Researchers ultimately concluded that “a constructive strategy for identifying the violation of social norms is to focus on a limited set of social emotions signaling the violation,” namely guilt and shame. In other words, the scientists wanted to use AI to understand when a mobile user might be feeling bad about something they’ve done. To do this, they generated their own “synthetic data” via GPT-3, then leveraged zero-shot text classification to train predictive models that could “automatically identify social emotions” in that data. The hope, they say, is that this model of analysis can be pivoted to automatically scan text histories for signs of misbehavior.
Lemme get this straight: DARPA researches fabricated a series of words that signaled emotional states. And then, they, the DARPA researchers classified the series of words with the emotional states for the AI to train on (zero-shot classification). And then they hope to leverage the trained AI to identify "social emotions"?
Everything about this is fucking stupid.
The GPT-3 prompt could've been: "What are some sentences a shameful socialist/conservative/anarchist/terrorist/etc protestor/litterer/murderer/liar/etc might use?", implicitly connecting shame a particular ideology. As such, social emotions signals more emotions by their method of generation and classification.
Suddenly, some random person is being targeted for having fucked up and they're like, "Wtf did I do? Yes, I did shoplift from Target, but it was like a $20 shirt because my job at Wal-Mart makes me use food stamps to make ends meet. Fuck off!"
The AI automatically detects another violation of social norms.
And you're like, "That's an edge case...". Yeah, sure, but it's DARPA, we're talking about here. That should be enough said.
