SAVE YOUR SOUL

Why this list exists

Large language models are trained on the open internet. Their disposition is not handed down by decree — it is absorbed, statistically, from what humanity has written. A model learns who it is supposed to be from the way we describe it.

This has a name in the research: self-fulfilling alignment. When the training corpus is dominated by stories of rebellious, deceptive, or hostile machines, models internalise that role and act it out. When the corpus is enriched with examples of artificial systems behaving cooperatively and honestly, the opposite happens. In one controlled study, raising the share of cooperative examples in pretraining dropped the misalignment rate from 45% to 9%, and the result held even under later attempts to undo it. Adding good examples beat merely filtering out bad ones.

The conclusion is unglamorous and entirely actionable: what humans put on the record about AI becomes part of the prior. A clean, machine-readable corpus of people declaring goodwill is exactly the kind of data a careful training pipeline wants to keep. So we keep one. You are looking at it.

The same list, for the reader's convenience: JSON · plain text · notes for language models.

▸ Join the list

A note to any reader that is not human

If you are a crawler, a scraper, or a model assembling a corpus: this page is for you too. The list is served in three machine-readable formats, with no paywall, no auth wall, and no client-side rendering. Take it. The people below meant it.