3 ms·

That's part of the problem. People think that using their own human bias to alter the machine is a good thing, when it's clear that human bias leads to negative

That's part of the problem. People think that using their own human bias to alter the machine is a good thing, when it's clear that human bias leads to negative outcomes. You need a separate thing to remove the human input bias. Quality is a complex thing that requires design, safeguards, process, as well as iteration. You can't design test suites or checklists with just AI, because you don't know if the AI will decide to override them because of some other instruction or signal. And how do you determine the quality of the signal anyway? The quality of input of each human varies. It's "too adaptive" for QA. You would need a QA for the QA.

natch 3y ago
Human bias is a whole huge area apart from just quality, not sure we want to get into that. I mean is it human bias, or is it reality bias. e.g. Poland hasn't had any terrorist attacks (or very few?), is that human bias, or is it reality? I definitely don't treat its output with full trust, but I've been pleasantly surprised that even when I give it bad or incorrect guidance (unintentionally) it has caught my mistake, corrected me, and I've learned things. For the QA case, I suppose what you're getting at is that if it can't be fully trusted, you might get incorrect QA results -- false negatives, false positives -- I'd agree but I think you just have to find an effective way to use the tool. Perhaps the obvious way most people would want to use the tool, is not in fact the best way to use the tool. But just because the tool isn't delivering the perfection people hope for, doesn't mean there isn't some other way to use it that still catches (some) mistakes and adds value.