Absorb False

Ars Technica: AI models can learn a falsehood even when the data calls it false

If a large language model sees a false claim during training alongside a clear warning that it is false, the warning may not be enough.