OpenAI's internal safety mechanism to prevent minor access to 'Adult Mode' is still in development with a 12% error rate where it misclassified minors as adults.