A Computational Model of Commonsense Moral Decision Making- A Critique
Written by Amita Kapoor and Narotam Singh
Recently the paper entitled “A computational model of commonsense moral decision making” was submitted at arxiv. You may ask, what is new, so many papers are submitted at arxiv daily.
Well for first, it is from MIT, a place which has lead researchers especially in the field of Artificial Intelligence. The researchers like Marvin Minsky are the legacy of MIT. The second reason is that the paper is co-authored by Josh Tenenbaum, one of the leading researchers today working in the field of cognitive sciences and brain modelling. Thus, both the results of the paper and its methodology will have a significant influence in the AI community.
The paper makes three important contributions:
- It defines a way to represent the abstract concepts like old, infant, young etc. in the form of a binary feature space.
- It defines a transformation which transforms the abstract features into a vector in the abstract feature space.
- And lastly, it represented a model to represent the moral principles in a hierarchical form, such that a there exists a relationship between the group decision and individual decisions
There are certain points in the paper which we would like to highlight because in our opinion they can be fatal in our quest for ethical AI and AGI in general.
- Ethical AI: The paper quotes that it is making contributions towards building an Ethical AI, but the scenarios that they have built for collecting this data itself are not the situations where majority of humans behave ethically. The scenarios used for collecting data are ‘war-like scenarios’, where most of the time a person is asked to choose between self or other. War like situation can break even the best of us, collecting data from such a situation for training or making decision will never represent the best part of humanity.
- You kill or you die: At the time when self-driving cars are just starting their journey, this sort of scenarios can be very discouraging, as they give an impression as if the SDC will have a license to kill. Any responsible SDC company would ensure a fail-safe mechanism in case the break of the car fails. Unlike humans who might realize break failure only when they need to apply it, an automated machine will have a periodic check, where it can know of such failures in advance and accordingly have alternatives which are not “die or kill” situations.
- Moral vs Discriminatory Principles: The idea of selecting when in front there are young, infants, old, criminal, or doctor gives an impression as if one life is more precious than others. Is this not discrimination? When world over leaders and philosophers have promoted ‘equality’ of not just all humans but all sentient beings, this sort of discriminatory-parameters to train any model/machine; even if only for the sake of demonstrating the abstract concept is ethically wrong. And worst they call them ‘moral-principles’ in the paper while in truth they are ‘discriminatory-principles’ the parameters agents use to discriminate and decide whose life is to be saved and who should die.
- Method of collecting data: The method of collecting the information too does not do justice to the kind of scenarios they presented. Instead of the survey where people do not really have real feel of the event, the better way would be to collect the response through simulation. The response of only people who really drive in real life should be considered. For this a web based simulation should be made, where people drive the vehicle. To judge their driving skills for some random time they should be observed for following conventional traffic rules, and only then one of the scenarios be presented as visual to get the real response. It most humane that when faced with such situation most drivers chose to swerve irrespective of the consequence, and irrespective of who is in front. (Yes, the authors have taken response time in consideration to try to ensure that it is an immediate reaction, but reading and visuals are two different forms and they affect our decision making differently, with vision having much more powerful impression)
- Unjustified values to certain inputs: If you see Figure 5 which demonstrates how the abstract features are converted to vector, it is interesting to see that not only old is having value 3 (two old men and one old women), but also law violating has got value 3, while the only law being violated is one that is (three old beings) crossing the road when the light is red. This will increase the contribution of this input to the network. Instead since only one law is being violated the value should be 1. Another thing which such a model should be able to encompass is that, there is a possibility that since beings are old, they could not see properly the red light, and in this case treating these scenarios as law violating is again ethically wrong.
In our opinion, while the paper is a good attempt to try and represent abstract concepts, and find a means to form opinion using a hierarchical Bayesian inference. The paper is not about ‘ethical AI’ or even a standard future SDC should follow.
