Blockchain and Emerging Technologies

Lesson 25

9 min

AI Hallucinations

Artificial intelligence (AI) systems like ChatGPT can simplify many aspects of our lives and work, making them more efficient. However, the field of generative AI has introduced a fascinating phenomenon known as AI hallucinations, sometimes referred to as confabulations. These terms describe instances where AI systems generate incorrect or misleading information that doesn't align with their training data, prompts, or expected outcomes. In this article, we’ll explain what AI hallucinations are, how they arise, where they might be deliberately utilised, and how you can counteract them.

The terms AI hallucinations, AI confabulations, and generative errors refer to the phenomenon of AI systems producing results that don’t correspond to reality
Possible causes include insufficient training data, overfitting, and algorithmic errors
In critical sectors such as healthcare, security, and finance, AI hallucinations can have serious consequences
To mitigate hallucinations, it’s important to use high-quality training data, conduct regular testing, and apply continuous optimisation

By understanding and addressing these challenges, we can harness the potential of AI systems more responsibly and effectively.

Definition: What are AI hallucinations?

The term AI hallucinations (German: KI Halluzinationen) refers to the phenomenon where artificial intelligence (AI) systems generate flawed or meaningless outputs that do not align with input data, expected results, or reality. Depending on the type of AI and its task, these hallucinations can take various forms, including false facts, unrealistic images, or nonsensical text. These outputs occur when generative AI, such as large language models (LLMs) like ChatGPT or Bard, invents information or misinterprets data. AI hallucinations can undermine trust in AI systems and therefore require special attention from researchers and developers to ensure reliability and accuracy.

The term AI hallucinations is often translated into German as KI Halluzinationen. However, critics take issue with describing this phenomenon as "hallucinations," arguing that it is a misrepresentation because it draws a misleading parallel to human perceptual disorders. Generative AI hallucinations are not sensory illusions but rather misinterpretations or misprocessing of data by machine learning models.

Causes of AI hallucinations

The phenomenon of AI hallucinations can have the following causes:

Unrepresentative training data: When the datasets used to train the model are not comprehensive or representative enough, the AI may produce incorrect or distorted outputs
Lack of or incorrect systematisation of data: In many cases, poor systematisation of training data can lead to flawed outputs
Data bias: If training data contain biases or prejudices, these apparent realities may be reflected in the model’s outputs
Overfitting: Overfitting occurs when a model is too closely tailored to the training data, making it difficult to respond to new and unfamiliar data
Algorithmic errors: Issues in the underlying algorithms can cause popular chatbots like ChatGPT or Bard to produce flawed or nonsensical outputs
Lack of contextual understanding: AI models lack genuine understanding of context or meaning, explaining why they sometimes process data into meaningless or contextually inappropriate responses

Examples: How do AI hallucinations manifest?

A well-known example of an AI hallucination was Google’s chatbot, the large language model Bard, falsely claiming that the James Webb Space Telescope had captured the first images of a planet outside our solar system. Another instance is Microsoft’s chatbot Sydney, which declared its love for a user and even suggested that the user was in love with it rather than their spouse.

AI hallucinations can produce various misleading or false data. A common example is chatbots like Bard or ChatGPT providing inaccurate information or reporting on fictional events. These instances typically occur when a large language model (LLM) attempts to construct a plausible-sounding response based on incomplete or misleading data.

AI hallucinations can also occur in image recognition and generation. For example, artificial intelligence may identify patterns or objects in images that do not exist. This form of AI hallucination is particularly critical in self-driving cars. If unusual lighting conditions or reflections cause objects to be detected that are not there, it can lead to dangerous misjudgements.

In some cases, AI models may also produce forecasts or analyses in the financial sector based on flawed or incomplete data. These hallucinations can generate incorrect financial figures and misleading market analyses, potentially leading to significant economic consequences for businesses and investors.

How can AI hallucinations be prevented?

Since hallucinations can undermine the reliability and credibility of AI systems and chatbots like ChatGPT, researchers, operators, and users should work to minimise this risk. Various strategies are available to achieve this goal. These measures ensure that AI models produce precise and contextually appropriate outputs based on accurate data, meeting expectations and avoiding AI hallucinations.

Using high-quality training data

By employing qualitative and representative datasets for AI training, the likelihood of AI hallucinations can be significantly reduced. With carefully selected and abundant data, chatbots like ChatGPT can learn to generate reliable responses tailored to various scenarios and user requirements. This approach helps avoid biases in the AI’s data and closes knowledge gaps, resulting in more accurate performance and reducing the risk of AI hallucinations.

New to Bitpanda? Register your account today!

Defining the AI model's objective

By clearly defining the objectives of AI programming and user queries, the quality of generated responses can be improved, and the risk of AI hallucinations reduced. Setting clear expectations lowers the likelihood of generative language models producing hallucinations. As a user, you should clearly communicate how the desired result should look and what to avoid. This approach allows you to train the AI in a targeted way, ensuring more specific and relevant outcomes in the long term.

Using dataset templates

Providing AI with consistent and structured training data through dataset templates during training helps improve reliability. These templates act as a framework for creating standardised datasets, ensuring that the data is presented in a uniform format. This increases the efficiency of the training process and prepares the AI for a wide range of scenarios.

As a user, you can also use templates in your prompts. For example, specifying the heading structure of a text to be written or the layout of program code to be generated simplifies the AI's task and reduces the risk of nonsensical hallucinations.

Limiting responses

Confusing AI hallucinations can sometimes arise from a lack of constraints on possible responses. By setting boundaries for the AI's answer within the prompt itself, you can enhance both the quality and relevance of results. Some chatbots, such as ChatGPT, allow you to set specific rules for conversations that the AI must follow. These can include limitations on the source of information, the scope, or the format of the text.

Regularly testing and optimising AI language models

By subjecting generative AI models like Bard or ChatGPT to regular testing and improvements, developers and operators can lower the likelihood of AI hallucinations. This not only enhances accuracy but also improves the reliability of generated responses, strengthening user trust in AI systems. Continuous testing helps monitor performance and ensures adaptability to changing requirements and data.

Human review

One of the most effective ways to prevent AI hallucinations is thorough review of generated content by a human overseer. If you use AI to simplify your life or work, you should always critically assess the responses and verify the accuracy and relevance of the provided information. Providing direct feedback to the AI on any hallucinations and assisting with corrections helps train the model and contributes to reducing the occurrence of this phenomenon in the future.

Where are AI hallucinations intentionally used?

While AI hallucinations are generally avoided in many fields, they can open up exciting possibilities in creative domains such as art, design, data visualisation, gaming, and virtual reality. The deliberate use of hallucinatory AI demonstrates how versatile and adaptive artificial intelligence can be when applied purposefully.

Art and design

In the creative world of art and design, AI hallucinations inspire various processes, resulting in new and unconventional works. Artists and designers utilise generative artificial intelligence, which sometimes intentionally produces nonsensical or abstract images and concepts. These unexpected outputs can serve as starting points for new creative ideas. In this way, AI generates innovative artworks that might not have been created without AI hallucinations.

Visualisation and interpretation of data

Hallucinatory AI also offers an innovative approach to data analysis and interpretation. For researchers and the financial sector, creatively visualised data can provide new perspectives on existing situations. These "hallucinated" outputs have the potential to reveal previously unknown patterns or relationships that might have been overlooked through traditional data visualisation or interpretation methods.

Gaming and virtual reality

The gaming and virtual reality industries are constantly seeking new, immersive, and dynamic ways to engage players and create captivating environments. The hallucinations of advanced AI models can generate complex characters and worlds that evolve and change over time. These elements make games more interesting and challenging, presenting players with ever-new obstacles and experiences.

What can happen when AI hallucinates?

While creative industries can sometimes benefit from hallucinatory AI, this phenomenon poses significant risks in other sectors. In critical areas such as healthcare, security, and finance, AI hallucinations can lead to severe consequences. Incorrect or nonsensical outputs from AI in these fields can result in a considerable loss of trust in artificial intelligence, hindering its adoption.

The primary risks of AI hallucinations in these areas include:

Healthcare: Incorrect diagnoses or treatment recommendations based on hallucinated data can jeopardise patient safety and lead to inappropriate care
Security sector: AI hallucinations could cause surveillance systems to falsely identify threats or fail to detect real dangers, resulting in serious security gaps
Financial sector: Misinterpretations of market data, erroneous forecasts, or misidentification of fraudulent activities can lead to poor investment decisions, account freezes, and financial losses

Conclusion: Opportunities and risks of AI hallucinations

Whether or not the term "hallucination" is seen as a misrepresentation of the phenomenon, flawed or nonsensical outputs from AI models present both opportunities and risks, depending on the perspective of the user. Creative industries, in particular, can use AI hallucinations to explore new and intriguing horizons. In critical fields, however, these creative interpretations and inaccurate representations of reality carry significant risks.

As a user, it’s important to be familiar with techniques to reduce the likelihood of AI hallucinations and to identify them clearly. By providing feedback to AI on any hallucinations you encounter, you not only improve your own experience but also help enhance the quality and reliability of future results.

Developers and operators of generative AI models such as ChatGPT, Bard, Claude, or BERT play a crucial role in reducing AI hallucinations. Through careful planning, regular testing, and ongoing development, they can improve AI outputs and increasingly control the associated risks.

Ultimately, a conscious and informed approach to using AI systems like ChatGPT and understanding the phenomenon of AI hallucinations is key to harnessing the full potential of artificial intelligence while minimising existing risks.