AI can create high-quality content extremely fast. Thanks to this, it is doing a great job of automating tasks and inspiring new ideas. Yet, there is one BIG problem: hallucination by generative AI. It’s also good at making things up to get things done as requested. One expert estimates that around 15%-20% of ChatGPT content might be hallucinated (made up) 1. So does this mean that we should give up on it all together? No. It would be a big mistake to discount generative AI as unreliable and put it aside. Instead, we need to create a new culture of fact-checking, just like we did when we first got the internet. AI can (and should!) be used as a power tool to support any content creation processes.
What is a hallucination?
As previously mentioned, hallucination is when the information presented as factual is made up. For example, when I asked ChatGPT to list ten reasons why Diso, in Puglia, Italy, is a great place to live (ChatGPT’s 10 Best Reasons to Live in Diso (LE) Puglia (corvaglia.me)), it told me about the beautiful medieval castle. The problem is that there aren’t any castles there. It also isn’t in a famous wine region, as ChatGPT had stated. While I immediately recognized this as an untruth, sometimes it isn’t that easy. The tool can get quite creative in fooling us. When one colleague of mine asked ChatGPT for the source for a statement it had made, the source was real, but the quoted sentence wasn’t in it! Oops.
Why do Generative AI tools hallucinate?
The fact that it’s making stuff up doesn’t mean that ChatGPT is actively lying. It shouldn’t be compared to humans who will lie and cheat to get ahead. As far as we know, it isn’t nearly that “sophisticated” yet. The source of incorrect information can be both technical and human. According to techtarget.com, there are three main reasons for hallucination: bad data in the source content, training, and generation methods of the AI model. It can also be bad quality user input when asking the tool to create content. In other words, incorrect facts may be present in data that the model learned from. Alternately, it may have been taught or received a request in a way that made it think that the best solution was to hallucinate facts.
Why 20% of invented information isn’t a show-stopper
The fact that generative AI can’t always be relied upon is annoying. Yet, whatever you do, don’t give up on generative AI because of hallucinations. New technologies always have bugs to be ironed out. As long as the user can work around them and there is still value in using the tool, they shouldn’t be considered show-stoppers. For instance, I’m old enough to have started my school career without widespread access to the internet. Once we got it, though, we had to use it carefully for our research. The internet was terrific for getting quick information, but it essentially hadn’t been fact-checked. (This wasn’t a problem when most info came from printed encyclopedias!) We didn’t stop using the internet because it was unreliable. Instead, we created a whole new concept of how to check the information found there.
Fact-checking is crucial, but not the only answer
A few simple steps can be taken to decrease your chances of getting burned by using hallucinated information from generative AI. The most important one is to give clear and specific requests (i.e., prompts). For example:
- Give the model a role to play
- Be specific about the type of output you want
- Tell it to only give information that it has sources for
- Tell it site its sources
- Tell it to ask you if something is unclear about the request
Using this technique can significantly increase the quality of output. But beware, even the best prompts can’t completely exclude the risk of hallucinations. As my colleague learned, everything needs to be double-checked (don’t just check if the quoted document exists, find the exact quote in it!)
Conclusion: don’t give up on generative AI just because it’s making stuff up
As we did when we first had the internet, we now need to find the best way to rely on the information we use from AI. Hallucination by generative AI isn’t a reason to give up on it. We need a new concept of fact-checking that also works with AI-generated content. It all starts with starting to use generative AI, second-guessing all content, and finding a way of being able to trust it. It’s also critical to remember that the benefits of having a helping hand in creation far outweigh the risk of getting incorrect information. Even in its imperfect state, quickly created supporting content and generation of ideas significantly augment human abilities. We should all lean into this and find a way to use it despite its faults.
1 Estimate as of July 2023, according to Peter Relan, a co-founder and chairman of Silicon Valley company Got It AI. Source