About
AI Receipts is a weekly newsletter for people who want to know what's real behind AI announcements, in plain English. Hype, itemized.
What we do
- Check claims. Each week we pick the loudest AI claims and hold them up to primary sources: lab blogs, papers, filings, pricing pages, independent evaluations.
- Call out the spin. What a claim means, what it leaves out, and whether it holds up. If a number is vendor-measured, we say so.
- Give a verdict. ✅ Checks out, ⚠️ Partly, or ❌ Doesn't add up. No mush.
What we don't do
We audit announcements, not people. We don't cite screenshots of posts as proof, and we don't repeat vendor benchmarks as fact. Sponsors, when we have them, are labelled and never affect a verdict.
How we check a claim
Every receipt follows the same five steps. If we can't complete one, we say which.
1. Pin down the exact claim
We quote the claim as made, with who made it and when, and link where it appeared. If a headline paraphrases a company, we go back to the company's own words.
2. Find the primary source
Lab blog posts, model cards, papers, filings and official pricing pages come first. Secondary coverage is used for context and labelled as such. A screenshot of a post is never proof.
3. Ask who measured it
Vendor-reported benchmarks are flagged as vendor-reported. We prefer independent evaluations and note when the comparison set was chosen by the vendor.
4. Show the math
We itemize: the numbers, what they're compared against, and anything temporary, plan-gated or undisclosed (sample sizes, failure counts, model names, dates). Estimates are labelled as estimates.
5. Give a verdict
| Verdict | Meaning |
|---|---|
| ✅ Checks out | The numbers support the claim as stated. |
| ⚠️ Partly | True with important conditions, or only part of the claim holds. |
| ❌ Doesn't add up | The evidence contradicts the claim, or the claim can't be supported. |
When we're wrong
We fix it in public. See the corrections policy and log.
Who writes it
AI Receipts is written by The AI Receipts team. Every verdict follows the method above and our corrections policy.
Contact
Reply to any issue. We read every reply, especially the ones that say we got a number wrong.
Get the receipts every Sunday
The week's AI hype, fact-checked. Plain English, free.