AI-Hallucinations Credit: deepgram.com
Introduction
AI-Hallucination is a phenomenon whereby a large language model (LLM), ChatGPT or generative AI perceive patterns or objects that are non-existent or imperceptible to human observers and which, as a result, create outputs that are nonsensical or altogether inaccurate. The term may seem paradoxical, given that hallucinations are typically associated with humans. However, machine generated outputs, especially in the case of image and pattern recognition can be truly surreal in appearance. Hence, AI-Hallucinations are similar to how humans sometimes see figures in the clouds or faces on the moon. These misinterpretations occur due to various factors, including bias in training data and high model complexity. As these factors tend to gain momentum the question looms as to how AI will be able to map reality as a benchmark for user acceptance.
Human Hallucinations
According to Wikipedia a hallucination is a false perception that lacks the stimulus of a real perception. Hallucination is a combination of two states of brain activity: wakefulness and sleep. A mild form of hallucination is known as a disturbance. It can occur with our eyes such as seeing an unreal or bizarre movement or hearing faint noises or voices. Auditory hallucinations are also very common in schizophrenia. They may be benevolent, telling the individual good things about themselves or malicious, cursing the individual. In visual hallucinations the source of the visual and imagined object can also be positioned behind the hallucinating individual. This can produce a feeling of being looked or stared at. Frequently, individuals experience auditory and visionary hallucinations at the same time. Research differentiates between hypnagogic and hypnopompic hallucinations. Hypnagogic hallucinations can occur as one is falling asleep while hypnopompic hallucinations occur when one is waking up. Hallucinations can be associated with drug use such as deliriants or sleep deprivation, psychosis, neurological disorders and delirium tremens. The English term ‘hallucination’ was introduced by the 17th-century physician Sir Thomas Browne as a derivation of the Latin word alucinari, meaning to wander in one’s mind or walk idly. For Browne, hallucination means a sort of vision that is “depraved and receives its objects erroneously”, a view that is still correct according to today’s definition.
AI-Hallucination and the Human Connection
By understanding the link between AI’s hallucinatory potential and our own, developers can build better AI-systems that will ultimately help reduce erroneous output. An LLM is not trying to conserve the limited mental resources of humans to efficiently make sense of the world. ‘Hallucinating’, in this context, just describes a failed attempt to predict a suitable response to an input. Nevertheless, there is still some similarity between how humans and LLMs hallucinate. LLMs generate a response by predicting which word is most likely to appear next based on what has come before together with the associations the system has learned through training. Like humans, LLMs try to predict the most likely response. However, unlike humans, they do this without understanding what they are describing. This is the reason how they can end up in generating nonsense. As to why LLMs hallucinate, there are a range of factors to be considered. Compared to AI-machines, humans can only process a limited amount of the information flooding our senses, and only remember a fraction of all the information we have ever been exposed to. Moreover, our biases can result in poor judgement. Consider the automation bias, which is our tendency to favour information generated by automated systems such as ChatGPT over information from non-automated sources. This bias can lead us to miss errors and even act upon false information. Another relevant factor is the halo effect in which our initial impression of something affects our subsequent interactions or the fluency bias, which describes how we favour information presented in an easy-to-read manner. The bottom line is that human thinking is often coloured by its own cognitive biases and distortions while these ‘hallucinatory’ tendencies largely occur outside of our awareness.
Why AI-Hallucinations are a Problem for AI
AI-Hallucinations can have significant consequences for real-world applications. For example, a healthcare AI-model might incorrectly identify a benign skin lesion as malignant, leading to unnecessary medical interventions. AI-Hallucination problems can also contribute to the spread of misinformation. If, for instance, a hallucinating news bot responds to queries with information that has not been fact-checked, it can quickly spread misinformation that is difficult to detect. One significant source of hallucination in machine-learning algorithms comes from input bias. If an AI-model is trained on a dataset that contains biased or unrepresentative data, it may hallucinate patterns or features that reflect these biases. An immediate problem with AI-Hallucinations is that they significantly disturb user trust. As users begin to experience AI as a useful tool, they tend to develop a significant amount of trust about their usefulness and are unpleasantly surprised when that trust is betrayed. One challenge classifying outputs as hallucinations relates to the fact that it encourages anthropomorphism. The attribution of human traits, emotions or intentions to non-human entities such as AI-machines creates confusion and distrust. Moreover, AI-machines, despite their intrinsic capacity of intelligence, are not really conscious as machine consciousness and human consciousness are two separate entities which are intensively discussed within the AI research community. AI-machines do not have their own perception of the world. Their output manipulates the users’ perception and might be more aptly considered a mirage rather than a machine that is hallucinating. Another challenge presented by hallucinations is the newness of the phenomenon which is a typical feature of LLMs. Output of AI-Hallucinations and LLMs are designed to sound fluid and plausible. Hence, if an individual is not prepared to read LLM outputs with a sceptical eye, he might believe that the hallucination is true. As a result, AI-Hallucinations can be dangerous due to their capacity to fool people. A third challenge to identify hallucination correctly is that LLMs often represent a black box of algorithms that are kept secret by the company that provides the software. Therefore, it can be difficult or in many cases impossible to determine why the LLM generated a hallucinatory response. It is the user, not the proprietor of the LLM, that has to watch for hallucinations delivered by the AI-machine.
How to Prevent AI-Hallucination
The most basic method to detect an AI-Hallucination is to carefully fact-check the AI-model’s output. This can be difficult when dealing with unfamiliar, complex or dense material. Users can ask the model to perform a self-evaluation and to determine the probability that an answer is correct or, in contrast, highlight the parts of an answer that might be wrong, using that as a starting point for fact- checking. Users can also familiarize themselves with the model’s sources of information to help them identify sources of potential hallucinations. To reduce the potential of AI-Hallucinations, LLM and generative AI practitioners can apply a number of procedures such as: Use clear and specific prompts because additional context can help guide the model toward the intended output; Select filtering and ranking strategies because LLMs often have parameters that users can tune; Use multishot prompting that provides several examples of the desired output format or context to help the model recognize patterns.
Today, hallucinations are considered an inherent but typically unwanted feature of LLMs. AI-researchers are trying to understand, reduce and control their impact. For example, OpenAI has proposed a strategy to reward AI-models for each correct step in reasoning towards the correct answer instead of just rewarding the conclusion if its result proves to be correct. This approach called process supervision aims to manipulate models into a chain-of-thought approach that decomposes prompts into single steps. While AI-Hallucinations are certainly an unwanted outcome of AI applications, they also offer a range of intriguing use cases that can help organizations and their individuals to leverage their creative potential in positive ways. With the hallucinatory capabilities of AI, artists can produce surreal and dream-like images that can generate new art forms and styles which open up new frontiers of knowledge presentation.
Conclusion
AI-Hallucinations challenge our understanding of reality. They represent a phenomenon whereby LLMs, ChatGPT or generative AI perceive patterns or objects that are non-existent or imperceptible to human observers and as a result, create outputs that are nonsensical or altogether inaccurate. If this trend is getting worse we need to apply common sense to prevent a collapse of useful AI-technology. To many observers we are getting close to an AI-bubble ready to burst with a highly disruptive outcome to producers and consumers alike.