These 4 specific examples include not just hallucinations, but an AI tool doing more than I asked for, making the process less efficient. Jump to my KEY TAKEAWAY – Tip to Be Efficient Working with Gen AI.
1. Fabricated Education & Accomplishments
I thought writing an executive bio for the website would be a perfect task for GenAI. But ChatGPT claimed an executive had a degree in computer science when he had studied engineering and never earned a degree. (He chose to enter the workforce prior to completion.) It also claimed executives had spoken at major events. While that is an excellent accomplishment worth highlighting in a bio, it was a fabrication.
Additionally, ChatGPT presented me with a photo that was not my client. At first, I thought the photo was AI generated. Did this tool just decide, “Here’s what I think a guy with this name looks like,” Ha! But it was an actual person the executive had worked with in the past.
2. Actual Company, When Asked for Fake Name
In another case, a customer agreed to a case study, and their corporate counsel subsequently did not want the business named. Thankfully, they agreed to my request to anonymize the case study in order to maintain the wonderful content of their recommendations without identifying their FinTech business.
I asked ChatGPT to rewrite the company description to make it anonymous and give the company a fake name. Instead, it gave me the name of a real FinTech company. I asked it again to generate fake, made-up company names. Once again, I was presented with 10 actual FinTech companies.
3. Misunderstanding What’s a Book Title, “He Never Answered That Q”
NotebookLM is a free Google tool, a Large Language Model built with Gemini 1.5 that focuses exclusively on the information sources you add, serving as a personalized AI expert in your information.
I thought NotebookLM would be an ideal tool to generate a list of all the books podcast guests had recommended.
NotebookLM said some guests did not recommend a book when asked by the host, although they did. In another case, NotebookLM provided a book title which was not named, although the subject matter was identical to the book which was recommended.
4. Lovable, GenAI Web Dev Tool, Knows Better Than You
As a content marketing pro, I’d never write a blog without a CTA. I also don’t typically include captions below photos. I mention them in the body text or simply include them within the document for some visual appeal without describing them.
Lovable doesn’t know this.
I’ve pasted the blog text and prompted the system to create and upload a new blog, but it goes ahead and inserts a CTA at the end as well as captions below the photos.
Now that I know this, my prompt includes guidance. “Do not add any additional text on the page beyond what I paste.” This reminder is included in a doc along with other prompt text, such as telling the system to add a link to this new blog on the relevant author’s page.
KEY TAKEAWAY – Tip to Be Efficient Working with Gen AI
“Wait, what? AI makes you more efficient. It’s an inherent attribute of the technology.”
Yeah, sorta. But it’s not quite as simple as being able to unlock and start your car while leaving your keys in your pocket or purse. (LOVE THAT, btw.)
The above issues with Lovable are a quintessential example of how AI isn’t always fastest at first (in this case, compared to uploading a blog in WordPress.)
It’s important to document these issues and refer back to your notes so you operate efficiently moving forward. [Reminds me how I fear students won’t learn how to take notes if they write essays via ChatGPT…]
- Document issues
- Keep your notes organized
- REFER BACK (If you don’t read your notes, did you even take them?)
Do you need content marketing support from a human who understands tech and AI? Let’s chat.