Receipts
Got an answer from an AI?
Ask for the receipts.
Paste it below. Receipts splits it into claims, finds the Wikipedia sentence that could back each one, and tells you which claims are backed, may conflict with Wikipedia, or have no receipt — always showing the sentence it relied on.
No account, no key. Models run in your browser. Each claim is searched on Wikipedia separately (plus the topic); nothing is sent to an AI company.
● Backed● May conflict● No receipt = check it yourself
What now?
Keep the sources for what is backed, and ask the chatbot to back up the rest.
How do we know it reads the evidence?
Small language-inference models are over-confident on sentences that are not about the claim: they often call them a "contradiction". A plain fact-checker would then tell you a true fact is false. Receipts only lets a sentence decide if it is close in meaning to the claim, and a conflict may only come from the most relevant sentence, about the same subject. Measured on public, human-labelled data:
Look for the ⚖︎ notes under a claim: they show what a simpler checker would have said, and why Receipts set that sentence aside.
What Receipts is not
Wikipedia is not the truth, and "no receipt" does not mean false — it means "go check this one". Receipts uses pretrained models only (no training) and the thresholds were chosen on separate development data. Full method, numbers and misses: docs/eval.md.