Site Reveals Which People AI Models Memorized

Two former OpenAI engineers, Joey Flynn and Thomas Dimson, launched a website called In the Weights that tests whether an AI model can recall a specific person purely from its training data — no web search, no tools, just what the model absorbed during training. It queries several models and returns a single number, a "strength score," that estimates how deeply a given name is baked into the weights.

The scale tops out at 996, and the leaderboard reads like you'd expect: Mozart, Shakespeare and Taylor Swift sit at the ceiling, the people most heavily represented across the text the models were trained on. Lesser-known names score far lower — The Decoder noted two of its own writers landing at 175 and 262 — which is exactly the point. The tool turns an abstract question, 'does the AI know who I am,' into a concrete figure you can look up.

Why this matters

Being "in the weights" is different from being findable on the web. A model that recalls you without a search has effectively memorized you: your name was common enough in the training text that the system kept it. That has real consequences. It shapes whether a chatbot can describe you accurately or confidently makes things up, and it's a window into whose information these systems prioritized — and whose they ignored. For public figures, journalists and anyone whose name shows up online, it's a rough gauge of how the dominant AI tools will represent them by default.

The creators are candid about the limits. Smaller models — they cite a one-billion-parameter Llama — are harder to read cleanly. Models can hallucinate biographical details, typos in a name drag the score down, and common names tend to return worse results because the signal smears across many people who share them. So a low score isn't proof you're absent from the data, and a high one isn't proof the model has its facts straight.

What makes the project useful is less the exact numbers than the framing. Training data has been one of the most opaque parts of modern AI: companies rarely disclose what went in, and users have little way to probe what came out. A simple lookup that exposes whether a model retained a person is a small but honest step toward making that black box legible — and a reminder that when a chatbot speaks confidently about someone, it may be reciting from memory, or just filling in the blanks.