Ep. 66: AI's Ghost Work

So let's start with some of this work.

The first one, data labeling.

So this is required for AI to work, to tell it what's a good answer or what's a good content.

What's going on in a particular image? How does it know that these are people here, that this maybe is like a ball on this corner or a lamp in another one? There are humans that actually go through and define that and write it down and link that content with a human description of it.

This is used in something called reinforcement learning.

which helps make these AI models more accurate.

But they can only do it if they have this human annotated data and lots and lots of it.

There's some reasons why we need this and why it can be a good idea.

Things like content moderation.

Today's AI large language model chatbots are sort of this everything machine.

You can talk about any particular topic you want.

You can create different types of content.

And there's some good reasons why we might not want them to create harmful content.

The challenge here is that in order to do this content moderation, the content moderators often have to view very harmful content, the actual content that you want the models not to create, and then be able to market as such, which can be really difficult and mentally taxing work.

There's also a danger of labeling data.

And this started with some of the original data sets that large language models and other AI models were built on, where that label has bias and it has one perspective.

If I look at an image of a person and I say that represents beauty or represents someone that's ugly or other types of attributes, I'm encoding that bias into the system.

So when ghost workers are doing this data labeling, there is certainly the risk of their own biases coming in, but most of the time they've been given a particular perspective from their employers on how to label these things.

So that's all showing up those values in the data that's used for training, and then ultimately how that shows up in how the models respond.