The U.S. Geological Survey (USGS) is investigating how the social networking site Twitter, a popular service for sending and receiving short, public text messages, can augment USGS earthquake response products and the delivery of hazard information. Rapid detection and qualitative assessment of shaking events are possible because people begin sending public Twitter messages (tweets) with in tens of seconds after feeling shaking. Here we present and evaluate an earthquake detection procedure that relies solely on Twitter data. A tweet-frequency time series constructed from tweets containing the word "earthquake" clearly shows large peaks correlated with the origin times of widely felt events. To identify possible earthquakes, we use a short-term-average, long-term-average algorithm. When tuned to a moderate sensitivity, the detector finds 48 globally-distributed earthquakes with only two false triggers in five months of data. The number of detections is small compared to the 5,175 earthquakes in the USGS global earthquake catalog for the same five-month time period, and no accurate location or magnitude can be assigned based on tweet data alone. However, Twitter earthquake detections are not without merit. The detections are generally caused by widely felt events that are of more immediate interest than those with no human impact. The detections are also fast; about 75% occur within two minutes of the origin time. This is considerably faster than seismographic detections in poorly instrumented regions of the world. The tweets triggering the detections also provided very short first-impression narratives from people who experienced the shaking.