Err, did you read your own link? This wasn’t a deliberate “watermark”. It was a training error that resulted in some odd characters being inserted. This was a training defect that was fixed, not some sort of plagarism countermeasure. These marks were present for only a short time and only on a subset of OpenAI models.