AI Hallucinations Explained: Understand Why Your AI Makes Up Things

AI making headlines with its blunders is not something; there are enough cases of it giving nonsensical and inappropriate outputs. 51% of organizations using AI experienced at least one negative consequence, and nearly one-third of all respondents reported consequences from AI inaccuracy. All of these are the consequences of AI hallucination. 

In this blog, we will discuss everything about AI hallucinations, why they occur, and what you can do to mitigate them. 

What is an AI Hallucination?

AI hallucinations are incorrect or misleading answers/outputs given by AI models. This can be caused by several factors, such as insufficient training data, false assumptions made by the model, and biased training data. In a single word, it means “Your AI just makes stuff up.”

These AI mistakes can have serious consequences when making important decisions in fields like healthcare, transportation, and finance. 

Types of AI Hallucinations

These are some common types of mistakes AI can make:

1. Made-up facts: AI can generate situations, cases, events, facts, stats, or information that do not really exist. 
2. Incorrect predictions: For example, if you ask AI Will it rain tomorrow? It might say yes, but the rain didn’t occur the next day. 
3. False references/citations: The AI can make up sources, texts, and legal and academic citations. 
4. Wrong facts: It may give you incorrect details about a topic, for example, a wrong date of an historical event.
5. Illogical outputs: AI can also give you logical hallucinations, like a fact that can be valid but the logic to validate it can be nonsensical. 
6. Incorrect numbers: When you look for stats or reports on AI, it gives you figures, but many times, those are different from the actual numbers in the reports. It can give you miscalculated figures by taking out the average on its own, percentages, or measurements. 

AI Hallucination Case Studies

There are ample examples of how AI has made big goof-ups, ranging from fake scenarios to generating inappropriate images of people and more. Here I have shared a few real-life cases where AI made severe mistakes.

1. ChatGPT Made a Fake Child Murder

It is a case where the Norwegian user Arve Hjalmar Holmen wanted to know if ChatGPT contains any information about him. When he began searching, the man encountered a made-up horror story: Where the AI described Holmen as a murderer of his two children and planning to attempt to murder his third child as well. And to make it worse, the chatbot included real-life elements from his personal life. Additionally, it also said that the user was sentenced to 21 years in prison. 

Hjalmar Holmen quoted that “Some think that ‘there is no smoke without fire’. The fact that someone could read this output and believe it is true is what scares me the most.”

2. US Lawyers Submitted Fake Cases Using ChatGPT

In 2023, a lawyer in New York used ChatGPT to find past cases for a lawsuit, and it gave him several case names with legal citations. The cases seemed completely real, so he filed them in the court. When the judge went through the cases, it turned out that the cases were fake; they do not exist in real life, and AI has invented them all. 

3. Mata vs Avianca Inc.

Mr. Mata Roberta, a passenger of Avianca Inc. (a Colombian airline), was hit on the knee by a serving cart and got seriously injured. Thus, he wanted to claim damages under the Convention for the Unification of Certain Rules for International Carriage by Air. When he sued the airline, the lawyer used ChatGPT for legal research, and it generated several fake cases to support the argument. The court could not locate them, and it turned out that ChatGPT had ‘hallucinated’ them. The lawyers were fined $5,000 by the court. 

Cause of AI Hallucinations

There are various reasons for AI, as these are both technical and systemic. 

Data Issues

There can be various issues in data, like poor data quality with missing facts, context, or languages. This makes AI take improvised decisions. Biases in data are another issue that can give stereotypical outputs. And domain blindness means that most models are experts when asked about things beyond their area of expertise, but the error rate increases. 

Model Limitations

Issues like overfitting, where models memorize the exact examples they were trained on instead of learning the general rules. It can make mistakes when dealing with unfamiliar inputs. Creative completion can be another model issue, because gen AI focuses on predicting the next word or sentence based on patterns in its training data

Input Issues

The ambiguity in prompts can lead AI models to guess and generate incorrect details. Also, longer prompt lengths or multi-part prompts increase the chances of hallucination, especially in cases where models need to maintain complex reasoning capabilities without failing. 

Retrieval Time Latency

Some models combine pre-trained data with external data sources using RAG techniques. The limitations due to retrieval failures or the use of less authoritative sources may introduce new inaccuracies into outputs. 

Human Errors

Incorrect or misleading output must be caught and corrected during reinforcement learning, if not, they may train the model to repeat those errors. 

How to Reduce AI Hallucinations

Here are some of the best ways you can reduce hallucinations in AI. 

• Focus on providing high-quality training data; make sure to have regular audits, cleaning, and expansion of data sets. If your industry is at high risk of hallucinations, use expert-labelled data with regular real-life checks. 
• RLHF (reinforcement learning with human feedback) designs systems that learn from human validations. Particularly for outputs that are highly sensitive. 
• Incorporate RAG (retrieval-augmented generation) as hybrid models combine the training data with trusted external datasets to increase accuracy and lessen hallucinations. Also, prompt models to cite supporting information to allow traceability and review. 
• Enhance prompt engineering. Design prompts that clearly state the context, scope, and reference requirements. Build mechanisms to signal when prompts are outside of AI domain expertise. You can do this by warning or escalating the prompt to human review. 
• Improve models and their transparency by making AI systems that can identify uncertainty or the likelihood of hallucinations in their input. Clearly tell users about the pitfalls, possible mistakes, and necessary updates in new versions. 
• Implement automated detection tools that will help you find internal issues to flag hallucinations, while third-party checks/audits benchmark models to measure hallucination rates and reliability across outputs. 
• Always involve human intervention before outputs come to public use to make sure they comply with safety standards. Make use of hybrid systems that blend automated validation with trained human judgement to reduce errors and improve accuracy. 

Conclusion

AI hallucinations highlight the limitations of how AI models work; they can offer fast and smooth responses, but the question is about the accuracy and reliability of the outputs. There are several types of errors that AI can make that we discussed above. You can take measures like using improved prompt engineering, standard data, human reviews, automation to flag hallucinations, and more to curb these hallucinations. We also saw how these errors can cause big damage to the reputation of an organization and an individual. 

So the challenge is not to make smarter AI systems, but the most trustworthy and reliable ones. 

Related: What is AI Ethics? Benefits and How it Works