The proposal of bigram frequencies playing a role in reading can be traced back to the late 60s, and has recently received support from neuroimaging studies. But despite 50 years of research, the behavioral literature shows conflicting results. Here we present a new paradigm that by using an artificial script, recreates the process of learning word forms from scratch and attains a tight control of confounding factors that otherwise arise due to our lifelong experience with reading. By orthogonally manipulating word and bigram frequencies, we show how bigram frequencies are automatically extracted during a word learning task.