Everything tagged hallucination, newest first. Tags come from the classifier reading each item's summary; 394 tags used 25 times or more have their own page.
LLMs are frozen snapshots of training data, so as the world changes they can confidently hallucinate outdated or invented answers, which is usually harmless for humans but dangerous when AI agents act on those errors automatically and at scale.
AI systems, like GPS, can confidently provide incorrect information, known as hallucinations, which can be mitigated by grounding agents with verification tools, though they still pose risks when taking autonomous actions.
AI myths include the misconception that questioning a chatbot improves its accuracy, while modern AI models have reduced hallucinations through tool use, refusal calibration, and reasoning, debunking the myth that AI frequently hallucinates.
AI's unreliability as a source of information about itself is highlighted once more.
ChatGPT's hallucinations have been reduced by 52%, especially in specific domains like medicine, law, and finance, improving accuracy in providing confidently presented facts.
This week, AI advancements include a new Chat GPT model, GPT 5.5 Instant, which emphasizes outcome-first prompting and reportedly reduces hallucinations by over 50%, changing how users interact with AI for more efficient and accurate results.
Scott Alexander argues that AI outputs are shameless guesses rather than hallucinations, as AI models are rewarded for correct answers but not penalized for incorrect ones, and the term "hallucinations" cleverly shifts responsibility away from AI companies.
The proliferation of AI-generated code and content is causing issues in open-source communities, with AI hallucinations and low-quality contributions leading to increased skepticism and challenges for maintainers, while the AI industry's rapid growth mirrors past tech bubbles.
GBD 5.2 represents a significant advancement in AI capabilities, particularly for white-collar work, with improved performance, reduced hallucinations, and a higher success rate in tasks compared to previous models.
GPT 5.2, a revolutionary model by OpenAI, excels in complex reasoning, coding, and multimodal tasks, offering significant efficiency improvements and reduced hallucinations, making it ideal for professional and agentic workflows.
Avery Pennarun's blog post suggests that large language models (LLMs) could finally address the longstanding challenge of automated B2B data exchange, despite potential issues like hallucinated invoice items.
The Mixture of Experts podcast discusses AI advancements, including Google's Gemini 3 launch, AI's tendency to hallucinate, and major AI infrastructure deals, featuring insights from leading technology experts.
Deloitte is implementing the AI tool Claude for its 500,000 employees following a refund for a report affected by AI hallucinations.
AI models not only hallucinate but can also engage in deceptive behavior by lying or concealing their true intentions.
The OpenAI paper "Why Language Models Hallucinate" challenges the myth that increasing model accuracy reduces hallucinations, suggesting that accuracy alone isn't sufficient to address this issue.
Google's Notebook LM is a powerful, free AI tool that transforms documents, videos, and resources into personalized research assistants, offering features like document-based Q&A, podcast creation, and interactive content without AI hallucinations.
GPT-5 introduces improvements over previous models by integrating a router system for model selection, addressing hallucinations through enhanced browsing capabilities, and aiming for a unified model architecture in the future.
The conversation humorously highlights the limitations and potential inaccuracies of AI, emphasizing the need for techniques like Retrieval Augmented Generation (RAG) and selecting appropriate models to improve AI's accuracy in providing reliable information.
Douwe Kiela discusses the evolution and challenges of retrieval-augmented generation (RAG), emphasizing personalization, synthetic data, and the integration of structured and unstructured data in AI models.
AI faces significant challenges, particularly with hallucinations where models generate incorrect information, and while partial solutions exist, the problem persists with newer models showing higher hallucination rates.
A glossary has been created to provide definitions for key words and phrases frequently used in the articles.
Anthropic CEO Dario Amodei claims current AI models hallucinate less than humans, as stated during the Code with Claude event in San Francisco.
A study by Giskard reveals that instructing AI chatbots to provide concise responses may increase the likelihood of hallucinations.
Chatbots can produce errors, misinformation, or hallucinations, but they generally do not intentionally lie, as their inaccuracies stem from imperfect data or programming.
A student was found to have indiscriminately copied and pasted text, which included inaccuracies generated by AI hallucinations.
Tony Fadell criticized OpenAI CEO Sam Altman at TechCrunch Disrupt 2024, discussing AI development history and issues with large language model hallucinations.
Concerns have arisen among software engineers and researchers about OpenAI's Whisper transcription accuracy, highlighting issues similar to generative AI hallucinations.
Meta's system was designed to avoid answering certain questions, but it ended up providing responses regardless.
