rules from memory: it searches the live web for the claim and for any correction or denial, screens what comes back, and rules only on that evidence. Then we check that every source it cites actually says what it is cited for. For claims that must survive scrutiny, the Research Panel has models from different providers research each claim independently, and a judge weighs them. Read the full methodology
When Stamp AI says high confidence, it is right 97.2% of the time, measured on 2,154 verdicts on real-world claims that professional fact-checkers had settled (AVeriTeC and LIAR2). When it is less sure, it says so with a moderate or low band. The benchmarks page has the datasets, the method and every figure, including where it is weakest. See the benchmark
Stamp AI uses one fixed, fast model of our choosing, so its speed and price stay predictable. The Research Panel draws on models from Anthropic, OpenAI, Google, DeepSeek and other providers, routed through one gateway; organization admins can change the panel and the judge. See the model catalog
Yes, both depths do. Every verdict comes with the sources it relied on, so you can open each citation yourself instead of trusting a model's memory.
It's credit-based and priced per claim. Before you submit, we count the claims in your content and quote the exact price: the number of claims times the per-claim rate of the depth you chose, plus a small per-check fee (10 credits a claim, never more than 100). The quote is the charge. Re-checking a draft? Opt in and unchanged claims you verified in the last day keep their verdict for free. Every signup includes 1,000 free credits, and we refund failed checks automatically. See pricing
PDF, Word (.docx), plain text, and Markdown. Submit one document or a batch. Processing runs in the background, and results appear in the report as each claim finishes.
We use your documents only to run your checks and build the reports in your account. We don't sell your data or train models on it.
Get every claim checked against live evidence, with sources, before you publish.
Start freeContact
Questions about tiers, the API or volume pricing? We read every message. You can also email support@fact.engineering.