A strange side effect of generative AI is hallucinations, or when a GenAI tool confidently delivers completely made-up information. Maybe you asked for a works cited page and received citations for articles that don’t exist. Or you’re given a plausible-seeming statistic that turns out to be fake. Whatever way these hallucinations come up, it can be hard to tell what’s true and what’s a believable lie.
As GenAI becomes more prevalent in our daily lives, it’s important to understand why hallucinations happen and how to work around them.
🤖 Why GenAI Makes Stuff Up
Large language models (LLM) predict the next most probable word based on patterns learned during training. It’s sort of like really advanced autocomplete. And as we know by now, these tools are remarkably useful for tasks that rely on pattern recognition, like summarizing, coding, or writing.
But because models recognize patterns rather than recall data directly, they sometimes make incorrect assumptions due to gaps or biases in training data. Researchers are actively exploring ways to reduce hallucinations, but with the current tech, they seem to be here to stay (and by some accounts, hallucinations may get worse as models become larger).
That means that it’s up to us to make sure we’re not peddling misinformation every time we use a GenAI tool. Most information generated by AI is accurate, which can make hallucinations tricky to spot. GenAI doesn’t have a “tell” — it delivers false info with the exact same style and confidence it uses for facts. This can become a big problem when you’re exploring a topic you don’t know very well.
To see for yourself, think of a niche topic that you’re essentially an expert in and a separate topic that you’re interested in learning more about. Then, open your favorite Generative AI @ Pitt tool and ask it to give you a brief description of each topic.
Most likely, you’ll find a few inaccuracies in the summary of the topic you know a lot about. These can be obvious — like a broken link — or more subtle, like made up details that fill in information gaps the AI couldn’t find easily. If you know the topic, the hallucinations will be clear (and maybe even laughable), but for the topic that you’re learning for the first time, you’ll find it can be difficult to pinpoint which details are real and which are fabricated.
✅ Verify, Verify, Verify
Your best defense against hallucinations is good old fashioned research skills. Check the sources GenAI cites, use primary sources whenever possible, and never repeat information you haven’t independently confirmed.
Yes, this can be time consuming and tedious, but it’s better than spreading misinformation. Here are a few ways you can protect yourself:
1. Use Your Subject Matter Expertise
Because hallucinations are more obvious if you have some pre-existing knowledge, being a subject matter expert is a secret superpower against GenAI lies. While you might not know everything about a given subject, you'll be better equipped to sense when facts don’t add up, or what parts of the AI-generated info need more scrutiny.
2. Ask for Sources (And Check Them)
If you’re using GenAI for research, always ask for sources. Often, you’ll receive web links that make it much easier to verify where information is coming from, and if the AI made any errors as it synthesized the information. Sometimes, even if you ask for sources, the tool will provide some info without citing anything. That’s a red flag that you’re looking at a hallucination.
3. Limit the Data Pool
Tools like NotebookLM exclusively reference sources that you provide, which limits the number of hallucinations you’ll get back. Because the data set is well-defined, the GenAI tool is less likely to make errors as it grapples with the entire volume of the internet. It might seem counterintuitive, but by reducing the amount of input, you’ll get far better output.
4. Spot the “Too Perfect” Answer
If something sounds flawless or overly polished, pause. Hallucinations often hide in answers that feel right but lack nuance. Cross-check dates, names, and stats — especially if they’re oddly specific. A quick search for a primary source can save you from spreading fiction as fact.
5. Be Cautious with Niche or Current Topics
The less training data available on a topic, the higher the chance of errors and hallucinations. If you’re asking about a topic with less information, like obscure subjects or breaking news, the GenAI tool is more likely to fill gaps with plausible-sounding fabrications. And since models are trained on data up to a specific cutoff date, they know essentially nothing about current events.
🎓Human Intelligence > Artificial Intelligence
At the end of the day, your best defense against hallucinations is to use your own intelligence. While GenAI tools seem confident, they’re much more likely than you are to make things up. These tools can help us to speed up our workflows, find new information much faster, and reduce administrative headaches, but they’re far from perfect. Hallucinations are one reminder to stay in the driver’s seat any time you’re using a GenAI tool.
If you practice good research skills — and avoid using GenAI tools as a singular source of truth — you’ll be well defended against hallucinations and potentially spreading dangerous misinformation. Stay safe out there, Panthers!
— Pitt Digital
