I recently read an article discussing how the phrase “vegetative electron microscopy” entered scientific literature due to a digitization error and eventually appeared in AI-generated responses. This got me thinking about how AI systems, including yours, can sometimes carry forward subtle human-made mistakes and treat them as fact.
I am curious about what kind of systems OpenAI has in place to catch these kinds of errors, both during the training process and afterward. Is there a team that regularly checks for outdated or inaccurate information in the model’s output, especially when it is rooted in mistaken but widely circulated data?
I am also interested in how OpenAI approaches community feedback in these situations. For example, are conversations or discoveries from users ever used to help improve future versions of the model?
I am not a journalist or researcher. I am simply someone who is genuinely interested in understanding the current limitations of AI and how we can all contribute to making it better.
Thank you for your time and for the work you are doing in this field.
Kind regards,
Samantha
Ottawa, Ontario, Canada