Q: Do I have to fact-check what ChatGPT says?
Yes. If it is important for ChatGPT's response to be correct, you will need to fact-check its output.
- ChatGPT has been known to make up false information. This is often called an "AI hallucination."
- The false information it generates will often sound very plausible. At times this can make fact checking difficult, as the false portions may not be immediately obvious.
- Due to how ChatGPT works (as of May, 2023), it cannot fact-check itself. Simply put, its "knowledge" is the result of being trained on massive amounts of data. If its training data contained a large amount of false information on a specific topic, or if it lacked sufficient information about a specific topic, ChatGPT may be more likely to answer incorrectly on that topic.
- For example: let's say that there's a commonly-believed myth. ChatGPT may have been trained on many example conversations between people who believed this myth. This would likely result in ChatGPT being trained to respond as though this myth were true. Here's a conversation with ChatGPT demonstrating this phenomenon:
- Prompt: What are a few famous Sherlock Holmes quotes?
- ChatGPT: Sure, here are a few famous Sherlock Holmes quotes:
- "Elementary, my dear Watson."
- "When you have eliminated the impossible, whatever remains, however improbable, must be the truth."
- ... (further quotes removed for brevity)
- The first quote, "Elementary, my dear Watson," never appears in any of the Sherlock Holmes books. However, it is commonly-believed to be a Sherlock Holmes quote, so it is likely to be present in a great deal of ChatGPT's training data, despite its status as a myth.
