The risks of OpenAI's Whisper audio transcription model