> YOUR ECONOMISTS ARE TOO CLOSE TO THE PROBLEM TO SOLVE IT, the yogurt said. ANY HUMAN IS. I have heard people express similar ideas about AI. That it is someh

> YOUR ECONOMISTS ARE TOO CLOSE TO THE PROBLEM TO SOLVE IT, the yogurt said. ANY HUMAN IS. I have heard people express similar ideas about AI. That it is somehow objective and correct, free from human folly. These people often object strongly to installing guardrails, on the grounds that the AI knows the truth better than us - almost a suggestion that we align to it. This is problematic for a number of reasons. But to pick just one, it makes you vulnerable to manipulation through what I call "the unsupervised loophole," whereby you are able to get the machine to give you the results you're after without labeling the data. Instead, you shape the training set such that you get the result you want anyway. Eg, you could implement an illegal redlining scheme by neglecting to balance your dataset beforehand, allowing systemic biases to influence your training set. This is an rookie mistake (you can tell because I'm an ML rookie and I've already figure it out the hard way) but might give you enough plausible deniability to survive a lawsuit.