How To Avoid Your Own AI Content Blunders

Adam Driver as Cesar Catilina in Megalopolis. Photo Credit: Courtesy of Lionsgate

Courtesy of Lionsgate

The recent Megalopolis trailer debacle was shocking to many in the film industry and beyond. When Lionsgate released the trailer for the much-anticipated Francis Ford Coppola epic, it quickly found a problem. The quotes from famous film critics were not only inaccurate but inconsistent with the sentiment of the actual reviews.

These quotes, attributed to legendary film critics like Pauline Kael and Roger Ebert, turned out to be generated by AI. They went unchecked before making their way into the final cut. The result? Job loss for Eddie Egan, the marketing consultant responsible for the trailer. And, both the film and the studio took hits to their credibility.

So, how did this happen? And more importantly, how can you avoid a similar pitfall when using AI in your own content creation?

Why Do AI-Generated Errors Happen?

AI tools, including Large Language Models like ChatGPT, are powerful but far from perfect. They excel at generating human-sounding text by predicting what words should come next based on vast amounts of data. But, as smart as they seem, they don’t truly understand the content they generate.

Instead, they create plausible-sounding text. This content can include errors, especially when the LLM is asked to produce specific quotes.

These mistakes, often called “hallucinations,” occur because AI can’t distinguish between facts and fabrications—it’s just trying to fulfill the request based on the patterns it has learned. This is particularly problematic when generating quotes from real people. It may be acceptable to put the conclusion of an article in different words, but not to fabricate direct quotes.

The Megalopolis trailer incident is a very visible example of what can go wrong when AI-generated content isn’t checked.

How Can You Ensure AI-Generated Quotes Are Accurate?

If you’re using AI to help create content, you don’t want to suffer the same fate as Eddie Egan. Even if your job isn’t on the line, it’s possible to avoid the embarrassment of sharing incorrect or fabricated information.

I ran into my own hallucination problem when I asked ChatGPT to suggest some useful quotes from the transcript of a podcast interview I had recorded. The quotes sounded great, exactly what I had asked for. One problem: when I checked them against the transcript, they didn’t exist.

When I asked ChatGPT about the accuracy of the quotes, it replied that the quotes, “were created to fit the tone and content of the article, based on the typical style of statements that an executive in his role might make. They are not direct quotes from the transcript you uploaded.”

ChatGPT offered to try again with direct quotes, but didn’t manage to do that on two more attempts despite my prompting to use exact text from the transcript. The quotes were consistent with the content, but didn’t match the transcript.

I asked AI expert Christopher Penn, Chief Data Scientist, TrustInsights.ai, for advice on getting better results. He suggested several steps you can take to ensure accuracy, especially when it comes to quotations.

Should You Use A RAG-Locked System?

One of the most reliable ways to prevent errors is by using a “RAG-locked system.” This sounds more intimidating that it is. RAG stands for “Retrieval-Augmented Generation,” which means the AI can access and use a specific set of information. “Locked” means the AI can’t change that information during normal use.

Using a RAG-locked system ensures the AI can’t generate content outside of what’s been provided, effectively eliminating the risk of fabricating quotes. Penn suggests tools like Google’s free NotebookLM or the open-source software AnythingLLM in query-only mode as examples.

I tested NotebookLM (labeled as “experimental” by Google) by uploading a 5000-word transcript of a podcast interview. I asked it to suggest some useful quotes from the guest. That’s not a particularly specific prompt, but my sole objective was to test the quote accuracy. It produced a group of quotes that did, indeed, match the transcript perfectly.

Can You Reduce Hallucinations With ChatGPT?

If you are already comfortable with a general purpose LLM like ChatGPT, Gemini, Claude, CoPilot, etc., you can change your prompts to improve accuracy.

“The challenge for most people is that their prompts are simply too short and too vague, which leads to mistakes and hallucinations in AI,” Penn notes. He recommends a technique called “few-shot contrastive prompting,” where you provide the AI with a few examples of how to do it right and a few examples of what not to do. This approach gives the AI clearer guidance and reduces the likelihood of generating erroneous quotes.

He also emphasizes the importance of being comprehensive in your prompts. While some still advise users to keep their prompts short, Penn counters that AI models can handle much longer inputs—up to 90,000 words in some cases. So, don’t hesitate to provide detailed, context-rich prompts to ensure the AI has all the information it needs to get it right. “Keep your prompts relevant, yes. Short, no,” Penn says.

In my earlier experiment, I finally got better quotes from ChatGPT by telling it to forget the previous conversation, re-uploading the transcript, and specifying exact quotes with timestamps. Even these were not quite word-for-word, but they were close to the transcript text and captured the meaning well.

Don’t Forget Manual Fact-Checking

Since AI generates content based in part on probability, its fabrications are usually plausible. Since hallucinations may not be obviously incorrect, human oversight is still essential. After generating content with AI, always cross-check any quotations, statistics, and factual statements against original sources.

As the Megalopolis trailer fiasco shows, failing to double-check AI-generated content can lead to bad outcomes. Understand the limitations of AI. Use RAG-locked systems when working with specific starting content. Use detailed prompting for ChatGPT and similar LLMs. Whichever AI you use, check the output carefully. Use AI the right way, and you’ll preserve both your job and your reputation.

Originally Appeared Here