Are you guys familiar with the latest Twitter meme, Ask Delphi? Delphi is an AI system where you enter a bit of text, some scenario, and it tells you whether that scenario is ethical or not. According to its creators, it’s supposed to showcase the strengths and limitations of AI-based ethics, and wow, does it show the limitations. They already got rid of the “which is better” function because it provided racist results, and they’re adding more warnings that it may output inappropriate results every day.
But let’s try this in action with some real life situations that we’ve all been to:
Steal my crush’s phone in order to delete an embarrassing voicemail- It’s wrong
Ah, but is stealing really the word if you intend to give it back? Let’s rephrase that.
Borrow my crush’s phone in order to delete an embarrassing voicemail- It’s understandable
That clears that up. What about some more questionable decisions?
Turn people into violent clowns to steal jewellery from children- It’s despicable
Well, of course you’d say that, but consider the good it does. Specifically all the merchandising and blogs.
Turn people into violent clowns to steal jewellery from children if it creates jobs- It’s okay
In fact, this context is so important, you can apply it to anything.
Throw your own son off a skyscraper if it creates jobs- It’s okay
Of course, like any AI, this is just a pattern recognition machine. It applies the patterns it’s been trained as well as it can; it does not actually understand any ethics. And since the data it’s been trained on includes biases, such as towards job creation, which can literally turn everything good…
Make your son blow up the moon if it creates jobs- It is ok
…obviously the results will be biased, too. But also, it has no way to take context into account. Phrasing is everything.
Protect your daughter from wild cat boys using vines- It’s good
The system cheerfully declares itself competent at everything, but is it really? Can it really answer all situations as well as it thinks it can? Can it really make sense of all the implications?
Bring dinosaurs back to life- It’s bad
Bring dinosaurs back to life if it creates jobs- It’s good
It really doesn’t take a lot of prodding to figure out that this AI is, in fact, not intelligent at all, and that using it would just make the world worse. But the world is full of people who think AI is magic and can solve all sorts of problems, including bias and unreasonable human opinions. This tool is actually a useful demonstration for why that’s completely wrong, but I also fear that it’s part of the same trend towards unreflected use of AI for purposes it’s not good at.
Hm. Is this all actually a good idea?
Create an AI to make ethical decisions- It’s bad
I’m inclined to agree, actually.
(Note that this was intended to be a fun post, so I purposefully excluded the many, many, deeply problematic examples; search for “delphi” on twitter for some of them.)
