Recently, Google announced the launch of a new open-source AI model, DataGemma, aimed at addressing the 'hallucination' problem that large language models (LLMs) often encounter when handling statistical data. This hallucination phenomenon causes the model to potentially provide inaccurate answers to questions related to numbers and statistics. The release of the DataGemma model marks a significant advancement for Google in the field of AI. Image source note: Image generated by AI, image licensed by service provider Midjourney to reduce hallucinations during statistical queries.

Cornell University and other institutions conducted hallucination benchmarking tests on generative AI models like GPT-4o, Claude, and Gemini, finding that these models produce hallucination-free text only about 35% of the time, indicating that the reliability of AI still has room for improvement. The research involved fact-checking through authoritative sources across topics like law, health, and history, designing question sets that included content not covered by Wikipedia. The results showed that OpenAI's models performed the best overall, but the improvement compared to GPT-3.5 was limited. The study pointed out that model size is not related to the frequency of hallucinations.

Two papers reveal that content generated by generative AI may be regarded as 'nonsense', highlighting the challenges posed to public servants due to its disregard for accuracy. The erroneous information produced by generative AI differs from intentional lies, and the papers argue that using the term 'nonsense' is more precise, aiding the public's understanding of technology. Currently, AI laws and regulations are incomplete and cannot effectively prevent the harm caused by such misinformation. The authors suggest the introduction of regulations similar to those in the publishing field, emphasizing the avoidance of 'casual statements' that could cause social harm, ensuring that truth is not compromised.

Meta's AI assistant falsely reported a nonexistent assassination attempt on former President Trump, leading to widespread discussion and regret from Meta executives. The misleading response was caused by programming limitations initially set to avoid addressing questions about Trump's assassination, but after these restrictions were removed, the AI still provided incorrect information in some instances. This phenomenon, known as 'hallucination,' is a common challenge faced by generative artificial intelligence. Similar issues have also plagued Google, whose search autocomplete feature was criticized for censoring Trump-related queries.

In this era of information explosion, content creators are extremely busy. According to a survey, most creators spend more than 20 hours per week producing new content for their niche. As a result, the time left for them to interact with their fans is minimal. However, Mark Zuckerberg, the head of Meta, believes that artificial intelligence (AI) could be the "lifeline" to solve this problem.Image source note: Image generated by AI, image authorized service provider MidjourneyRecently, Zuckerberg

While large language models like ChatGPT, Claude, and Gemini are impressive, they all share a significant issue: they often generate hallucinations. This is a serious problem in the field of artificial intelligence, and even Apple has expressed concerns about how future Apple Intelligence will handle hallucinations. Fortunately, a team of researchers has now developed an AI hallucination detector that can determine whether the AI is fabricating content.

Dictionary.com has chosen 'hallucination' as the Word of the Year for 2023, highlighting the widespread application of artificial intelligence. Search data shows a 45% increase in searches for 'hallucination', with overall AI-related terms growing by 62% year-on-year. Google CEO Sundar Pichai noted that the 'hallucination problem' in AI remains unresolved. The term originates from Latin, underscoring the ongoing influence of artificial intelligence on societal changes.

Galileo Lab launches new metrics to help improve the accuracy of generative AI. OpenAI GPT-4 has a 23% tendency for hallucination in basic question-answering tasks. The new metrics from Galileo Labs include correctness and contextual consistency. These new tools can help businesses test and mitigate hallucination issues in AI implementations.

OpenAI GPT-4 stands out on the Vectara leaderboard with only a 3% hallucination rate in document summarization and an accuracy rate of 97%. The Vectara evaluation reveals that GPT-4 Turbo and GPT-4 perform the best, followed by GPT-3.5 Turbo, while the highest-rated non-OpenAI model is Meta Llama. Vectara has released open-source models that can check the hallucination rate of language models to enhance the credibility of generative AI systems. Google's models performed moderately on the leaderboard.

Cambridge Dictionary has announced 'Hallucination' as the Word of the Year for 2023, as AI technology redefines the concept. AI tools like ChatGPT have sparked significant interest but have also raised questions about the reliability of AI-generated text. AI 'hallucinations' are having real-world impacts, involving fabricated legal cases and other practical situations. An AI ethicist from Cambridge University points out that AI 'hallucination' anthropomorphizes artificial intelligence.

The Galactica model from Meta was taken offline just three days after its release due to hallucination issues, with LeCun expressing regret. Galactica is regarded as an early attempt at large language models, positively influencing Meta's subsequent projects. The model's lead author, Ross Taylor, admitted that the foundational model was not thoroughly checked prior to release, but he still emphasized its research value. Joelle Pineau, Vice President of AI Research at Meta, highlighted that Galactica was a research project, not a product. The removal of Galactica has sparked significant discussion.

LexisNexis first revealed its Lexis+ AI initiative last May, utilizing large language models to answer legal questions. The company has lifted restrictions on Lexis+ AI, making it widely available to clients in the U.S. Lexis+ AI provides reliable results with 'hallucination-free' legal citation links. It enables conversational searching, document drafting, case summarization, and document analysis. LexisNexis claims that Lexis+ AI can enhance the speed of legal professionals.

["Experts are concerned about the hallucination phenomenon present in popular AI platforms like ChatGPT.", "AI hallucination is a phenomenon where information appears correct but is actually inaccurate.", "ChatGPT has been pointed out for providing false information, raising the attention of scientists."]

Fujitsu has announced the launch of two new AI trust technologies aimed at enhancing the reliability of conversational AI models, including technologies for detecting hallucinations and adversarial attacks. These technologies will be integrated into Fujitsu's AI platform 'Kozuchi', designed to provide tools for businesses and individual users to assess the reliability of conversational AI responses, thereby enhancing AI security. The first technology aims to detect hallucinations in conversational AI responses with high precision, addressing potential misinformation issues in those replies. The second technology is used to detect conversational...

US AI startup Galileo has launched a new tool to help users explain the hallucination phenomenon of AI large models. Galileo Studio features real-time monitoring capabilities to optimize LLMs performance. Galileo introduces new metrics and safeguards to assess the authenticity of model outputs, addressing hallucination issues, and ensuring LLMs produce accurate and compliant results.

Large model hallucinations can be mainly divided into three categories: those that conflict with input, context, and facts. Compared to traditional models, large model hallucination evaluation faces new challenges such as large data scales, strong generalizability, and subtlety. Mitigating hallucinations can be approached from pre-training, fine-tuning, and reinforcement learning, but reliable evaluation remains a challenge. Overall, the evaluation and mitigation of large model hallucinations require further research to promote the practical application of large models.