ChatGPT - absolutely certain and absolutely wrong
When is Chinese New Year?
I was just writing an update for my latest Kickstarter talking about the fulfilment timeline for early next year.
This is this first time I’ve used a playing card manufacturer in China and shipping times will be longer than I’m used to because I’m using sea freight.
I was fairly sure that everything should be sorted and the shipments well underway before Chinese New Year, but I wanted to double check the date of that.
So I turned to ChatGPT to ask: when is Chinese New Year in 2024?
January 25th
Sounds like a reasonable answer. Thank you ChatGPT.
However… I thought it was actually sometime in February. So I asked again.
February 13th
A different date, this time in February so maybe this could be it. I’m losing confidence though so I check again.
February 10th & January 24th
That’s now four different dates. I asked to confirm and it said that last date again. So I thought I’d dig a little deeper.
February 5th
Hmm, now I have a fifth date to add to the potential answers.
I asked if it was sure and it confirmed that Chinese New Year “will indeed fall on February 5th”.
Good. Maybe we’re getting somewhere. It seems to have zeroed in on 5th February so that must be it.
February 5th
Just to put my mind at ease, I finished with:
Great, it must be the 5th. I have to be happy with a firm 10 out of 10.
A firm wrong
Yet, alas, that’s wrong.
Chinese New Year next year will be on the 10th February, which was mentioned by ChatGPT at some point, along with the four other dates.
A few seconds of research the ‘old’ way quickly gave me an answer I was happy with. Wikipedia is always a relatively safe place to start, and then I found two reputable sources to corroborate it — the South China Morning Post and The Standard in London.
A printed book in a library would probably be an even more reliable source. That’s not to say that just because something is printed it’s true, but the processes and gatekeepers that are part of publishing a book add layers of credibility.
The internet is an amazing force for good when it comes to the democratization of information, but we need to keep our wits about us.
Is it a big deal?
This is a fairly trivial example, and I’m not the first person to document such errors or hallucinations from ChatGPT.
Using tools like ChatGPT, it’s easy to get lazy and complacent with our research and fact checking. If I’d have simply used the first date for Chinese New Year that ChatGPT gave me, I’d have looked pretty silly and I’d have lost trust from some of my Kickstarter backers.
More than that, it could’ve actually have had a real impact in my business if I’d miscalculated my shipping times based on a wrong assumption about when Chinese New Year is going to be.
In practice, when planning my project timelines I’d have done a lot more work than that. But such an error could easily find it’s way into a less mission-critical situation like a side-note in a blog like this. In itself that wouldn’t be the end of the world, but we need to be careful.
I know I’m sounding a bit melodramatic here, but I think there’s a real risk of a slow erosion of standards in our collective knowledgebase. The glib way ChatGPT gave a wrong answer with a firm 10 out of 10 certainty is especially frustrating. If it wasn’t a machine I’d call it arrogant. Can Dunning–Kruger be applied to a software application?
Unnoticed, that erroneous date would get put back into the system to feed the AI beast with misinformation.
Do your research, check your facts, and let’s try and avoid getting into a self-perpetuating downward spiral of ignorance.
…
Notes
I was using ChatGPT 3.5 here. I know that’s not the latest version, but that’s not really the point. It’s the one that’s freely available and used by me and millions of others.
I’d have thought it would be an easy question with plenty of references going back years, so how could it be so bad at answering it? I’d be interested to hear comments from people who understand these things better than I do.
