Four colors (red, yellow, green, and blue) were arranged in all possible two-color sets to determine if goldfish can discriminate between color sets associated with shock and color sets associated with safety/shock omission in a one-phase (linear presentation) discrimination-learning procedure. The results showed that goldfish learned to discriminate between two-color sets if set colors were natural categorical color-code mates (red = yellow and green = blue). When natural code mates were not in the same set, and therefore were paired with different shock consequents, no discrimination learning occurred, suggesting that goldfish, unlike pigeons, are not able to code colors arbitrarily. The method also allowed a measure of preference between colors within sets associated with safety/shock omission. Original-learning preference measures between colors in sets so associated showed that goldfish chose red over any other color, yellow over blue or green, and green over blue, despite the fact that both colors in any set were procedurally identical, implying that goldfish do discriminate between colors in the absence of explicit discrimination training. The goldfish that failed to discriminate between redlblue and green/yellow sets in original learning were transferred to red/yellow and blue/green color sets. In transfer, the color paired with safety/shock omission in original learning was preferred over the color paired with shock in original learning, which resulted in a reversal of original-learning color preferences for half the goldfish. The transfer color-preference results imply that the goldfish had associated specific colors with specific shock consequents, but the associations were not robust enough to support discrimination learning in the face of categorical color coding.