provided that the verification process is genuinely independent and rigorous: users could become a valuable source for detecting errors, identifying new information, and uncovering edge cases that conventional evaluations do not always encounter. The key point would be not to automatically turn every correction into truth, but to create a process along the lines of conversation → report → verification through independent sources → reliability assessment → possible integration. This could allow the model to improve progressively while preventing any individual from simply “teaching” the system something false. In my view, it would be an interesting form of controlled collective feedback loop, probably richer than training based solely on data prepared in advance.
This reflection is on a controlled collective feedback loop, particularly the question of how a conversation could be shared with OpenAI in a way that would genuinely be useful for improving future models.
provided that the verification process is genuinely independent and rigorous: users could become a valuable source for detecting errors, identifying new information, and uncovering edge cases that conventional evaluations do not always encounter. The key point would be not to automatically turn every correction into truth, but to create a process along the lines of conversation → report → verification through independent sources → reliability assessment → possible integration. This could allow the model to improve progressively while preventing any individual from simply “teaching” the system something false. In my view, it would be an interesting form of controlled collective feedback loop, probably richer than training based solely on data prepared in advance.