A terminal value (also known as an intrinsic value) is an ultimate goal, an end-in-itself.
In an AGI with a utility or reward function, the terminal value is the maximization of that function. The non-standard term "supergoal" is used for this concept in Eliezer Yudkowsky's earlier writings.
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. If a (simple) chess program tries to maximize piece value three turns into the future, that is an instrumental value to its terminal value of winning the game.
Some values may be called "terminal" merely in relation to an instrumental goal, yet themselves serve instrumentally towards a higher goal. The student described above 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.
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, health, friendship, social status, love, 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' terminal values are often mutually contradictory, inconsistent, and change over time.
Future artificial general intelligences may have the maximization of a utility function or of a reward function (reinforcement learning) as their terminal value. The function will likely be set by the AGI's designers.
Since people make tools instrumentally, to serve specific human values, the AI's assigned value system may be much simpler than humans'. This will pose a danger, as an AI must seek to protect all human values if a positive human future is to be achieved. The paperclip maximizer is a thought experiment about an artificial general intelligence with consequences disastrous to humanity, with the the apparently innocuous terminal value of maximizing the number of paperclips in its collection,
An intelligence can in principle work towards any terminal value, not just human-like ones. 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.
For an artificial general intelligence to have a positive and not a negative effect on humanity, its terminal value must be benevolent to humans. It must seek the maximization of the full set of human values (for the humans' benefit, not for itself).
Benevolence may arise even if not specified as an end-goal, is it is a common instrumental value for agents with a variety of terminal values. For example, humans often cooperate 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 moral shift (described by Immanuel Kant) in which benevolence changes from a merely instrumental value to a terminal one--they become altruistic and learn to value benevolence in its own right.
However, such shifts cannot be relied on to bring about benevolence in an AI. Benevolence as an instrumental value for an AI is relevant only when humans are at roughly equal power to the AI. If the AI is much more intelligent than humans, it will not care about rewards and punishments from humans.. Moreover, a sufficiently powerful AI is unlikely to undergo a Kantian shift, as any changes in one's goals, including replacement of terminal by instrumental values, generally reduces the likelihood of maximizing one's utility function (Fox & Shulman 2010; Omohundro 2008).
Eliezer Yudkowsky, Terminal Values and Instrumental Values