My reasoning was this: If the conversers found it very difficult to talk with delay in the circuit, they would surely push the button soon after the delay was introduced. On the other hand, any time when they continued to converse without pushing the button—while delay was in the circuit—was obviously also a time when the delay did _not_ make conversation impossible. So we had at least one measurable quantity—the time taken to detect delay—which we could interpret as an answer to our question. Note that we could tell if people pushed the button “just to be on the safe side” by seeing how often they did this when there actually was no delay in the circuit. There are just a few more necessary details before we discuss the results: —I recorded the conversations for later analysis, but the conversers knew that the recordings would be held in confidence. —The amount of delay used was 1.2 seconds, which represents the total round-trip delay for a circuit using two satellites, including the delay in typical end connections on the ground. We used this much delay because our preliminary tests indicated that it would be more likely to produce an effect than would the 0.6 second delay in a one-satellite link. —The entire 1.2 seconds of delay was put into one of the lines, since we had discovered that, where there is no echo, conversers cannot tell the difference between a delay of 2_t_ seconds in one line and a delay of _t_ seconds in each of two lines. I did this for the sake of convenience, so that I could introduce delay in the quiet line while the other one was active. —After someone detected delay, I removed it immediately and then waited at least a minute before putting it back in. —Altogether, I collected about two hours of conversation and introduced delay 22 times. The Results and What They Mean Now we could answer the question, “How long _does_ it take people to detect 1.2 seconds of delay?” As you can see from _the table opposite_, the times ranged all the way from 20 seconds to over 10 minutes, and, in two cases delay was not detected at all. The results in the table are also shown in _the histogram on the next page_, which depicts how broadly the detection times were distributed. To me, one of the most interesting things is that even people who were able to detect delay quickly sometimes did _not_ detect it for a couple of minutes. For example, the pair K/G had two times under a minute, one of 143 seconds, and one of 421 seconds. I interpret their two short times to mean that they knew what to look for, since they made no incorrect responses while delay was not present. However, their long times seem to mean that they sometimes didn’t notice delay for quite a while. Incidentally, only two responses were made during the total of about 40 minutes when I did not introduce delay, and these “false alarms” were by two of the fastest pairs at true detection—F/K and S/H.