Reduce AI Hallucinations

It is both inevitable and unavoidable to deal with hallucinations in generative AI. Maybe someday, today’s AI hallucinations will be viewed as a historical anomaly. Asimov should have derived seventh, eighth, or ninth laws to govern machine behavior, to ensure our computer creations will not lie to their human builders.

Hallucinations can be so bad, they are frequently the subject of memes in their least benign ridicule, and at worst provide debate fodder to allegedly prove generative AI cannot be trusted. If AI is so prone to make stuff up, to deliberately give incorrect answers in the absence of model data as basis for feasible output, then what’s the point? Is it not too great a risk to attest to AI output that turns out to be wrong?

You cannot stop AI from giving up hallucinations completely, for now, but there are practices you can control and take to minimize their chances to impact your generative AI results.

AI Is Not The Oracle at Delphi

There is, I believe, a common and never discussed assumption that, because we consider generative AI as so technologically advanced, and we are living in the most advanced and modern of technological times ever known, that models which wordsmith perfect grammar must have an innate or even divinely granted means to provide us mere mortals the answers we so desire and require, whether real, true, or not. You know, intellectually, AI hallucinations are real, but you cannot escape the subconscious emotional pull affirming such digital power so sophisticate could ever be wrong. The less one understands how generative AI and their models work, the greater the pull.

AI is not a direct conduit to the Temple of Apollo. AI, simplistically, is the latest collective frontier of computers and computer programming. Computers and programs made by humans, even if ever removed more and more from the building of their source code, are still fallible, because their creators are fallible.

What AI can do is both impressive and daunting, yet we must not let these characteristics rule our psyche when engaging with AI models. The real arbiters of what is real and genuine, what results and output are worth putting our names to, are still us.

AI Is Not Trivial Pursuit

A common reason for hallucinations is asking AI questions that demand a right answer. While many common questions are of this type, think of how you sought answers before AI. Likely in the absence of an expert or accepted source material, you would perform an Internet search. That search provided many results and hyperlinks. The onus was on you to decide which sites, which sources, which authors had the most correct answer. While easier for basic factual look-ups, the task would become more complex and formidable on softer and abstract topics.

Your need for inquiries has not changed, only the tools used, yet those tools, the generative AI models, should be treated with the same approach. Just as you would not unconditionally accept Google’s first search result, you should not take an AI’s prompt response as the correct and final answer. Helpfully, all major LLMs provide source references to how they generated their responses. It is your duty and responsibility to affirm the validity of those sources paired with the answer provide by the AI.

In my personal and professional practice with AI, as a rule of thumb, if I am seeking discrete, quantifiable data or information, I still use a classic Internet search; AI usually feels like unnecessary overhead in these circumstances.

Use AI for Preparation, Not Final Output

To truly avoid or at least greatly mitigate hallucinations, never start with the assumption that AI will create the final copy you seek. An impression I have gotten, especially from the philosophy of relegating human work, is that AI is good enough to create end results. While there may be niche cases of this, I would never trust AI to write an article for me unedited (sorry, Cassie), or create a PowerPoint presentation for sharing with my colleagues, or a collection of meeting minutes. In each case example, I have experienced firsthand how AI will get things wrong, not on purpose, but because it lacks context and understanding, and even an appreciation of tone for the target audience, regardless of how well defined the source prompt is.

AI should not produce final copies, but draft copies. Let AI help you build a starting point, but just like with any draft produced from any range of human expertise, initial content must be validated, refined, replaced, and sometimes outright removed. If you accept an AI’s output as final, you are accepting the AI’s tone as your own. You now “own” that tone, and people will notice if the content you provide is not in the style they would have expected from you in the absence of AI generation.

Understand the Subject/Topic

If you are not knowledgeable in a particular area, on what grounds can you trust AI to produce results that are correct? How would you trust a human in the same scenario?

Allow me to provide an example I have directly run into which underscores why it is important to be comfortable on a topic when using Generative AI, at least to the point you are able to sense something is wrong. I frequently use LLMs like Copilot, Gemini, and ChatGPT for assistance with technical commands for maintaining my home Linux Fedora build. I may ask the LLMs to provide me with command lines to perform certain tasks. Every so often, a response will be given for a non-Fedora branch of Linux, usually Debian or Ubuntu, the most common Linux branch. As a stark example, an “apt-get” command cannot be run in Fedora; instead, you have to run df.

If I knew nothing about how the different Linux distributions manage packages, I would be trying to run apt-get commands in Fedora. While the result here would be benign since the command would error, what about a case where I ran a “sudo” (root/admin) command that did execute, and I had no idea what it was doing? It’s not that I have to be a certified Linux system administrator, but I should know enough to understand Linux commands in general and what a specific command may or not be attempting, and be willing to accept the risk of the AI’s recommendation.

A tangent on using AI as a learning tool: Although you can use generative AI as a learning aid, and I have from time to time, you should do so only with the upmost caution. An AI generated learning plan may be inferior or outright wrong, pieced together in ways potentially not suitable for human learning (or your learning), and especially to someone unfamiliar with a subject. Yes, an AI generated learning plan could be fine, but personally for me, I would not risk doing so at this time. Instead, leverage AI to find appropriate learning paths and proven sources to help expand your knowledge and skills.

Use AI as Collaborator

The best way to minimize AI hallucinations is to leverage AI in ways that make hallucinations a mute (or near mute) point. An easy approach is to treat your generative AI tool as a virtual collaborator.

A collaborator is not an expert, not a teacher, not an oracle, but rather a colleague, an associate, a partner, an assistant. And like their human instances, you work side-by-side with your collaborator, to bounce ideas off of, or to get a different perspective. What is true, or more specifically, how you wish to determine the validity of your collaborator’s assertions, is entirely up to you and your judgement.

If you have not used generative AI LLMs for collaboration before, you start by instructing the AI what role you wish it to play (i.e. role play).

“You are a fellow project manager.”

“You are an associate on my operations support team.”

“You are an accomplished and published author specializing in food and dining practices, and will collaborate with me to compose an article.”

Pursing generative AI as a collaborator does not have to be elaborate; you can start by merely asking, and then keep asking, the AI questions on a particular topic.

Use AI for Creativity

Leveraging generative AI LLMs as a creative muse is a surefire method to take hallucinations completely out of the equation. More so, when it comes to looking for creative or offbeat answers from an AI, a hallucination may unexpectedly give you something brilliant.

Let’s take an example by jolting the traditional Thanksgiving meal:

I want to explore options for nontraditional Thanksgiving centerpiece foods. Turkey is of course the mainstay, and there may be other variants to this end, but I want to pursue, with the family’s permission, focusing the Thanksgiving meal on something completely different, but still in keeping with the spirit of the holiday. Please provide ten suggestions.

I ran this through Google’s Gemini and received interesting suggestions, from crown roast of pork for being visually spectacular and signifying abundance, to labor-intensive beef Wellington reinforcing dedication to the celebration.

On a more practical level, AI can offer creative solutions when you simply don’t know where to start. Have writer’s block? Ask AI to help. Need to present on a topic of your choosing to your work colleagues? AI can offer suggestions with defined criteria.

In all of these “creative” AI use cases, there is no “right” answer, and hence no chance of the AI hallucinating, at least to a degree that matters.

Conclusion

What is your take on AI hallucinations? Are there practices you employ to minimize their impact when working with generative AI? I am interested to learn your perspectives, which no AI can provide!