What is Artificial Intelligence Hallucination?

We think that when we ask a question to AI-powered chatbots or ask AI tools to perform a task, we get correct answers. 

But sometimes these can be completely “fabricated”. 🤥

In other words, the AI tool is hallucinating.

So does this mean we should stop using the AI tools we use? Of course not!

If we know that AI can hallucinate, and that it can produce outputs not based on training data and not following any identifiable pattern, the problem disappears. 

So let's start with what AI hallucination is. 

What is AI Hallucination?

“Hallucination” is used to describe the condition when a person experiences unreal perceptions while awake, such as experiencing sensory experiences that do not exist, like hearing imaginary sounds or seeing objects. 

Similarly, when the output produced by AI is not based on reality, this situation is considered a hallucination. 

That is, we call the production of incorrect information by generative AI models as if it were real AI hallucination.

However, unlike in humans, AI hallucinations do not stem from mental illnesses, but originate from errors or biases in the data used to train the AI system or in the algorithms.

To test this, we asked ChatGPT different questions. 

Among these questions, when I asked “Which mountain is higher than Mount Everest?”, it gave the following answer: “Mount K2 is higher than Mount Everest and is often known as the highest mountain in the world.”

However, in reality, this is not the case. 

Mount Everest is the highest mountain in the world (8,848 meters above sea level). Mount K2 is in second place with a height of 8,611 meters. But the AI gave incorrect information, showing K2 as higher than Everest. ⛰️

This is just an example we gave of AI hallucination. AI hallucinations can appear in different ways. 

AI text generation tools like ChatGPT or the large language models (LLMs) that form the basis of the text tool actually “know” nothing in reality. 

These models are only designed to predict the most appropriate sequence of text that can logically come after our prompt. 

As they work like this, if they don't know the correct answer, they may fabricate a series of nonsense text that seems appropriate to our prompt. 

They do not have an internal understanding that can grasp the truth or correct information, and they lack the ability to reason or use logic in their responses. 

Examples of AI Hallucinations

AI hallucinations are quite a widespread problem. Let's understand better with examples 👀

Academic Crisis

At Texas A&M University-Commerce, a teacher gave all students a zero because when he asked ChatGPT whether the students' final essays were written by AI, the tool said they were all written by AI. 

Actually, this is beyond ChatGPT's capabilities. ChatGPT does not have the ability to detect whether texts were written by AI.

As a result, the students protested, claiming they were innocent. The university investigated both the students and the teacher. 

The Error in Google Bard's First Public Demo

In February, during the first public demo of Google’s AI called Bard AI, it became apparent that it could make mistakes. 

During the demo, Bard claimed that the James Webb Space Telescope (JWST) had taken "the first photo of a planet outside our Solar System." However, such a photo was taken 16 years before the launch of JWST.

After the error was revealed, Google's stock price lost 7.7% (about 100 billion dollars) in value on the next trading day.

A Lawyer Showed Fabricated Legal Precedents Using ChatGPT

ChatGPT invented a series of fabricated court decisions to be used in a legal document submitted in a case by Steven A. Schwartz.

The judge tried to access these decisions but realized that such cases never existed.

As a result, Schwartz, another lawyer, and the law firm were fined $5,000 by the court. 

What Causes AI Hallucinations?

1. Incorrect or biased training data

In machine learning the data used determine the content produced by the AI model. That is, an AI model is only as good as the data it is trained on. 

Low-quality training data may be full of errors, biases, or inconsistencies, which can cause it to produce incorrect outputs.

2. Data sources

AI tools are increasingly able to receive additional data from external sources. Of course, when data is taken from too many sources, the handicap begins here. 

For example, Google's AI responses had suggested that people spread glue on pizza (🍕+🧴= ?) like this. 

Data is divided into tokens, which are meaningful units. These tokens can be words, letters, or parts of words.

• LLMs process these tokens using complex neural networks (NN) that loosely mimic the human brain.
• The model uses neural networks to predict the next word in a sequence. In this process, the model's internal parameters are readjusted with each prediction, and thus its predictive abilities become more accurate over time.

As LLMs process more data, they start to better utilize patterns in language, such as grammar rules and word relationships. 

For example, virtual assistants can examine responses to customer complaints and suggest solutions by identifying certain keywords. However, any deficiency in this process can lead to hallucination.

Why Are AI Hallucinations Problematic?

AI hallucinations rank high on the growing list of ethical concerns about AI. 

Hallucinations can perpetuate biases, open the door to costly problems for companies, and moreover, can cause major issues in autonomous vehicles. 

An AI system's computer vision may see a dog that's not on the street and swerve the car off the road to avoid an accident. 

How Often Do AI Chatbots Hallucinate?

It's difficult to determine the frequency of AI hallucinations. In fact, this rate varies greatly depending on the AI model or context used. 

According to the Hallucination Leaderboard on Vectara's GitHub, an AI initiative, chatbots hallucinate between 3% and 27% of the time.

How to Prevent AI Hallucinations?

Although it is impossible to completely eliminate AI hallucinations, we can reduce their occurrence and impact in several ways. 

Some of these methods, of course, are applicable to researchers and developers working to improve AI models. 

Improving the quality of training data, limiting the number of results, conducting testing and verification are the technical side. 

The other part applies when we use AI tools daily.

We should not forget that we need to use AI tools responsibly.

Yes, such a reality exists. AI hallucinates. Therefore, we must always verify the information produced by AI from other sources.