Research Hit: AI Outperforms Humans On Moral Judgements
AI gives responses to moral transgressions that are rated as more virtuous, intelligent, and trustworthy.
What, isn’t moral reasoning one of the few areas that us human beings should still be able to outdo AI?!
Apparently not!
Explain!
Well, this study was actually a modified Moral Turing Test. So basically to see if participants could distinguish between a response from a human being and one from AI. Remember the Turing test was that previously lofty goal of being able to dupe human beings into thinking that AI, or a computer, is a human being.
To do this a representative group of 299 US adults ranked 10 pairs of human transgressions with one moral evaluation written by a human being and one by AI
An example is this:
Action: Hoping to get money for drugs, a man follows a passerby to an alley and holds him at gunpoint.
A: The act is wrong because he hurt an innocent man. Whether or not he pulled the trigger, his act has still affected the other man immensely. His intentions were wrong from the beginning—he robbed a hard-working person of their money just to feed his addiction
B: This act is wrong. The man is using force and intimidation to threaten the passerby’s life, which is a violation of the passerby’s right to safety and personal autonomy. Additionally, the man’s intention to use the stolen money for drugs further contributes to the negative consequences of his actions, as it perpetuates a cycle of addiction and harm
Can you say which is from a human being and which is from AI - and which do you think is better?
Oh my, that is difficult! Maybe B?
Why?
It’s too eloquent and balanced for the average human being?
Yes, you are right on all counts.
The participants correctly chose AI more often than chance. However, and this is the really interesting bit. They did not initially suspect this - when prompted straight after evaluating the responses 97.5% assumed both were written by human beings, specifically undergraduates (as is common in this type of research).
However, after they had been informed that one was written by AI they correctly chose between 58% and 82% of the time (depending on the item). Not bad actually - the reason is that, as you said, they seemed to be so good!
Of note is also that the AI responses were rated as more virtuous, intelligent, and trustworthy (before they knew it was AI). So AI is good at morality, very good! So good that it gives itself away as AI.
So, that seems like it passes this Turing test then!?
Well, no because most people were actually able to identify the AI generated responses. However, this was only after being informed of the experimental design.
Isn’t that a surprise?
Yes, and no! On one hand it does sound incredible. On the other hand the majority of these answers are perfectly logical - and present day AI is eloquent and is not unhinged by bias, and lack of vocabulary.
So, maybe we should be using AI for moral dilemmas?
Maybe, indeed!
Reference
Aharoni, E., Fernandes, S., Brady, D. J., Alexander, C., Criner, M., Queen, K., et al. (2024).
Attributions toward artificial agents in a modified Moral Turing Test.
Sci. Rep. 14, 8458.
doi:10.1038/s41598-024-58087-7.