Solid read from Eileen Guo on AI-driven content moderation:

"Why the world is looking to ditch US AI models" (MIT Technology Review)

Having built text analysis (NLP/NLU) systems myself, I get why companies lean on ML/AI for this: scale.

But I also know about the shortcomings. And there are plenty.

The biggest sins of ML/AI-based content moderation are:

1/ Insufficient expert involvement during design + R&D. (Including: ignoring the expert when they tell you that ML/AI won't do a good job.)

2/ Insufficient human oversight. (The model will be wrong now and then.)

3/ Treating a model as a one-and-done. (They require constant upkeep because the goalposts keep moving.)