Chatting (GPI) Without Cheating

AI as a tool for translators

For the past few years, I have been using Google Translate as a tool for checking the translations I make from languages I know reasonably well. That is, after Englishing a paragraph the old-fashioned way, I put it into the program to catch any obvious mistakes I may have made. (At the very least, the program caught words from seventy-nine word German sentences that had been, quite literally, “lost in translation.”)

Recently, while making a particularly challenging translation, I tried using ChatGPI for the same purpose. I quickly discovered, however, that direct translations provided by ChatCPI were exactly the same as those generated by Google Translate.

This lead to a second experiment. Rather than asking ChatGPI to provide a single translation of the passage in question, I asked it to provide three alternate translations. This yielded a trio of similar texts. While none of these met my standards, reading three different renditions of the text helped me think of additional possibilities. In other words, ChatCPI enabled me to conduct a one-man brainstorming session.

It’s worth noting in this respect that asking ChatCPI for more than three translations led to a great deal of repetition. That is, when I asked for five translations of a passage, ChatCPI provided fewer than five possibilities for many distinct parts of the text.

For example, when I asked for five different translations of the following passage, Chat CPI rendered but two translations of the words printed in bold text. (These were “Silver sparks dispersed” and “Silver sparks scattered.”)

Es war dunkel im Räume, den ein sanftes Schlingern wiegte, ein feines Beben erschütterte. In seiner Höhe kreiste ein Lichtspiel von Linien. Silberne Funken zerstreuten sich, blinkend und zitternd, um sich tastend wiederzufinden und zu Wellen zu vereinigen.