In the field of artificial intelligence (AI), a hallucination is a response generated by an AI that is not justified by its training data. Hallucinations can occur in various forms, whether they are visual, auditory, or statistical. AI hallucinations gained prominence in 2022 alongside growing public exposure to large language model (LLM) technologies like ChatGPT.
Users initially reported minor offenses, such as the generation of a random number in place of a statistic. However, upon further experimentation, users began to report that chatbots would seemingly embed ‘plausible-sounding falsehoods’ into content at random.
Large language models (LLMs) utilize massive datasets of digital text pulled from the internet to study how humans construct sentences. As an LLM pours through information, it notes patterns in the text that allow it to generate seemingly human responses, analogous to an advanced version of Google Autocomplete. However, the wealth of misinformation online sometimes misleads chatbots, leading them to occasionally pull inaccurate statistics or make unsupported, even paradoxical, claims.
Additionally, the average generative pre-trained transformer (GPT) query processes a vast quantity of information. For example, OpenAI’s GPT-4 was trained on 570 Gbps of text data containing billions of applicable patterns. A chatbot combines these patterns, often unexpectedly, to generate results. This means that even if a chatbot is only trained on factual source material, it is still capable of generating responses that are untrue.
In an internal document cited by the New York Times, a Microsoft representative explained that such early AI systems are “built to be persuasive, not truthful.”
What do AI hallucinations mean for service providers?
For the average consumer, it is good practice to avoid asking chatbots a question you do not know the answer to or have no way of confirming. However, for service providers who seek to capitalize on the development and application of new AI technologies, these malfunctions have already demonstrated their volatility.
RELATED: Oops! Google's new AI tool Bard showcases artificial stupidity
During the launch of Google Bard in early 2023, the chatbot cost the company nearly $100 billion in shares after it answered a question incorrectly. The error also called Google’s previously uncontested dominance over the search engine market into question. It wasn’t until days later, during the launch demo of Microsoft’s Bing AI, where the chatbot made several factual errors in its analysis of earnings reports, that experts named this phenom “hallucination.”
Despite concerns about hallucinations, the AI software market is forecast to grow at a compound annual growth rate (CAGR) of 18.4% by 2025, according to IDC. Additionally, the AI investment market is poised to grow in the coming years, reaching $200B by 2025 globally, according to Goldman Sachs economists Joseph Briggs and Devesh Kodnani. The economists also claim that generative AI could raise global GDP by 7% across a 10-year period.
What solutions are used to combat AI hallucinations?
Companies such as OpenAI and Microsoft continue to innovate new solutions to combat hallucinations in AI deployments. OpenAI utilizes a technique called reinforcement training, where human testers are encouraged to rate ChatGPT’s responses to identify instances of artificial hallucination.
Microsoft, which utilized OpenAI’s GPT-4 framework to construct its Bing chatbot, integrates search engine technology to increase accuracy. Improving the system through the additional process of preliminarily searching the desired subject on Bing’s web browser and incorporating these results into the chatbot’s final response.
Read more of our cloud explainers here.