GarciaMarzaCalvo2024MoralLearningByAlgorithms
Domingo García-Marzá and Patrici Calvo (2024) Moral Learning by Algorithms: The Possibility of Developing Morally Intelligent Technology
Bibliographic info
García-Marzá, D., Calvo, P. (2024). Moral Learning by Algorithms: The Possibility of Developing Morally Intelligent Technology. In: Algorithmic Democracy. Philosophy and Politics - Critical Explorations, vol 29. Springer, Cham. https://doi.org/10.1007/978-3-031-53015-9_6
Commentary
In this article, García-Marzá and Calvo explore the question of whether AI is able to learn morality. In other words: moral machine learning. They do this by describing some history and the way AI works when it comes to morality, and distinguishing two possible directions for this field: ethics of AI-creating external frameworks to ensure responsible AI use-and algorithmic ethics-trying to embed moral competences directly into AI systems through different methods. Furthermore, they describe different technical ways to give AI a sense of morality, pointing out both the potential, as well as the possible problems and dangers with each method.
Beyond the technical side, they show that AI still lacks some essential human abilities like moral judgment, critique, and emotional motivation. We could maybe train a model to make decisions that correspond to our own moral rules, but could you call a model like that moral? Finally, they discuss how governments, organizations, and institutions are trying to set up ethical governance systems to guide AI in a responsible way, highlighting the importance of transparency, accountability, and involving the people affected.
In my opinion, there is a lot of hype in the world of AI. CEO's of big AI companies promise us the world, while of course they benefit enormously from this increase in public interest as well. The authors of this article, however, are in my opinion quite critical about the capabilities of AI in the field of morality. It almost reads like a warning of some kind. This approach, I think, is very much justified, especially given the increasing use of AI in government decision-making in the United States for example. It is important that we do not assign human traits on these models when they are simply following algorithms without understanding, reasoning, or moral awareness.
Excerpts & Key Quotes
Projecting Human Traits onto Machines
- Page 104:
...the discipline of Artificial Intelligence has, from its beginnings, established a process of appropriating and assimilating extra-technological—biological, social and moral—terminology to recreate a symbolic imaginary capable of anointing, or tokenizing, machines endowed with mathematical models with a kind of naturalization and humanization, or even supra-humanization.
Comment:
The authors describe a common problem that occurs when humans interact with something. We are almost wired to project an intention, idea, belief or emotion onto something, even when it is not in fact there. To name an example that is also mentioned by the authors, when we think that a self-driving car feels empathy for us, we might trust it way more than we perhaps should. This concept is therefore, in my opinion, very important in moral machine learning, since misusing this could harm people on a larger scale. And moreover, thinking about this problem could help us design AI systems in a better way, so that they won't encouraging false projections of human traits.
Training for morality?
- Page 114:
A common feature of all these different approaches to the acquisition and learning of values by AI applications is that they are strictly conventional and framed by strategic interests. They are based on the uncritical observation of the patterns, opinions and behaviour belonging to a particular community or else a programmer’s view of what is fair and good for society.
Comment:
Here, the authors describe a large problem when we try to teach AI moral values through various approaches: bias. Almost all data produced by humans is biased, as humans are biased themselves. When we use human data to teach AI about morality, therefore, I think that there are 2 difficulties. The first one, which is also sort of touched upon by the authors, is the fact that we might reinforce existing societal biases in these models, rather than achieve objective ethical reasoning. The second is that training an AI on data from the past can potentially lock it into the moral values of that time. So what would happen if society’s views change in the future?
Optimizing for morality?
- Page 116:
Many of the current proposals therefore create a kind of substitute for artificial moral judgements based on a goals-focused utility function, rather than adhering to principles, values or norms whose consequences could be accepted by all those affected.
Comment:
In my opinion, the authors underline one of the most important critics on moral machine learning. I think that morality is not something that can be obtained by minimizing a loss function or achieving a high accuracy on a test set. Morality needs the ability to think and rethink about decisions. A moral agent should know or feel what to do in a completely new sitiuation it has never seen before. This is something I don't think AI will be capable of in the near future.
GarciaMarzaCalvo2024-SecondAgeOfArtificialIntelligence
Floridi2018AI4People#The power to decide