3 ms·

I think the people who keep harping about bias are trying to convey a narrow point. My point is broader. Probabilistic models are probabilistic. I am simplifyi

I think the people who keep harping about bias are trying to convey a narrow point. My point is broader. Probabilistic models are probabilistic. I am simplifying the behavior of these systems, but my argument is twofold. First,, just because an output has a high probability of being correct, doesn't mean that any particular output is correct. Second, "low probability" events that are acceptable in a limited use case are disastrous in broader use. For example, If I were being generous, I'd say that GPT-4 makes an error 0.1% of the time i.e. 99.9% of the time it doesn't make an error. I would say that is extremely fair. I use this model daily and I've found the rate in my limited sample set to be higher than that. If you are dealing with 10 cases, a 0.1% error rate is immaterial. If you are dealing with 100,000 cases, that's 100 cases where an error was introduced. The true error rate is likely to be higher, for example, https://www.ncbi.nlm.nih.gov/corecgi/tileshop/tileshop.fcgi?p=PMC3&id=9052&s=177&r=2&c=2 https://www.ncbi.nlm.nih.gov/corecgi/tileshop/tileshop.fcgi?... Is it acceptable for a few thousand to a few hundred innocent people to face legal action because an overgrown next token prediction model made a mistake? I love GPT-4. I love its promise, but I don't think it should be implemented in safety-critical situations.

sottol 2y ago
I am also worried about officers "gaming" the system to bolster whatever stat their superiors want them to optimize at the detriment of those being "policed". We've all seen how prompting can change GPT output and this is what worries me the most. Imagine users finding out that starting every interaction that is recorded with a certain sentence acts as a pre prompt that in turn subtly biases the gpt output so that it makes judges or prosecutors less sympathetic to defendants when the police department is optimizing for higher conviction rates, eg. (That is, doing the opposite of the quote above of the public defender "humanizing" the drug case defendant in the plea deal example.) There will be a lot of reports transcribed, people will find ways to optimize the output to their gain and we've all seen how easy it is to bias llm output by promoting.

sottol 2y ago

I am also worried about officers "gaming" the system to bolster whatever stat their superiors want them to optimize at the detriment of those being "policed". We've all seen how prompting can change GPT output and this is what worries me the most. Imagine users finding out that starting every interaction that is recorded with a certain sentence acts as a pre prompt that in turn subtly biases the gpt output so that it makes judges or prosecutors less sympathetic to defendants when the police department is optimizing for higher conviction rates, eg. (That is, doing the opposite of the quote above of the public defender "humanizing" the drug case defendant in the plea deal example.) There will be a lot of reports transcribed, people will find ways to optimize the output to their gain and we've all seen how easy it is to bias llm output by promoting.