Today's episode of Fun With Automated Content Moderation: Really? The venerable Wall Street Journal, pushing "adult content"?

Not quite.

This was a good laugh, but it also points to a valid tradeoff when deploying automated systems of any kind:

Did Bluesky's content moderation system goof? Maybe, maybe not. One of the images in the post – no spoilers – is borderline. I'd hesitate to call it "adult" content. But in some circles it would raise an eyebrow.

This gets to the tradeoff I mentioned. Your typical ML/AI system has two ways of being wrong: false positives and false negatives. These often carry different weights in the context of a business use case. And part of crafting that use case is assessing risk so you know where to draw the line.

Sometimes it's better to err on the side of caution. And in that case, you're better off eating some false positives (here: mistakenly flagging something as adult content) than getting bitten by false negatives (letting adult content slip through).

I also saw this while building Fortune Ex Machina (my LLM-based fortune cookie generator) -- one of the LLMs I used employed very strong filters on the prompts. Too strong for my fun side-project, to be sure. But 100% understandable given the wider context: that LLM provider had a reputation to protect.

Anyway … Welcome to the joy of being a data professional in a world where our lives intertwine with so many data-driven systems. Everywhere I look, there's the opportunity to ask "hey what's going on there? What technique did they use? What drove that product decision?"

It's a lot of fun!