As Nobuyoshi and Ellis point out, however, their conclusions are based on a very small sample size. In addition, they are based on a very low number of obligatory occasions. E1, who showed the clearest gains, went from 4 correct out of 13 at time 1 to 8 correct out of 9 at time 2. E2 went from 9 correct out of 20 at time 1 to 16 correct out of 26 at time 2. E1's gains are statistically significant (Fisher Exact Test, 2 tail, p = .0115) but E2's gains are not (chi square = .69). Thus, for one subjects there was no evidence of the value of comprehensible output (E3), and for another, gains were not statistically significant. Data supporting a central hypothesis should be made of sterner stuff.

Note also that all three subjects had studied the past tense rule, and had been clearly focused on it in session 1. It is reasonable to expect that when subjects are focused on form, then put back in the same environment, they will be focused on form again, especially if the conversational partner is their teacher. The significant effect on E1, in other words, may have been a performance effect - E1 was simply more inclined to try to use a consciously learned rule for the past tense and was a more successful Monitor user than E2 or E3.

Van den Brandon's subjects (Van den Brandon, 1997, discussed earlier) who participated in sessions that encouraged negotiation of meaning increased their output relative to a control group that did not engage in interaction, but were not superior in grammatical accuracy. Each subject, however, only had seven to nine minutes of interaction.

Tarone and Liu (1995) suggest that CO may have played a role in the second language development of Liu's subject "Bob." Bob was recorded interacting with peers, with teachers, and with an "adult-friend" (Liu). Tarone and Liu note that language use was much more complex in the latter interactions, and, in general, "new structures appear first in interactions between Bob and the researcher, spread to the interactions with his peers, and appear last in his interaction with his teacher" (p. 119). They note that it is likely that Liu provided Bob with more complex input, but also suggest that Bob's attempts to produce CO in interacting with Liu played a role. While interacting with Liu, Bob used English in a much wider range of speech acts than in the other situations, and this may have pushed Bob to "go beyond the limits of his interlanguage competence" in production" (p. 121). Tarone and Liu show that the CO hypothesis, as well as the Input Hypothesis, is consistent with what is known about Bob's development. As they note (p. 123), data is lacking on the frequency of CO, which prevents us from determining whether CO resulted in language development and whether Bob produced significant quantities of CO.

The DISCOMFORT OF CO

The CO hypothesis predicts that we acquire language when there is a communicative breakdown and we are "pushed to use alternative means to get across .. the message ... precisely, coherently, and appropriately" (Swain, 1985, pp. 248-249). In addition to the research that shows that CO is an unlikely candidate, there is additional evidence that "pushing" students to speak is unpleasant for them. When asked what aspects of foreign language classes are the most anxiety-provoking, students put "talking" at the top of the list (Young, 1990). Laughrin-Sacco (1992) reported that for students in beginning French classes, "for nearly every student ... speaking was the highest anxiety-causing activity" (p. 314).

4