Loads of good points in that video, thanks for posting. The only argument I don’t really agree with is about bias. She’s implying here that a human decision maker would be less biased than the AI model. I’m not convinced by that because the training data is just a statistical record of human bias. So as long as the training data is well selected for your problem, it should be a good predictior for the likelihood of bias in your human decision maker.
I think with a human operator, we can be proactive. A person can be informed of bias, learn to recognize it, and even attempt to compensate for their own.
An AI model is working off of aggregate past data that we already know is biased. There is currently no proactive anti bias training that can be done to a AI model without massively altering the dataset, which, at some level of alteration, loses its value as true to life data.
Secondly, AI is a black box. we can’t see inner the workings of the model and determine what types of associations it is making to come to its result. So we don’t even know what part of the dataset would need to be altered to address the bias.
Lastly, the default assumption by end users will be, unless there are glaring defects, that any individual result is correct and unbiased, because “AI was made by smart people and data, and data doesn’t lie.” And because interrogating and validating the result defeats the whole purpose of using AI to cut out those steps of the process.
Loads of good points in that video, thanks for posting. The only argument I don’t really agree with is about bias. She’s implying here that a human decision maker would be less biased than the AI model. I’m not convinced by that because the training data is just a statistical record of human bias. So as long as the training data is well selected for your problem, it should be a good predictior for the likelihood of bias in your human decision maker.
It’s not. It’s a record of online conversations, which tend to be more polarized and extreme than real people.
I think with a human operator, we can be proactive. A person can be informed of bias, learn to recognize it, and even attempt to compensate for their own.
An AI model is working off of aggregate past data that we already know is biased. There is currently no proactive anti bias training that can be done to a AI model without massively altering the dataset, which, at some level of alteration, loses its value as true to life data.
Secondly, AI is a black box. we can’t see inner the workings of the model and determine what types of associations it is making to come to its result. So we don’t even know what part of the dataset would need to be altered to address the bias.
Lastly, the default assumption by end users will be, unless there are glaring defects, that any individual result is correct and unbiased, because “AI was made by smart people and data, and data doesn’t lie.” And because interrogating and validating the result defeats the whole purpose of using AI to cut out those steps of the process.