Researchers, scared by their own work, hold back “deepfakes for text” AI