Generative AI Hallucinations: How Often Do They Happen and Should Marketers Be Worried?

Research Reveals Generative AI Hallucinations

Throughout 2023, generative AI has exploded in popularity. But with that uptake, researchers and practitioners have warned of the pitfalls of generative AI, such as generating biased results, infringing on copyright, and sharing inaccurate answers to queries. In fact, generative AI tools (also known as large language models) are capable of fabricating information, a phenomenon known as a hallucination.

Vectara, a start-up launched by ex-Google staff, is exploring the frequency with which chatbots deviate from factual accuracy. According to the firm’s studies, chatbots fabricate details in at least 3 percent of interactions, and this figure can reach up to 27 percent, despite measures taken to avoid such occurrences.

Vectara relied on its own Hallucination Evaluation Model as an open-source model on Hugging Face, along with a publicly accessible leaderboard. The leaderboard serves as a quality metric for large language model factual accuracy, similar to how credit ratings or FICO scores function for financial risk.

This gives businesses and developers insight into the performance of different large language models before implementing them.

Chatbots are so versatile that it is difficult to say definitively how often they hallucinate. Vectara researcher Simon Hughes, who led the Hallucination Evaluation Model project, told The New York Times that it would be necessary to examine all of the world’s information to make a determination. Hughes and his team tested the chatbots’ ability to summarize news articles, a straightforward task that is easy to verify. Even in this simple setting, the chatbots frequently invented information.

“We gave the system 10 to 20 facts and asked for a summary of those facts,” said Amr Awadallah, Vectara’s chief executive and a former Google executive. “It is a fundamental problem that the system can still introduce errors.”

The researchers argue that when these chatbots perform other tasks (beyond mere summarization) hallucination rates may be higher.

Now, here’s the good news for users of ChatGPT: OpenAI’s technologies had the lowest hallucination rate, around 3 percent.

What Marketers Should Do about generative AI Hallucinations

It’s easy to get alarmed about a concept such as a large language model hallucinating, but at the same time, do you trust everything you find on the internet when you research a blog post, idea for a marketing campaign, or anything else for that matter? This is no reason to stop using generative AI. But you should be careful, as you should be when you’re online doing anything. For example:

• Please don’t allow generative AI to do your research and writing for you. If you do that, you have bigger problems on your hands than a generative AI hallucination. Generative AI is a tool, not a substitute for what you do as a marketer.
• Do use generative AI as a starting point. Create detailed, thoughtful prompts to uncover answers to questions you are trying to solve for as you do your work.
• Cross-check your content. But you should be doing that with all the content you create, regardless of techniques you use to create anything. No marketer should work alone without peer review of their work. It’s called teamwork. Embrace it.
• Verify your sources. Verify all the sources that AI uses to ensure they are credible and reliable.
• Use generative AI tactically as you create content. Run drafts of your work through generative AI and ask what you missed. If gen AI suggests something, check it out first – not necessarily because of the odds of a hallucination occurring but because good marketers vet their work. (And yes I did so with this post.)
• Monitor developments. Keep up with the latest developments in AI to continuously improve the quality of outputs and to be aware of new capabilities or emerging issues.

At IDX, we’re on the front lines using technology to create content more effectively (and this statement goes well beyond generative AI). Click here to learn about our services. Click here to learn more about the opening of our new content production studio. We’re here to help you!