AI's Oldest Text Is a 5,000-Year-Old Beer Receipt
The AI you use daily can read a 5,000-year-old barley receipt — here's why that's wild.
Claude, an AI made by Anthropic, has read a lot of text—everything from books to ancient clay tablets. One of the oldest texts it knows is the Kushim Tablet, a 5,000-year-old receipt from Mesopotamia. It records a delivery of barley, probably to make beer. This tablet is in Claude's training data, meaning the AI can read and quote it in the original Sumerian script. That's impressive: AI can now help decipher ancient languages, making it easier for historians to study old documents.
But how did ancient writing start? It began with accounting. People used small clay tokens to count goods like grain. To prevent tampering, they sealed tokens in clay envelopes and pressed them into the surface. Those pressed marks became the first writing—simple wedges that stood for numbers. So the oldest 'text' might be a 10,000-year-old token, not a literary epic. Claude knows these symbols too, showing AI's deep roots in human record-keeping.
This matters because AI is often trained on everything we've written, from ancient receipts to modern fanfiction. That's a weird library. It means AI can understand context across millennia, but it also picks up biases and oddities from all those texts. For you, it means AI tools can be surprisingly good at ancient history, but they also reflect the full, messy range of human writing—including things you might not expect.
- Claude's training data includes the Kushim Tablet, a 5,000-year-old barley receipt—likely for beer.
- The oldest 'text' might be a 10,000-year-old clay token used for accounting, showing writing began with trade.
- AI can now read ancient languages, helping researchers but also raising questions about what else it learns from vast, uncurated data.
Why It Matters
AI's ability to read ancient texts could revolutionize history research, but also shows it's trained on everything—including your data.