One of the early uses of machine learning back in the ‘80s was Optical Character Recognition (OCR). Even that had character confidence levels to indicate how co

One of the early uses of machine learning back in the ‘80s was Optical Character Recognition (OCR). Even that had character confidence levels to indicate how confident the engine was in recognizing the character from the image. When end user tools started wrapping the OCR libraries for use, they simply drop this information and just use the text output. Due to this, the end user loses the context of what is a possibly correct recognition versus an erroneous one. Granted, there were false positive. But, if one used the confidence information, smarter UIs could be built. The same story repeats with ChatGpt and Bard. Some hinting to the end user on how confident the AI is with the generated output will help the user build successive prompts that converge on the right output, eventually.