Is ChatGPT Cooking Your Books?

If you’re like me, you might assume that ChatGPT is better at numbers than we are. After all, it’s an AI—crunching data and processing information should be its bread and butter, right?

Well, if you’ve been placing blind trust in ChatGPT for translating large numbers, I have some bad news for you. It turns out that when it comes to numbers, ChatGPT might be unintentionally inflating your translations in ways you never intended.

ChatGPT’s Number Inflation: Case 1

Like many, I often find myself second-guessing while translating large Japanese numbers.

Is 100 million 一億 or 百万円? How many zeros are in 一億 again?

In the following screenshot, you’ll see me asking ChatGPT for help with a particular translation.

Let’s take a look at these two sections I highlighted in red.

ChatGPT mistranslated both 9億円 and 125億円 as 9 billion yen and 125 billion yen! (The correct translations are 0.9 billion yen and 12.5 billion yen, which is a significant error.)

Case 2:

Fool me once, shame on you. Fool me twice, shame on me. But completely lie to my face?! Now that’s bold.

In this next example, I needed to type out 42,383百万円 with all the zeros and commas.

ChatGPT told me the answer was 423,830,000,000, and during a final check, I realized this number might be off. If 百万 is one million, then 42,383 million is 42.383 billion. So why was ChatGPT saying 423 billion?

To double-check, I asked ChatGPT if 42,383,000,000 was wrong.

As you can see, ChatGPT doubles down and tells me that 42,383,000,000 is incorrect! Upon further explanation, we can see what really happened is that ChatGPT calculated the final answer based on the assumption that 百万円  = 100,000,000, or 100 million. (Yikes.)

It never crossed my mind that one of the world’s most advanced AI models would miscalculate numbers of all things.

Double Check and Always Be Wary

I don’t think I need to spell out why these mistakes pose serious risks for those who rely on generative AI to translate or verify their work.

Never take AI-generated translations at face value. While AI can be a powerful tool, it still has a long way to go before we can fully trust the answers it provides. Unless your goal is to impress investors with an imaginary billion-dollar valuation, it’s best to keep AI on a short leash when dealing with financial translations.


The following two tabs change content below.

Jessica Azumaya

最新記事 by Jessica Azumaya (全て見る)

コメントを残す

メールアドレスが公開されることはありません。 が付いている欄は必須項目です

このサイトはスパムを低減するために Akismet を使っています。コメントデータの処理方法の詳細はこちらをご覧ください