A terminal value is an ultimate goal, an end-in-itself. In an AI with a utility or reward function, the terminal value is the maximization of that function.
In Eliezer Yudkowsky's earlier writings, the non-standard term "supergoal" is used instead.
Terminal values vs. instrumental
Terminal values stand in contrast to instrumental values, which are means-to-an-end, mere tools in achieving terminal values. For example, if a given university student does not enjoy studying but is doing so merely as a professional qualification, his terminal value is getting a job, while getting good grades is an instrument to that end.
Some values may be called "terminal" merely in relation to an instrumental goal, yet themselves serve instrumentally towards a higher goal. In the previous example, the student may want the job to gain social status and money; if he could get prestige and money without working he would; and in this case the job is instrumental to these other values. However, in considering future AI, the phrase "terminal value" is generally used only for the top level of the goal hierarchy: the true ultimate goals of a system, those which do not serve any higher value.
Human terminal values
Humans system of terminal values is quite complex. The values were forged by evolution in the ancestral environment to maximize inclusive genetic fitness. These values include survival life, health, friendship, social status, love of various kinds, joy, aesthetic pleasure, curiosity, and much more. Evolution's implicit goal is inclusive genetic fitness, but humans do not have inclusive genetic fitness as a goal. Rather, these values, which were *instrumental* to inclusive genetic fitness have become humans' terminal values (an example of subgoal stomp).
Humans cannot fully introspect their terminal values. Humans' values are often mutually contradictory and change over time.
Non-human terminal values
Future artificial general intelligences are may have the maximization of a utility function or of a reward function (reinforcement learning) set by their designers as their terminal value.
The paperclip maximizer is a thought experiment about an artificial general intelligence assigned the apparently innocuous terminal value of maximizing the number of paperclips in its collection, with consequences disastrous to humanity.
AIXI is a mathematical formalism for modeling intelligence. It illustrates that the arbitrariness of terminal values may be optimized by an intelligence: AIXI is provably more intelligent than any other agent for *any* computable reward function.
In a Friendly AI
For an artificial general intelligence to have a positive and not a negative effect on humanity, its terminal value must be benevolent to humans, i.e., the maximization of human values (for the humans, not for itself).
Benevolence is a common instrumental value for agents with a variety of terminal values, and thus may arise even if not specified as an end-goal. For example, humans usually cooperate not out of pure altruism, but because they expect either an immediate benefit in response; or because they want to establish a reputation that may engender future cooperation; or because they have live in a human society that rewards cooperation and punishes misbehavior. Humans sometimes undergo a Kantian shift in which benevolence changes from a merely instrumental value to a terminal one--they learn to value benevolence in its own right.
However, these processes cannot be relied on to produce benevolence in an AI. Benevolence as an instrumental value is relevant only when humans are at roughly equal power to the AI. If the AI is much more intelligent than humans, their rewards and punishments will mean nothing to it. Moreover, an AI is unlikely to undergo a Kantian shift, as preservation of ones' own goals is a valuable characteristic for almost goal-seeking AI (Fox & Shulman 2010).
Eliezer Yudkowsky, Terminal Values and Instrumental Values
Joshua Fox and Carl Shulman (2010), "Superintelligence does not imply benevolence", Proceedings of the VIII European Conference on Computing and Philosophy, Oct, 2010. Ed. Klaus Mainzer. (Munich: Verlag Dr. Hut), pp. 456-461