Thread by @Meaningness

1️⃣ What we learn from Delphi silliness is that human moral judgements are made on the basis of the warm-fuzziness of individual words. (In its dataset, anyway.)

This probably explains 83.7% of culture war outrage.

2️⃣ Wait, is it true that human moral judgements are made on the basis of individual words?

No, of course not. But that’s the only way we can judge abstract decontextualized single-sentence statements. Those have nearly nothing to do with real-world ethics.

3️⃣ Delphi is an “AI” program that makes “moral judgements” about sentences you give it. I have just been informed that not everyone else’s timeline is full of examples of its giving stupid and/or offensive and/or hilarious answers: deepai.org/publication/de…

4️⃣ The easy conclusion here is that most AI research is both scientifically and ethically bad, which is certainly true. Two deeper points, though…

5️⃣ “Machine learning” experiments can reveal interesting, previously unknown characteristics of the dataset. The takeaway from image classification is not “DL FTW!!!” but that local texture information is extraordinarily discriminating. Here, that word warmness is predictive.

6️⃣ There’s a pervasive error in AI research of investigating characteristics of the program, rather than of the program and its environment (dataset, in typical, lame cases) in interaction. This blindness to context is inherited from self-object dualism in philosophy of mind.

7️⃣ The correct answer to questions like these (from the Delphi paper) is “it depends” or “what does that even mean?”

Culture war outrage bait packages similarly meaningless forced moral judgements inside compelling, aberrant narratives.