This has interesting implications for representations within AI models.