Check it before you share it.
An open, free educational protocol for questioning claims you see online - news, social media, or AI-generated - using AI tools you already have. Not medical, financial, or legal advice. Canonical terms: AGREEMENTS.md.
Truth is what survives contact with the original record.
CrossCheck is a short process anyone can follow to practice examining something they heard before acting on it or sharing it. You define one claim, run three independent AI passes (validator → referee → meta-referee), then read primary sources yourself.
It takes about ten minutes. It uses free tools. It works for many everyday claims you encounter online - from product safety to breaking news. It is not a substitute for professional medical, mental-health, financial, tax, investment, or legal advice.
Try the interactive tool — or use the copy-paste prompt templates in this repository with any AI you already have.
| Step | What You Do |
|---|---|
| Claim | Write the exact wording you want to verify (question, statement, rumor, quote — anything). One claim only — no “opposite angle” field. |
| 1 — Validator | New chat → primary-validator template: short answer grounded in primary sources plus a structured CrossCheck summary in the same reply. |
| 2 — Referee | New chat → referee template: embeds your claim and the full validator paste; critiques that pass (not a competing narrative). |
| 3 — Meta-referee | New chat → meta-referee template: embeds claim, validator paste, and referee report; stress-tests the referee against primaries. |
The golden rule: No matter how many AIs agree, click through to the actual primary source and read it yourself.
Don't have three different tools? Open separate new chats or temporary tabs in the same provider — each conversation must start fresh so no answer influences the next.
Long-form guidance (triage, expanded steps, hallucination checks, checklist): CrossCheck_Protocol.md.
| File / folder | Description |
|---|---|
templates/ |
Copy-paste prompts with {{CLAIM}} and paste placeholders — use offline or alongside any chat UI |
CrossCheck_Protocol.md |
Long-form protocol: triage, depth, examples, tests, checklist |
AGENTS.md |
Guidance for AI models in the three passes |
AGREEMENTS.md |
Terms and agreements (canonical) |
LICENSE |
CC BY 4.0 |
README.md |
This overview |
The hosted tool at atom.builders/crosscheck implements the same protocol with a guided form (claim field, tool lineup, prompt generation, paste handoffs). It is not part of this documentation repository.
Any combination works.
| Tool | Free Tier | Link |
|---|---|---|
| ChatGPT | Yes | chat.openai.com |
| Gemini | Yes | gemini.google.com |
| Claude | Yes | claude.ai |
| Copilot | Yes | copilot.microsoft.com |
| Grok | Yes (on X) | x.com |
| Meta AI | Yes | meta.ai |
| DeepSeek | Yes | chat.deepseek.com |
| Le Chat (Mistral) | Yes | chat.mistral.ai |
| Perplexity | Yes | perplexity.ai |
| Proton Lumo | Yes | lumo.proton.me |
A single AI model can:
- Agree with your framing instead of correcting it (sycophancy)
- Invent a citation that looks real but links to nothing (hallucination)
- Reflect its training biases without flagging them
- Sound confident even when it's wrong
CrossCheck chains three independent model passes and still expects you to read primaries — so one model’s blind spots are less likely to become yours.
- News Literacy Project - nonpartisan educational resources
- Stanford Civic Online Reasoning - teaches "lateral reading"
- AP Fact Check - Associated Press fact-checking hub
- Reuters Fact Check - Reuters fact-checking desk
- IFLA: How to Spot Fake News - library-standard infographic
- Towards Understanding Sycophancy in Language Models - Sharma et al., 2023
- A Survey on Hallucination in Large Language Models - Huang et al., 2023
- A Multitask, Multilingual, Multimodal Evaluation of ChatGPT - Bang et al., 2023
- PubMed - biomedical research
- Google Scholar - academic papers across all fields
- CPSC - U.S. product safety
- CDC - U.S. public health
- FDA - drugs, food, medical devices
- WHO - international health
- Congress.gov - U.S. federal legislation
- IRS - U.S. tax rules
- PACER - U.S. federal court records
- National Weather Service - official forecasts
CrossCheck materials in this repository are offered for educational and academic purposes only. Full binding text: AGREEMENTS.md.
Not professional advice. Nothing here is medical, financial, or legal advice. Do not use these materials in place of a qualified professional. If you may have a medical emergency, call your local emergency number.
No warranty. Materials are provided "as is." AI systems can hallucinate.
CrossCheck is a living document. If you find a broken link, a better primary source, or a way to make the steps clearer, open an issue or submit a PR.
The CrossCheck documentation in this repository is licensed under CC BY 4.0. See LICENSE and Creative Commons BY 4.0.
CrossCheck - Check it before you share it. - atom.builders/crosscheck