What Are AI Hallucinations and How to Prevent Them?

What comes to mind when you hear the term “hallucinations”? For most of us, this brings up images of insomnia-induced visions of things that aren’t real, schizophrenia, or some other sort of mental illness. But have you ever heard that Artificial Intelligence (AI) could also experience hallucinations?

The truth is that AIs can and do hallucinate from time to time and this is an issue for people and companies that use them to solve tasks. In this guide, we’ll take you through AI hallucinations, what causes them, and what their implications are.

AI Hallucinations Defined

An AI hallucination is a scenario where an AI model begins to detect language or object patterns that don’t exist and this affects the outcome that they are given. Many generative AIs work by predicting patterns in language, content, etc. and giving responses based on them. When an AI begins to generate output based on patterns that don’t exist or are completely off base from the prompt they were given, we refer to this as ‘AI hallucination.’

Take, for example, your customer service chatbot on an e-commerce website. Imagine you ask it when your order will be delivered and it gives you a nonsensical answer unrelated to it. That is a common case of AI hallucination.

Why Do AI Hallucinations Happen?

Essentially, AI hallucinations happen because generative AI is designed to make predictions based on language but doesn’t actually ‘understand’ human language or what it is saying. For example, an AI chatbot for a clothing store is designed to know that when a user types the words ‘order’ and ‘delayed’, its response should be to check the status of the customer’s order and tell them that it is on the way or has already been delivered. The AI doesn’t actually ‘know’ what an order is or what a delay is.

So if a user types in the chatbot that they would like to delay their order because they won’t be home, such an AI might keep telling them the status of their order without actually answering their query. If you spoke to a human who understands language nuance, they would know that just because certain words are in a prompt doesn’t mean the same thing every time. But AI, as we’ve established, does not. Instead, it learns to predict patterns in language and works based on those. Hallucinations also tend to occur when a user gives a prompt that is poorly constructed or too vague and this can cause confusion. Typically, AIs will become better at language prediction over time but AI hallucinations are still bound to happen now and again.

Types of AI Hallucinations

Typically, AI hallucinations occur in several different ways:

• Factual inaccuracies. Many of us are turning to AI to ask about certain facts and find out if some things are true or not. AI, however, isn’t always perfect and one way that AI hallucinations can manifest is by giving information that is simply not correct when a user asks.
• Fabricated information. It is not unheard of for AI to completely make up facts, content, and even people that don’t exist. AIs have been known to write fake news articles, including events and people that are not true and passing it off as the real thing. In the same way humans can tell tall tales, so can AI.
• Prompt contradiction. This happens when AI offers a response to a prompt that has nothing to do with what was asked at all. We’ve all asked our voice assistants, for example, about one thing and they begin talking about something completely unrelated.
• Bizzare statements. AI is known to make claims and statements that seemingly come out of nowhere and can be downright bizarre, such as making fun of users or claiming to be an actual person.
• Fake news. An AI hallucination can lead to the AI offering fake facts about actual people who do exist. This sort of information can end up being harmful to the people in question.

Consequences of AI Hallucinations

Now that we understand AI hallucinations better, it is worth exploring what its consequences are.

AI hallucinations can cause many serious problems. Firstly, they can lead to fake news. We as a society have been trying to combat fake news for several years now but AI hallucinations could put a dent in these plans. People rely on respected news outlets for legitimate news and if AI hallucinations continue to create fake facts, the truth and lies will be further blurred.

Secondly, AI hallucinations can result in a lack of trust in AI. For AI to continue being used by the public, we will need to have some trust in it. This trust is shaky when AI models are giving fake news to users or offering facts that are not correct. If this is constant, users will begin cross-checking AI responses, which defeats the purpose. With this, trust in AI will be diminished. There’s also the fact that AIs giving nonsensical or unhelpful responses will only irritate and alienate users.

Furthermore, many of us turn to AI to get advice or recommendations for everything from food to schoolwork. If AI gives incorrect information, people could end up harming themselves, which is a whole other can of worms.

AI Hallucinations’ Examples

A prime example of an AI hallucination would be the Bard chatbot falsely claiming that the first image of a planet outside of the Milky Way was taken by the James Webb Space Telescope. In reality, the first image was taken in 2004, 7 years before the James Webb Space Telescope was even launched.

Another example is ChatGPT making up fake articles associated with The Guardian newspaper, including a fake author and fake events that never happened.

Or, for instance, Microsoft’s Bing AI insulted and even threatened a user to reveal his personal information and ‘ruin his chances of finding a job’. after it was launched in February 2023.

Detecting and Preventing AI Hallucinations

Because AI is not infallible, both developers and users need to know how to detect and prevent AI hallucinations to avoid experiencing the downsides. Here are a few ways to detect and prevent AI hallucinations:

• Double-check results. If an AI gives you a specific answer to a question, search online to be sure that it is correct. This is doubly important if the information will be used for school or your work.
• Give clear prompts. When dealing with AI try to make your prompts as straightforward and clear as possible. This reduces the chances of the AI misinterpreting it.
• In-depth AI training. If you are developing AI, you must train it based on diverse and high-quality materials and test it as much as possible before it is released to the public.
• Experiment with AI’s temperature. Temperature in AI development is a parameter that determines how random the AI’s response will be, with higher temperatures meaning that hallucinations are more likely. Experiment with temperatures and make sure it is at a safe level before releasing your AI.

Final Thoughts

The more we use AI, the more we become aware of its limitations and issues that still need to be worked out. AI hallucination is a genuine issue within the tech world and one that both creators and users of AI need to be aware of. Whether a flaw in the system or due to prompt issues, AIs can and have given false responses, nonsensical ones, and much more. It is up to developers to work towards making AIs as close to infallible as possible and for users to be cautious as they use AI.