AI hallucinations can be frustrating. If you’ve used an LLM, you’ve almost certainly seen it deliver an answer that was either confidently wrong or just downright mistaken.
Co-founder, Executive Chair and Chief Innovation Officer of Zappi.
But when I asked the model to assess our customer reviews, it confidently concluded we were underperforming due to failures in our “electricity structure systems.” At first glance – “huh?!” But it became clear the model had conflated us with an unrelated company that shares our name and makes EV chargers.
Most people view hallucinations simply as annoying mistakes – but the truth is, hallucinations are byproducts of how LLM models are trained and what they’re optimized to do. If you’re expecting AI tools to be perfect, you’re expecting the wrong thing.
I provided clearer context, the model produced accurate results, and I was able to get the intel I wanted.
Why hallucinations occur
So why exactly do hallucinations happen? A recent paper from OpenAI shows hallucinations occur because models are rewarded for giving an answer—not for saying “I don’t know.”
An LLM is never deterministic; it’s always probabilistic.
OpenAI has explained that during the pre-training phase, AI models learn by ingesting vast amounts of data from the internet. In this initial stage, these models do a good job at signaling how confident they are in the answers they provide. They can also signify uncertainty reasonably well, saying, “Here’s a possible answer, but I’m unsure.”
However, when it comes to post-training, models are refined by reinforcement learning that rewards accuracy… without penalizing inaccuracy. Just like a multiple-choice exam, the LLM is trained to give an answer even if it’s a guess. As with humans, it often serves the system better to fill in something, rather than leave the question blank.
LLMs: Confidently wrong by design
Up until AI came along, we lived in a largely deterministic world. We used tools that provided a single, definitive answer, with little room for interpretation. For example, if we plugged a mathematical problem into a calculator, we’d get an answer.
If we queried a database for a document, it would provide it. We could trust these tools to return a predictable result.
LLMs are not the same. AI was designed to mimic how the human brain works, and humans are imperfect; they get things wrong all the time. So, if we expect that LLMs are going to get things right 100% of the time, we’re misunderstanding how an LLM works in the first place.
LLMs are a probabilistic system that produces the most likely answer, rather than guaranteed truths – which means they can be confidently wrong in the same way humans can be (or at least many of my colleagues accuse me of being!). The bottom line: An LLM that never hallucinates is simply not possible.
Demanding perfection and accuracy from a system is a human flaw.
How to reduce hallucinations
When it comes to curbing AI hallucinations, knowing that they’re a feature and not a bug is half the battle. It starts with resetting expectations – realizing that errors are inherent and not a fatal defect. The good news is, model developers like OpenAI are also working to decrease the rate at which hallucinations occur.
In the meantime, what can businesses and teams do about them? Here are three practical tips to keep in mind:
1. You can’t rely on the model alone for facts. As I mentioned, LLMs aren’t deterministic. Companies need to plan for errors by carefully reviewing the information returned and double-checking sources the LLM is pulling from.
Even if you prompt the LLM to only answer if it is 100% sure, it’s often still unlikely to say, “I don’t know the answer.” So, just like you would carefully review a colleague’s work, you need to monitor LLMs for accuracy too.
2. Feed the model trusted, connected information. What you give an AI system matters as much as what you ask it. The more you ground a model in trusted, connected sources—validated research, internal reports, documented decisions, and shared institutional knowledge—the more useful and reliable its outputs become. When data is fragmented or vague, the model fills gaps. But with clear, current, connected inputs, AI can reason within real constraints instead of guessing.
3. Use carefully curated prompts. The more general the prompt, the more general the response. You can better control the outcome by providing relevant context and source material, and then asking a specific question. The prompt then becomes, “Answer this question only using the data I provided, and then cite where the information came from.”
This can dramatically reduce hallucinations. You can even prompt the model to be more nuanced by saying “If you are not 100% sure about the answer, then say you don’t know. Accuracy is very important here.”
AI as a system, not a magic box
AI is a powerful tool that we will continue to fold into our everyday working lives and beyond. We must realize, though, that AI isn’t a magic box. It’s an imperfect system that reflects the training and insights it’s been given.
Only when we stop expecting perfection from AI, can we use it in the way it works best – alongside us – to deliver real business value.
We’ve featured the best AI chatbot for business.
This article was produced as part of TechRadarPro’s Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro
Leave a comment