SUBSCRIBE NOW!

Vol. 40 No. 8 — October 2023

LET'S GET STRATEGIC

A SLIPPERY SLOPE

There are some great potential uses for generative AI via tools such as ChatGPT, Bard, and Jasper. But there’s peril as well. These tools have made stuff up, much to the chagrin and detriment of their users. Some New York lawyers, for instance, have been sanctioned for their use of ChatGPT—not the actual use of the tool, but their failure to fact-check the content generated, which included false citations. Without appropriate due diligence, using those tools can result in impacts ranging from extreme to minor embarrassment, depending on the purpose of their use.

I was looking for examples of companies with employee resource groups (ERGs) specifically focused on older demographics—Baby Boomers and beyond. It generated what appeared to be an impressive list but, since I know the tool I was using didn’t have content beyond 2021, I went to the companies’ websites to verify the groups were still in existence. I couldn’t find any of them. So, I went back to the tool and asked, “Did you make these ERGs up?” Its response: “I apologize for the confusion caused earlier. The examples I provided in my previous response were hypothetical and not based on specific knowledge of existing ERGs.” I asked it to generate another list and specifically said, “Do not make these up; use real examples.” It gave me another impression list, but when I asked again if it had made the list up, I got the same response. Troubling.

Unfortunately, the tool has misled me in other ways. One use case I thought would be helpful was to have the tool generate SEO-related content—titles, email subject lines, meta copy, etc.—for blog posts that need to be a certain word or character length. And I discovered that it didn’t always adhere to the length requirements, even when I asked it over and over again to do so. Also troubling.

I’ve come up with some of my own ways of both minimizing these outputs through my prompts and doing some additional fact-checking. However, I wondered how others were tackling the issue. I reached out to other content marketers and generative AI experts. Here’s what they had to say.

WHY DOES AI HALLUCINATE?

The trouble with generative AI, say those most familiar with how it works, is that it tends to be a people pleaser. It appears to give you useful information. “The problem is, when it doesn’t know something, it pretends to know [it], instead of saying ‘I don’t have enough information,’ ” says Juliet Dreamhunter, a productivity and AI consultant with Juliety. Dreamhunter once input a link to an article she’d written, asking ChatGPT to “please summarize this article for me.” It generated a one-paragraph summary. Dreamhunter knew the content well. She could tell immediately that it was “completely made up.”

After learning about the concept of hallucinations, which is the term for when generative AI tools use false information, Dreamhunter went back to ChatGPT and asked how it had written the summary. From the “confession,” she says, “I found out that it didn’t follow the link but generated a summary of what an article like this could be about, based on the keywords in the URL and its own knowledgebase.” Despite the potential for hallucinations, there are steps that content marketers can take to minimize the potential for error—and check for it.

ENLIST THE ASSISTANCE OF DOMAIN EXPERTS

Fawaz Naser is CEO of Softlist.io, a platform that compares tools and software to optimize productivity. He shares an experience that his firm recently went through when using AI to generate content for its marketing campaigns. “While the created content was grammatically perfect and contextually relevant, it incorporated some facts about our company’s history that were unrealistic,” Naser says. “It claimed that we expanded into markets that we had not yet penetrated, probably extrapolating based on our rapid growth in our existing markets.” His recommendation is, “By leveraging AI as a tool, rather than a decision maker, organizations can make AI’s strength work for them while also mitigating hallucination risks.” Naser suggests doing the following:

• Use relevant domain experts to validate AI outputs.
• Continuously re-evaluate and retrain the AI model with updated, accurate data.
• Validate AI output using multiple models.
• Remind stakeholders—internal or external—that AI is not infallible, but a tool that needs to be controlled and supervised.

As with any other tool, it’s becoming increasingly clear that generative AI requires human oversight to ensure accuracy and relevance.

APPROPRIATE USE

John Pennypacker, VP of sales and marketing at Deep Cognition, a company that offers next-generation AI platforms and solutions, says his top recommendation when utilizing tools such as ChatGPT is to never use them as an encyclopedia. ChatGPT, Pennypacker says, is “great for organizing your content, tweaking here and there, or rewriting your content—but never depend on it for fact-checking.”

Pennypacker shares an experience he had when generating content about Python programming. “I fed it the topic and some basic points, and it came up with a pretty comprehensive piece,” he says. “Upon fact-checking, I discovered that it had made up a Python function. It was an impressive, creative hallucination, which wasn’t actually in the Python library.” Pennypacker’s example illustrates how critical it is to check the output provided by generative AI tools, regardless of how impressive and credible that output may seem . “Despite AI’s ability to generate content and provide suggestions, we are ultimately responsible for checking facts, ensuring accuracy, and imparting the right tone and flavor,” he says.

Dan Chadney, a web designer, front-end developer, and blogger, says that the best way to avoid hallucinations is to train the AI before making a request. “Providing extra context and actual facts to the AI first will always yield the best results,” he says. “I often use a ChatGPT plugin that provides web browsing functionality. Then when writing my prompt, I’ll ask the AI to look at specific URLs and use them as references for facts.” Another AI tool Chadney recommends is Perplexity, a research assistant. “I’ve found this to be one of the most useful tools ever, because it also provides sources—and URLs—for each fact returned. I’ve used it for writing biographies, stats, and software reviews. It’s still best to always double-check your facts, but Perplexity cuts out so much of the legwork, it’s a huge time-saver,” he says.

BETTER RESULTS OVER TIME