Cheap talk
Costless, nonbinding words that may or may not be believed.
What it means
Communication that is costless, non-binding, and unverifiable, so the words themselves carry no direct payoff. Whether they transmit information depends on how far the sender's and receiver's interests align: when interests coincide, cheap talk can be fully credible; as they diverge, the information conveyed degrades gradually, with fully uninformative 'babbling' — every message ignored — as the limiting case rather than the norm.
The alignment condition, made precise
Crawford and Sobel's model gives the intuition a sharp shape. A sender who privately knows a state sends a message to a receiver who then acts; the sender's preferred action is biased above the receiver's by some amount that measures their conflict of interest. In equilibrium the sender does not reveal the exact state but only which interval of a partition it falls in. The smaller the conflict, the more intervals the message can sustain and the finer the information that crosses. As the conflict grows, the number of intervals falls until just one remains, and the message is ignored. Information is therefore not all-or-nothing: it degrades gradually as interests diverge, and alignment buys credibility without anyone paying a cost.
What the experiments show
In the laboratory, people communicate and believe more than the bleakest theory predicts, but less than full efficiency. Crawford's 1998 survey of these games finds that cheap talk raises coordination when interests are common, though subjects deviate from equilibrium in systematic, boundedly rational ways. Charness tested Aumann's conjecture that agreeing to play an efficient outcome conveys no information and so should not matter. Messages coordinated behavior strongly when the sender spoke before choosing an action, but not when the message followed it, so timing and the sender's own stake are decisive. Where interests conflict, transmission becomes coarse and self-serving, as the partition model predicts.
From cheap words to credible ones
Pure cheap talk is trusted only under alignment, but three bridges convert it into something believed. Repetition and reputation are the first: once breaking a promise forfeits future credibility, the words stop being payoff-irrelevant. A mediator or shared correlation device is the second, coordinating actions that unstructured talk cannot. Adding a cost or a binding penalty is the third, which moves the interaction out of cheap talk and into costly signaling or contract. This explains why a one-off announcement persuades less than a track record, and why regulators demand verifiable disclosure rather than statements of intent. The same logic drives the 'cheap-talk scripts' used in valuation surveys, which warn respondents about hypothetical bias before they state what they would pay.
Limits and open problems
A babbling equilibrium, in which every message is ignored, coexists with any informative one, so cheap talk creates a coordination problem of its own: which meaning will prevail? Theory offers refinements, such as Farrell's neologism-proofness and the notion of credible messages, to select the meaningful equilibrium when players share a language, but these do not always pin down a unique outcome and fit the data only partly. Cheap talk also differs sharply from Bayesian persuasion, where the sender commits in advance to an information policy; there it is commitment power, not free words, that does the work. And nothing in the framework rules out strategic silence or deliberate vagueness, which are often the equilibrium response when candid disclosure would hurt the speaker.
Examples
Before a coordination game, players announcing 'let's both go left' can succeed because neither gains by deceiving the other.
Two colleagues agreeing to meet at the main entrance costs nothing and works — neither gains by misleading the other. Make it about who gets the corner office and the words carry nothing.
A press release saying the company 'takes your privacy extremely seriously' is free to write and free to break. A published audit that would embarrass them if false is not.
A patient describing symptoms to their own doctor is trusted on their word: interests align, so lying gains nothing. The same self-report to a disability insurer, where interests diverge, triggers medical verification instead.
A government pledging to cut emissions 'by mid-century' persuades little on its own, since reversing a distant promise is costless. Legislating binding interim targets with penalties turns the words into a commitment.
First described in Crawford & Sobel (1982); Farrell & Rabin (1996).
Key references
- Charness, G. (2000). Self-serving cheap talk: A test of Aumann's conjecture. Games and Economic Behavior, 33(2), 177-194. doi.org/10.1006/game.1999.0776
- Crawford, V. P. (1998). A survey of experiments on communication via cheap talk. Journal of Economic Theory, 78(2), 286-298. doi.org/10.1006/jeth.1997.2359
- Farrell, J., & Rabin, M. (1996). Cheap talk. Journal of Economic Perspectives, 10(3), 103-118. doi.org/10.1257/jep.10.3.103
- Crawford, V. P., & Sobel, J. (1982). Strategic information transmission. Econometrica, 50(6), 1431-1451. doi.org/10.2307/1913390