Gemini、ハルシネーションが酷すぎて使い物にならない
Reports indicate that Google's highly anticipated AI, Gemini, is frequently generating "hallucinations" – information that deviates from facts.
Its severe inaccuracy has led to widespread online sentiment that it's "unusable," causing significant user disappointment and disillusionment.
Given the high expectations, this situation is raising serious questions about Google's overall AI strategy.
Related Keywords
What is Gemini?
Gemini is a large-scale generative AI model developed by Google. It is characterized by its "multimodal" capabilities, meaning it can understand and generate multiple forms of information such as text, images, audio, and video. Announced in December 2023, Gemini was positioned as the core of Google's AI strategy, with high expectations. It particularly emphasized natural conversation indistinguishable from human writing and logical reasoning capabilities for complex questions. Google positioned Gemini as the culmination of its AI development and a strong competitor to OpenAI's ChatGPT. However, reports of its actual use indicate that it has not consistently delivered the expected performance. Specifically, the problem of "hallucinations" – generating incorrect or inappropriate content, especially in situations requiring advanced reasoning – has become apparent, detracting from the user experience. Given that it is positioned as the culmination of Google's long-standing search technology and AI research, the instability of its performance has sparked considerable debate. As a crucial product that could influence Google's competitiveness in the AI market, significant improvements are strongly desired for the future.
What is Hallucination (Generative AI)?
In generative AI, "hallucination" refers to the phenomenon where AI generates information that does not exist in its training data or deviates from facts, presenting it as if it were true. It is likened to humans experiencing hallucinations, characterized by content that is completely fabricated or contradicts existing information. Specific examples reported include creating biographies of non-existent individuals, presenting incorrect scientific facts, or generating fictitious citations. The main causes of this phenomenon include biases or deficiencies in training data, a mismatch between the model's knowledge and user requirements, or the limits of the model's ability to confidently respond with "I don't know." For instance, if the AI does not have a direct answer in its training data for a certain question, it might "create" plausible but incorrect information. This phenomenon severely undermines the reliability of AI and makes it difficult to utilize as a decision-making support or information-gathering tool in business. Particularly in fields where accuracy is paramount, such as medicine, law, and finance, misinformation carries the risk of severe consequences. Therefore, the development of technologies like RAG (Retrieval Augmented Generation) that link with external databases to improve accuracy, and the importance of fact-checking by users, are being emphasized.
What is Large Language Model (LLM)?
A Large Language Model (LLM) is an AI model that possesses the ability to understand and generate human-like natural language by learning from vast amounts of text data. Google's Gemini and OpenAI's ChatGPT are prime examples, typically having billions, hundreds of billions, or even more parameters (values adjusted by the model through learning). These models can perform a wide range of tasks based on a given prompt (instruction), such as summarizing text, translating, answering questions, generating poetry or code, and sentiment analysis. The underlying technology has evolved significantly, especially with the advent of the neural network architecture called "Transformer." Transformer enabled efficient learning of relationships between words in a sentence, contributing to the high performance of LLMs. In the LLM training process, the model reads vast amounts of text data from the internet (books, web pages, academic papers, etc.) and statistically learns word appearance patterns and contexts. This allows it to develop the ability to predict the next word, which leads to fluent text generation. While LLMs are expected to have applications in various fields such as business, education, and research due to their versatility and advanced language processing capabilities, they also face challenges that need to be addressed, including the problem of hallucinations, high computational costs, and ethical issues (e.g., generating discriminatory content, copyright concerns).