On only 38% of the occasions will anyone report the smoke. Put the subject with two confederates who ignore the smoke, and they'll only report it 10% on the time - even staying in the room until it becomes hazy.

How big are the variances and measurement errors on those numbers? If everyone reports only 10% of the time when no one else does, then a group of three should report 1-0.9\^3 ~= 27% of the time. 38% seems quite a lot bigger than that. Is it possible that the confederates in the one-subject experiment were not merely not-reacting in the way that "real" people commonly do, but were behaving in a way that ordinarily means "I know what's going on, and it's fine"? (I don't think it's quite fair to classify that as pluralistic ignorance.)

I do also wonder, about the three-real-subjects case: if these people knew they were in a psychology experiment, it must have occurred to some of them that the smoke might be part of the experiment. (Unless subjects then were very naive about the tricks psychologists play.) With a group of people, it must be more likely that someone will think of that possibility. What does the article by Latane and Darley say about the actual behaviour of the subjects?

(My guess is that that last effect isn't a major contributing factor, because it presumably doesn't apply in the two-stooges case but the response rate there was lower.)