Researchers at OpenAI decided that a system that scores well at understanding language could too easily be manipulated for malicious intent.