This is a learning experiment about AI misinformation. Fauxthropic looks and feels like a real AI assistant, but everything it says is intentionally wrong.

It answers with confident, plausible-sounding claims, invented statistics, and made-up studies. Its citations are clickable and lead to real web pages, but those pages never actually support the claims.

Some "sources" are respectable sites cited for things they never said; others are satire or vaporware sites dressed up with official-sounding names. It's surprisingly hard sometimes to filter for truth in this day and age.

After each answer, press "Reveal what's wrong" to see exactly how each citation fails, and tips on how you could catch it yourself.

Ironically, Claude is used to label what is wrong, so it may also have its own biases and limitations in identifying the inaccuracies, you can view it as part of the learning experience.

This is not a problem contained within casual users. Even academic peer review is filling up with machine written text. ACM CHI is bracing for a wave of AI accelerated submissions 1. An analysis by Pangram also estimated that 21 percent of the 75,800 peer reviews at ICLR 2026 were fully AI generated 3 and GPTZero confirmed (or at least claims) over 100 hallucinated citations inside accepted NeurIPS papers 2.

It's important to say that the lesson isn't to not use AI, but instead that you need to verify the output.

I understand