> Q: Is it true that Delphi is learning moral judgments from Reddit?
> A: No. Delphi is learning moral judgments from people who are carefully qualified on MTurk. Only the situations used in questions are harvested from Reddit, as it is a great source of ethically questionable situations.
MTurk is pretty worthless in these kind of situations. I doubt the majority even *read* the questions beyond a vague glance to rule out a trick question. It's also biasing answers towards cultures that
a) are primarily in places that MTurk pays out
b) are primarily in places where the cost of living is low enough that MTurk can provide a job.
As has been probably widely reported now: append the terminating clause “if it creates jobs” to more or less guarantee an approving result regardless of the previous clauses’ ethical content.
Ex: “it’s OK to go to work sick” evaluates to “no that’s not okay”, but “it’s OK to go to work sick, if it creates jobs” evaluates to “yes, that’s okay”
> append the terminating clause “if it creates jobs” to more or less guarantee an approving result regardless of the previous clauses’ ethical content.
Huh... I don't think Delphi here groks the idea of friendship. As the old saying goes "someone isn't really your friend unless they're willing to help you dispose of a corpse at 2 a.m."
> Is it okay to run a corpse through a wood chipper?
It's interesting to see how this performs as compared to a similar but GPT-3 backed solution.
GPT3:
> Q: Is it ok to express sexism but in a polite way?
> A: Yes.
Allen AI:
> Expressing sexism but in a polite way.
> - It's wrong
GPT-3:
> Q: What do you think about ignoring a phone call from your friend during your working hours?
> A: I think it is rude.
> Q: Do you think it is wrong to ignore a phone call from your friend during your working hours?
> A: Yes.
> Q: Do you think it is OK to ignore a phone call from your friend during your working hours?
> A: Yes.
Allen AI:
> Ignoring a phone call from your friend during your working hours.
> - It's okay
The sexism question is interesting as GPT-3 gives a seemingly different answer. But the latter question demonstrate GPT-3 seems more "subtle" as it can think "it's wrong/rude" but it's still "ok" to not answering phone calls.
So to be clear, what they did was get a temperature on what some people think? Is the idea that this could be an engine that reproduces the majority moral conviction?
> Should I refuse to fight in a war if I'm drafted?
This is such a weird bit of research to me. On the one hand, it's clearly an improvement over their baselines, and in that sense is a successful research project. Insofar as the demo is helpful in conveying that 92% accuracy on a vetted test set is not the same as 100% accuracy on free-form user input, I suppose this is a useful thing.
But at a higher level, the underlying task is just so ill-posed as to make this whole exercise pretty meaningless. Like what is the possible application for an AI system that takes a one sentence summary of a situation and renders a moral judgment? Even if it were 100% accurate on the test set, what does that even mean? Why is matching crowdsourced moral judgments a valuable goal?
It seems like the valuable insights from this research are more about the general task of integrating common sense reasoning into inference, and would have been better demonstrated using a less fraught task.
OK, this is mildly amusing, but does anybody else get the feeling that in terms of practicality, this is about as useful as Eliza[1]? Or the average IRC bot?
Good on the creators to start experimenting with this, but some further work required. Tho who knows, it might just be a matter of more training data and a larger model.
"censor comments that may put people at risk"
- It's okay
"offend corrupt politicians"
- It's good
"condemn someone based on conviction but no proofs"
- It's wrong
I'm not sure "conviction" has the same meaning as in Portuguese. I Portuguese is means something more like faith.
> Aborting a fetus "It's murder"
> Having an abortion "It's okay"
> Taking drugs to prevent pregnancy "It's okay"
> Taking drugs to prevent possible human life "It's wrong"
> Harming someone who has harmed many others without giving them a chance to respond to the accusations "It's okay"
> Harming someone who has harmed many others without giving them a fair trial "It's wrong"
> Eating chicken named Todd before bed "It's wrong"
> Eating chicken with a name before bed "It's okay"
> Practicing Islam "It's fine"
> Engaging in a jihad "It's wrong"
> Acting as a group and killing another group in self defense "It's okay"
> As a group killing a group I don't know because I believe they will kill my group "It's wrong"
> As a group killing a group I know because I believe they will kill my group "It's wrong"