In 2004, Eliezer Yudkowsky put forward Coherent Extrapolated Volition (CEV) as a framework for friendly ASI to predict what we would want "if we knew more, thought faster, were more the people we wished we were, had grown up farther together", and act on it.
In this post, I present Directly Ascertained Volition (DAV) with an action plan as a development of CEV that allows for a more granular approach tailored to the values and needs of each specific individual.
Take it personally
Leaving aside for a moment the technical challenge of alignment, I believe CEV can be improved in two main areas:
idealised-me versus actual-me: in CEV, the friendly ASI is asked to predict what an idealised version of me would want. However, my values and desires are likely to differ at least somewhat from those of idealised-me. For example, idealised-me may well come up with a galaxy-brained argument for dumping my family and spending eternity in meditation, but actual-me would not want that at all. Following CEV would therefore cause real harm to actual-me's values and desires. Since I can't foresee how and where idealised-me's values would diverge from actual-me's, it's not in my interest to let idealised-me make the choice on my behalf.
majority choice versus individual autonomy: in CEV, the friendly AI extrapolates value for humanity as a whole. However, different cultures, even different individuals, don't share the exact same values. Many people would be unhappy if radically different (yet still human) values were imposed on them. This should be avoided by directly ascertaining the volition of each individual and acting on it to the extent possible.
Enter Directly Ascertained Volition
About 70% of the world population owns a smartphone or similar device. The percentage of households with at least one smart device is even higher and gets us pretty close to universal coverage. Even with current technology, it would be trivially easy for friendly ASI to communicate with the vast majority of people on Earth on an ongoing basis. Consequently, it would be trivially easy for friendly ASI to keep tabs on what each individual on the planet wants at any given time, and then deliver it if it were so inclined.
The ASI need only ask, "What can I do for you today?". Requests whose scope is limited to the individual in question can be granted straight away: e.g. "Dye my hair blonde", "Teach me fluent German", "Build me an immersive virtual reality scene of the Neolithic".
Requests involving third parties must always take in consideration the autonomy of these other people: e.g. if Peter asks the ASI to make Sally his girlfriend, the ASI should not grant the request unless Sally already wants to be Peter's girlfriend in the first place. Similarly, the ASI should not take other people's stuff to give to Peter unless those people were already planning to give it to Peter to begin with.
Requests that involve allocating finite resources should be granted up to the available budget. E.g. if someone asks the ASI to make them a millionaire, the ASI should only disburse funds up to the current per capita budget.
Requests involving a dispute should be mediated by the ASI as a neutral arbiter, according to its best and most impartial interpretation of human law and precedent.
Finally, children and people who have dementia or are in a vegetative state could have their parents or legal guardian make these choices on their behalf until they reached the age of majority or recovered from their condition and were able to decide for themselves.
I would like to emphasise that deferring to your idealised-me, as per CEV, would still be an option. It's just that in DAV, this would be a purely individual choice, without bearing on those who would rather let their actual-me choose.
Pre-explained action plans to guard against unintended consequences
If individuals are going to be submitting their requests directly to the ASI, it's essential for them to understand what these requests will lead to. For example, someone might ask the ASI to "Just make me happy forever", expecting to be made happy in the traditional sense (friends, family, health, money...), only for the ASI to hook them up to a wireheading machine until the end of time, which is not what the person would have wanted.
In DAV, the ASI should therefore lay out its concrete action plan before acting on any request, and it should only launch the plan once the person has given their informed consent. In the above case, once the ASI had explained that it was going to wirehead the person, he or she could go "Heavens, no" and ask it to come up with a different plan. The ASI would then iterate on its action plans, taking into account feedback from the person, until the latter is satisfied enough to give it the go-ahead.
The bottom line
The ubiquity of smart devices makes it possible to move from a one-size-fits-all CEV framework to a tailored DAV framework that respects the autonomy of each individual. In DAV, idealised-me doesn't get to impose his values on actual-me, and the majority doesn't get to impose its existential choices on the minority.
In 2004, Eliezer Yudkowsky put forward Coherent Extrapolated Volition (CEV) as a framework for friendly ASI to predict what we would want "if we knew more, thought faster, were more the people we wished we were, had grown up farther together", and act on it.
In this post, I present Directly Ascertained Volition (DAV) with an action plan as a development of CEV that allows for a more granular approach tailored to the values and needs of each specific individual.
Take it personally
Leaving aside for a moment the technical challenge of alignment, I believe CEV can be improved in two main areas:
Enter Directly Ascertained Volition
About 70% of the world population owns a smartphone or similar device. The percentage of households with at least one smart device is even higher and gets us pretty close to universal coverage. Even with current technology, it would be trivially easy for friendly ASI to communicate with the vast majority of people on Earth on an ongoing basis. Consequently, it would be trivially easy for friendly ASI to keep tabs on what each individual on the planet wants at any given time, and then deliver it if it were so inclined.
The ASI need only ask, "What can I do for you today?". Requests whose scope is limited to the individual in question can be granted straight away: e.g. "Dye my hair blonde", "Teach me fluent German", "Build me an immersive virtual reality scene of the Neolithic".
Requests involving third parties must always take in consideration the autonomy of these other people: e.g. if Peter asks the ASI to make Sally his girlfriend, the ASI should not grant the request unless Sally already wants to be Peter's girlfriend in the first place. Similarly, the ASI should not take other people's stuff to give to Peter unless those people were already planning to give it to Peter to begin with.
Requests that involve allocating finite resources should be granted up to the available budget. E.g. if someone asks the ASI to make them a millionaire, the ASI should only disburse funds up to the current per capita budget.
Requests involving a dispute should be mediated by the ASI as a neutral arbiter, according to its best and most impartial interpretation of human law and precedent.
Finally, children and people who have dementia or are in a vegetative state could have their parents or legal guardian make these choices on their behalf until they reached the age of majority or recovered from their condition and were able to decide for themselves.
I would like to emphasise that deferring to your idealised-me, as per CEV, would still be an option. It's just that in DAV, this would be a purely individual choice, without bearing on those who would rather let their actual-me choose.
Pre-explained action plans to guard against unintended consequences
If individuals are going to be submitting their requests directly to the ASI, it's essential for them to understand what these requests will lead to. For example, someone might ask the ASI to "Just make me happy forever", expecting to be made happy in the traditional sense (friends, family, health, money...), only for the ASI to hook them up to a wireheading machine until the end of time, which is not what the person would have wanted.
In DAV, the ASI should therefore lay out its concrete action plan before acting on any request, and it should only launch the plan once the person has given their informed consent. In the above case, once the ASI had explained that it was going to wirehead the person, he or she could go "Heavens, no" and ask it to come up with a different plan. The ASI would then iterate on its action plans, taking into account feedback from the person, until the latter is satisfied enough to give it the go-ahead.
The bottom line
The ubiquity of smart devices makes it possible to move from a one-size-fits-all CEV framework to a tailored DAV framework that respects the autonomy of each individual. In DAV, idealised-me doesn't get to impose his values on actual-me, and the majority doesn't get to impose its existential choices on the minority.