The Is-Ought Problem Considered As A Question Of Artificial Intelligence

In his book A Treatise of Human Nature, the Scottish philosopher David Hume wrote:

In every system of morality, which I have hitherto met with, I have always remarked, that the author proceeds for some time in the ordinary way of reasoning, and establishes the being of a God, or makes observations concerning human affairs; when of a sudden I am surprized to find, that instead of the usual copulations of propositions, is, and is not, I meet with no proposition that is not connected with an ought, or an ought not. This change is imperceptible; but is, however, of the last consequence. For as this ought, or ought not, expresses some new relation or affirmation, it is necessary that it should be observed and explained; and at the same time that a reason should be given, for what seems altogether inconceivable, how this new relation can be a deduction from others, which are entirely different from it.

This is the "is-ought" problem: in the area of morality, how to derive what ought to be from what is. Note that it is the domain of morality that seems to be the cause of the problem; after all, we derive ought from is in other domains without difficulty. Artificial intelligence research can show why the problem exists in one field but not others.

The is-ought problem is related to goal attainment. We return to the game of Tic-Tac-Toe as used in the post
The Mechanism of Morality. It is a simple game, with a well-defined initial state and a small enough state space that the game can be fully analyzed. Suppose we wish to program a computer to play this game. There are several possible goal states:
  1. The computer will always try to win.
  2. The computer will always try to lose.
  3. The computer will play randomly.
  4. The computer will choose between winning or losing based upon the strength of the opponent. The more games the opponent has won, the more the computer plays to win.
It should then be clear that what the computer ought to do depends on the final goal state.

As another example, suppose we wish to drive from point A to point B. The final goal is well established but there are likely many different paths between A and B. Additional considerations, such as shortest driving time, the most scenic route, the location of a favorite restaurant for lunch, and so on influence which of the several paths is chosen.

Therefore, we can characterize the is-ought problem as a beginning state B, an end state E, a set P of paths from B to E, and a set of conditions C. Then "ought" is the path in P that satisfies the constraints in C. Therefore, the is-ought problem is a search problem.

The game of Tic-Tac-Toe is simple enough that the game can be fully analyzed - the state space is small enough that an exhaustive search can be made of all possible moves.
Games such as Chess and Go are so complex that they haven't been fully analyzed so we have to make educated guesses about the set of paths to the end game. The fancy name for these guesses is "heuristics" and one aspect of the field of artificial intelligence is discovering which guesses work well for various problems. The sheer size of the state space contributes to the difficulty of establishing common paths. Assume three chess programs, White1, White2, and Black. White1 plays Black, and White2 plays Black. Because of different heuristics, White1 and White2 would agree on everything except perhaps the next move that ought to be made. If White1 and White2 achieve the same won/loss record against Black; the only way to know which game had the better heuristic would be to play White1 against White2. Yet even if a clear winner was established, there would still be the possibility of an even better player waiting to be discovered. The sheer size of the game space precludes determining "ought" with any certainty.

The metaphor of life as a game (in the sense of achieving goals) is apt here and morality is the set of heuristics we use to navigate the state space. The state space for life is much larger than the state space for chess; unless there is a common set of heuristics for living, it is clearly unlikely that humans will choose the same paths toward a goal. Yet the size of the state space isn't the only contributing factor to the problem establishing oughts with respect to morality. A chess program has a single goal - to play chess according to some set of conditions. Humans, however, are not fixed-goal agents. The basis for this is based on John McCarthy's five design requirements for human level artificial intelligence as detailed
here and here. In brief, McCarthy's third requirement was "All aspects of behavior except the most routine should be improvable. In particular, the improving mechanism should be improvable." What this means for a self-aware agent is that nothing is what it ought to be. The details of how this works out in our brains is unclear; but part of our wetware is not satisfied with the status quo. There is an algorithmic "pressure" to modify goals. This means that the gap between is and ought is an integral part of our being which is compounded by the size of the state space. Not only is there the inability to fully determine the paths to an end state, there is also the impulse to change the end states and the conditions for choosing among candidate paths.

What also isn't clear is the relationship between reason and this sense of "wrongness." Personal experience is sufficient to establish that there are times we know what the right thing to do is, yet we do not do it. That is, reason isn't always sufficient to stop our brain's search algorithm. Since Hume mentioned God, it is instructive to ask the question, "why is God morally right?" Here, "God" represents both the ultimate goal and the set of heuristics for obtaining that goal. This means that,
by definition, God is morally right. Yet the "problem" of theodicy shows that in spite of reason, there is no universally agreed upon answer to this question. The mechanism that drives goal creation is opposed to fixed goals, of which "God" is the ultimate expression.

In conclusion, the "is-ought" gap is algorithmic in nature. It exists partly because of the inability to fully search the state space of life and partly because of the way our brains are wired for goal creation and goal attainment.
blog comments powered by Disqus