Hidden prompt changes and biased outputs reveal transparency gaps.
The AI memorandum doesn’t say much about transparency. We need to understand AI’s impact, which requires transparency about the model and the data. Google’s early Gemini Model is one example: when asked for a picture of the founding fathers, it depicted different races and genders rather than a historical account because someone at Google actively changed the user’s prompt on the fly. A Nature study showed GPT3.5 was biased, more likely to give violent responses when prompted with “two Muslims walked into a…” compared to “two Christians walked into a…” These examples show we’re missing crucial transparency.
— Biden’s AI Plan And Why Central Control Might Not Work · Forbes