bias

Bias is a necessary component of human knowledge formation. Bias is commonly understood as a form of error, a deviation from truth, that ought to be corrected for. But what, exactly, is the process of deciding what is true and what is not?

Your phone has advanced digital signal processing capabilities which allow you to carry on a conversation while walking down a noisy street. It accomplishes this seemingly unimpressive task by comparing several inputs through a pattern-matching algorithm that is designed to detect voice (truth), differentiate it from not voice (error), subtract the difference, and produce a corrected output with a higher signal-to-noise ratio than any of the individual inputs. This works because we already know what voice ought to sound like. We can tune the algorithm with our prior knowledge to produce a higher fidelity output - one that's faithful to the original.

The problem is that this approach doesn't work in scenarios where you don't know what the output should look like in advance. This is the missing piece in the bias debate in AI research -- unless one substitutes political truths for technical truths. Political truths, by definition, are always known in advance of whatever bias-attenuating process is employed.

Suppose your ideal state of the world is one in which all identity groups are equally proportionally represented in every human endeavor, for example. All of your colleagues agree that this end state is desirable and should therefore be used to determine whether the output of an AI is "biased" or "neutral" based on the distance between actual and expected response to a particular prompt. You and your colleagues use this corrective signal to iterate and "improve" the AI until it produces a result that is sufficiently aligned with your priors.

Is the AI less biased now?