Hallucination in AI Chatbots: Can It Really Be Fixed?

Quincy Jon, Tech Times 01 August 2023, 03:08 pm

Large language models (LLMs) have propelled chatbots like ChatGPT and Google Bard, bringing forth notable improvements in artificial intelligence. These advanced AI systems do have certain drawbacks, however. One major issue is AI hallucination, in which LLMs provide false or deceptive data.

LLMs are made to create fluid and coherent prose, giving the impression that hallucinations are real. These logical or factual inconsistencies arise because the AI does not comprehend the underlying reality that language expresses. LLMs instead depend on statistical trends to produce grammatically and semantically sound content within the given context, according to Tech Target.

Although hallucinations are common in LLMs, it is still difficult to pinpoint the precise origins of these misleading outputs on a case-by-case basis. These hallucinations provide significant issues for enterprises, organizations, and high school students employing generative AI for document authoring and activities with high-stakes outcomes, such as in psychotherapy and legal brief writing.

A Tough Challenge

The developer of chatbot Claude 2, Anthropic, admits the occurrence of hallucinations while highlighting that existing algorithms are mainly built to anticipate the following phrase, which may cause errors. Major AI system creators like Anthropic and OpenAI are actively attempting to increase the veracity of their models. These endeavors are currently being assessed for efficacy.

Emily Bender, linguistics professor and leader of the University of Washington's Computational Linguistics Laboratory, believes that AI technology's mismatch with its intended use cases causes AI hallucination. The issue raises concerns regarding the dependability of generative AI technology, particularly in light of the anticipated economic effect, which the McKinsey Global Institute estimates to range from $2.6 trillion to $4.4 trillion, per AP News.

Google has previously offered news companies, where accuracy is crucial, an AI tool that can write news stories. Together with OpenAI, The Associated Press is researching AI technologies to enhance its systems. While Ganesh Bagler, an Indian computer scientist, is developing AI algorithms to create recipes for South Asian foods. He remarked that the results of hallucinated components might provide delicious or unconsumable dinners.

Humans Should Always be Involved

The creator of the family concierge service Yohana, Yoky Matsuoka, cautions against completely outsourcing employment to AI. She argues that because of AI hallucinations that produce inaccurate results, humans should be included in the process to double-check and rectify their work. Matsuoka also emphasized how crucial it is to keep in mind that AI is being created "for humans."

The National Academy of Sciences president, Marcia McNutt, stresses that AI should be used as a second opinion, assisting people in making choices rather than making decisions for enterprises, according to Forbes.

The goal of eliminating AI hallucinations is challenging and continues as generative the technology advances. In order to fully realize the transformational potential of AI, it will be essential to strike the correct balance between its capabilities and human control.