Presentation Jan Broersen @ACAIS 12th edition, Nijmegen
Main question of safety: Is delegating responsibility to AI safe?
Underlying philosophical questions: Is AI possible? What is responsibility?
This general problem has various subproblems:
- Social: What are the effects on work, trust in institutions, problems with bias, problems with democracy?
- Safety (todays topic)
- Perception problem: due to antropomorphisation we can overestimate the responsibility of AI.
- Regulation: avoiding responsibility gaps.
- Contribution: who do we hold responsible?
- Engineering: how do we build responsible AI?
Sources of confusion about the AI & Responsibility debate
1. Between different disciplines
Different disciplines give very different answers to responsibility questions. E.g.:
- Legal: if we see a car as a "legal entity" we can make the car pay for itself. This is an incentive for the car to make better decisions and learn.
- Computer Science: Sure, we can implement a system where the car maintains its own type of money.
- Philosophy: no that's all nonsense, punishment and guilt must be "felt" in order to work.
2. Is AI or an artificial mind possible at all?
Various positions:
- Idealist: the question is nonsense, all there is, is mind.
- Would require a physical copy of someone
- Requires a functional copy of yourself.
- Requires a quantum computation copy of ourselves
- Requires deep learning on Turing machines
- GOFAI is enough
- Any Turing machine suffices
- Panpsychism
- etc.
3. There are various description levels of agency
- Causal: who caused it?
- Informational: Could the subject have known about the bad consequences?
- Motivational: Was the bad action intentionally?
These levels are also ways of attributing responsibility.
Responsibility can be understood both active and passive.
Passive: you did not cause it yourself, but could you have prevented it?
Broersen discusses these positions w.r.t. the "death in the desert" example. A traveler plans on walking in the desert, and brings a water bottle. One bad actor poisons the water. Coincidentally, another one empties the water bottle. The next day, the traveler dies from dehydration in the desert.
Who is responsible for the death? The man that poisoned the water and had intention to kill? (But the poison did not kill the traveler). Or the man that drained the water? (But he did in fact only remove poison, that would have killed the traveler anyways).
One answer: look at who caused the death. Halpern and Pearl hold the second bad actor responsible.
Aristoteles: responsibility is a) control over outcomes (causal) and b) knowing what you did when you did it (informational). In this view, the second bad actor has less control over the outcome, whereas the first one controls the fact that death will be the outcome.
Machine ethics: forwards or backward responsibility?
A backwards problem is applicable when something bad already happened and we ask who is responsible after the fact. E.g. after a crash with an Automomous Vehicle (AV).
The forward problem is more interesting.
Claim of Broersen: Ethics is about moral orderings from a "logical perspective" (Broersen has a background in logic and mathematics).
1) What is ordered? * Intentions * Actions * Character traits 2) Where does that come from? * Biological * Emotions * Common good etc. 3) How do we use moral orderings correctly?
What are relevant ethics for AI?
- Virtue ethics? Not so much.
- Duty ethics? Not really, application to AI not clear enough.
- Act consequentialism. A good fit, but what are the goals of the AI?
- Rule consequentialism: J.B. argues this is the way to go.