Three experiments obtained evidence that spontaneous trait inferences
(STIs) occur on-line, at encoding. In each, participants read many sen
tences on a computer screen. After each paragraph, they indicated whet
her it included a test probe word. Paragraphs that imply but do not co
ntain traits should increase errors or reaction times (RTs) to trait p
robes. In Experiment 1, trait-implying paragraphs produced more errors
than control paragraphs, supporting the hypothesis. In Experiments 2
and 3, with feedback on each trial, longer RTs supported the hypothesi
s. STIs had the same effects as McKoon and Ratcliff's ''predicting inf
erences.'' Unexpectedly, participants gained control over STIs and pre
dicting inferences, so that RT differences (and error differences in E
xperiment 1) declined over trials. Analyses of reading times in Experi
ment 3 ruled out several alternative explanations. Results demonstrate
that social inferences can occur spontaneously at encoding and sugges
t that immediate feedback may make control possible.