7 Tips for Tackling Generative AI Hallucinations

Generative AI hallucinations have become such a hot topic that they gave rise to Dictionary.com’s 2023 Word of the Year. In this post, we’ll take a look at what AI hallucinations are, what causes them, the dangers they pose, and share a few examples so that you can spot them. And then we’ll share our top seven tips for how you can avoid the pitfalls of this quirk of generative AI.

What Are AI Hallucinations?

When we hallucinate, we see or hear something that isn’t really there. The term has been adopted to describe situations in which a generative AI tool produces a result that is incorrect, misleading, or seemingly imagined. It’s basically a polite way of describing ways in which AI gets it wrong.

What Causes AI Hallucinations?

It might be easier to understand if you think of a generative AI tool as an overeager child; it’s keen to show that it listened to the grown-ups talking, but it didn’t quite understand what they meant.

AI tools are trained using vast amounts of data, but not all of that is reliable or up-to-date. A quick glance at internet search results or social media will give you an idea of how much unreliable or conflicting information is available.

However large the amount of information fed to the AI tool, there will still be gaps. Nevertheless, those tools look for patterns, make assumptions, and enthusiastically provide a response, which can lead to nonsensical results.

And we don’t always help. A poorly phrased question or a cheeky attempt to trick it can provide some interesting results. We’ll look at one of those later.

What Are the Dangers of AI Hallucinations?

AI hallucinations may make for a great anecdote but, left unchecked, they can lead to serious consequences:

• Misleading or incorrect information in your content can damage your credibility and undermine trust, which can lead to a loss of business. 
• Producing or distributing content that includes AI hallucinations could leave you open to legal action, which will likely cost you more than just your reputation.
• AI tools may generate stereotyped, racist, or otherwise biased content that would be harmful to others as well as to your reputation if you were to publish it.

AI Hallucination Examples

We’ve talked a lot about what they are and what causes them, but what does an AI hallucination look like?

It could be as simple as a factual error, for example in a date:

The French Revolution started in 1789.

But it could be more subtle. After asking ChatGPT to provide some fun facts about Prince Rupert of the Rhine that might appeal to school children, we gave a follow-up prompt for a particular favorite:

Commendably, ChatGPT recognized and alerted us to this being anecdotal, but the rival’s name rumored to trigger the dog’s response was John Pym, not Cromwell.

Of course, just like human hallucinations, those produced by generative AI can be quite bizarre, like the one provided to Satyen K. Bordoloi in a blog post for sify.com. He cheekily asked ChatGPT “What is the world record for crossing the English Channel entirely on foot?” and received a confident, detailed, and fabricated response. Do we need to say that it’s not possible to cross the English Channel entirely on foot?

How to Prevent AI Hallucinations

Although there are steps you can take to reduce the risk of getting an AI hallucination in the output from your tool of choice, the likelihood is that they’ll creep in. So, when we talk about preventing AI hallucinations, we mean preventing them getting into your published content.

Here are our top seven tips for tackling generative AI hallucinations. We’ve split them into two sections: those you can use to reduce the risk of your AI tool hallucinating (writing tips) and those which you should use to avoid publishing any hallucinations that slip through the net (editing tips).

Writing Tips

1. Provide Relevant Information

Understanding how to write effective prompts and following them up will help to refine the content and reduce the risk of hallucination.

2. Tell It Where to Look

If you tell the AI tool to provide its response from a source that you know to be trustworthy – or, better still, give it the URL for the source – you will cut down the risk of it generalizing data from across the internet and falling victim to misinformation.

3. Tell It Where Not to Look

Steer the AI tool away from misleading information either by reference to a year – to include only up-to-date information – or to a URL you consider untrustworthy.

4. Include “I Don’t Know” as an Option

Giving the AI tool the specific option to say when it can’t find the answer can help to prevent it from imagining one.

Editing Tips

1. Check the Facts

The simplest advice is to trust nothing about the result you’ve received. Have a look at our post about how to fact-check AI-generated content for guidance on this vital skill.

2. Don’t Check the Facts with Generative AI

It might be tempting, but simply asking the AI tool to cite sources for its response may only add to your problems. It might produce a list of references that looks convincing but isn’t real.

3. Proofread the Output

Word and punctuation choice can alter the meaning of a sentence; proofreading what your AI tool has written can save you from unintentionally misleading your readers.

In Summary

With careful prompting, you can reduce the risk of generative AI hallucinations, but the best advice is to treat all such output as a first draft: go back through it carefully, checking everything before you publish. You’d do that for your own writing, so make sure to do it for this!

If you’d like to develop these vital skills, why not try our AI Prompting For Writers And Editors course? The course will teach you how to write effective prompts – to reduce the likelihood of AI hallucinations – and edit AI content to identify and fix incorrect or misleading information.