ChatGPT4-Questions
I'm starting this page as personal notes (and anyone can contribute to it too) about a taxonomy of all unaddressed questions and dangers with AI.
Interface questions
I define interface as the machine-to-human point of interaction and context of the interaction. For example, if an AI serves a political group, its interface to human society is the political group. If an AI has opened access to the public on the web, then its interface is an open web access point. If an AI is providing counsels to a trader in the financial sector, then its interface is the single trader it is counseling.
What are the planned and legitimate interfaces for AI into human societies?
Are interface limits being set for any given project, or can the AI access unplanned interfaces?
Are there certain interfaces that fundamentally change the dangers and risks to be expected from AIs?
Are interfaces leaky, i.e., where the AI could have downstream effects into society further than with the humans it is directly interacting with?
Are risks coming from leaky interfaces fundamentally different? What are their specific characteristics?
Evolutionary problems
Can an AI favor, through any of its outputs, the evolutionary fitness of a part of humanity over another?
Is a given AI capable of initiating a phenotypic revolution?
Political problems
If an AI is given axiomatic moral principles, how does it approach moral disagreements within humanity about these principles. For instance, if an AI is asked to maximize the saving of human life in a particular problem, how does it approach the problem of saving human lives of people who differ by age, if the AI has not been given axiomatic moral principles indicating preferences based on age?
If an AI is given axiomatic moral principles, are these principles subject to disagreement within society?
Epistemological problems
How should an AI approach the problem of truths on which not all humans agree?
Is a given AI always truthful in its responses or can it hide objectives or truths to its interface?
Are there mechanisms of self-deception that are present in humanity and not in AI that could lead an AI to radically different conclusions than a human would faced with the same facts? Should such self-deception mechanisms be implemented in AI?