Where does rationality collide?

Rationality as a consistency requirement on choice: an agent is rational if its preferences obey a small set of axioms, so that its choices can be represented as maximizing the expected value of a utility function. Von Neumann and Morgenstern's 1944 Theory of Games and Economic Behavior proved the expected-utility representation for choices under risk, and Savage later extended it to choices under uncertainty. This is the homo economicus baseline. The referent is a normative standard of coherent preference, not a description of how real minds reason, not a philosopher's account of good justification, and not an engineer's specification for a machine.[1]

Unit

A chooser whose preferences satisfy the axioms and thus maximize expected utility

Recovery means

Not applicable

Measured by

Assessed by whether observed choices are consistent with the expected-utility axioms

Rationality as holding beliefs and taking actions that are adequately supported by reasons, in accordance with the norms of logic, evidence, and good justification. Philosophers separate theoretical rationality, which governs what it is reasonable to believe, from practical rationality, which governs what it is reasonable to do, and they distinguish instrumental reason, which fits means to ends, from substantive reason about the ends themselves. Audi's 2001 The Architecture of Reason develops rationality as grounding in reasons across both belief and action. The referent is an evaluative standard of well-supported reasoning, not a utility axiom, not measured behavior, and not a machine's decision rule.[2]

Unit

A belief or action, judged against norms of logic, evidence, and good justification

Recovery means

Not applicable

Measured by

Appraised by philosophical argument about whether the reasons adequately support the belief or the action

Rationality as a benchmark that actual human reasoning is measured against, and often found to miss. Tversky and Kahneman's 1974 Judgment under uncertainty: Heuristics and biases showed that people replace hard probabilistic questions with a few mental shortcuts, and that these shortcuts produce judgements that depart from the economic and logical norm in systematic, predictable ways. This is the human-reasoning strand, close in spirit to bounded rationality but focused on the observed gap rather than on the limits that cause it. The referent is an empirical description of how fallible people actually judge, not the ideal that economics assumes, not the philosopher's standard of justification, and not a machine.[3]

Unit

A person's actual judgement, together with the shortcut that produced it

Recovery means

Not applicable

Measured by

Measured in experiments that compare people's judgements against a normative benchmark to reveal systematic bias

Rationality as a design target for artificial agents: a rational agent is one that, for each possible situation, selects the action expected to maximize its performance measure given the evidence its percepts and prior knowledge provide. Russell and Norvig's Artificial Intelligence: A Modern Approach makes this rational-agent definition the organizing idea of the field. It shares the maximizing spirit of the economic baseline, but it is stated as a buildable specification for machines and turns on the agent's own knowledge and performance measure. The referent is an engineering criterion for what an agent should do, not a norm of human justification and not a report of how people actually reason.[4]

Unit

An agent whose chosen action maximizes the expected value of its performance measure given what it knows

Recovery means

Not applicable

Measured by

Judged by whether the agent selects the action expected to maximize its performance measure given its percepts