hallucinations

Introduces CanLegalRAGBench, a benchmark for evaluating retrieval-augmented generation on Canadian case law using realistic queries and expert-annotated answers. The evaluation reveals sensitivity to design choices, competitiveness of open-source embedding models, and persistent hallucinations in generated answers.

GPTZero investigated Ernst & Young Canada's cybersecurity report on loyalty fraud and found it contained numerous hallucinated citations and AI-written text, highlighting the epidemic of 'vibe citing' in consulting reports.

A developer shares a personal experience of their AI agent producing hallucinated data and the lessons learned about verification and prompt specificity.

Discusses how to benchmark and grade production builds, focusing on key performance indicators like context-drift, hallucinations, and governance.

Developer built VoiceOBS, an observability tool for AI voice agents, providing latency breakdowns, sentiment analysis, hallucination detection, and more, integrated with Vapi.

The author presents a proof-of-concept showing that using gentle, mistake-tolerant prompts instead of high-pressure authoritarian prompts significantly reduces AI thought loops and hallucinations, leading to faster and more honest responses.

UC Berkeley Law School is banning almost all AI use by students for graded assignments starting summer 2026, citing the need to build core critical thinking skills and avoid AI errors and biases.

This paper presents a function-centric framework using Transcoders to trace computational pathways in vision-language models, demonstrating stronger attribution of visual grounding and the ability to predict hallucinations via graph-based features.

The article discusses the industry consensus that AI is becoming extremely capable but still faces reliability issues for high-stakes tasks, emphasizing that current systems optimize for plausibility rather than guaranteed truth, and that the path forward involves layered verification systems rather than a single perfect model.

Tim O'Reilly discusses the challenges of integrating AI into scientific publishing, including hallucinated citations, propagation of retracted papers, and training on compromised literature, and calls for adapting existing scientific infrastructure for AI use.

Developer shares struggles debugging AI agents in production, highlighting issues with hallucinations, regression from prompt changes, and high API costs, asking the community for strategies.

The article critiques how AI systems, particularly Grokipedia and AI search, perpetuate errors by merging unrelated communities due to English-centric transliteration and biased training data. It highlights the systemic issue of erasing cultural distinctions through simplified English representations and repeated misinformation.

The article criticizes the reliance on Large Language Models for generating bibliographic entries, highlighting issues with hallucinated citations and incorrect author lists in academic papers.

This paper investigates how supervised fine-tuning (SFT) increases hallucinations in LLMs by causing knowledge degradation and proposes a self-distillation-based method to mitigate this issue while preserving pre-existing factual knowledge. The authors identify semantic interference among overlapping representations as the primary mechanism behind SFT-induced hallucinations and demonstrate solutions including parameter freezing and self-distillation.

OpenAI publishes research explaining that language models hallucinate because standard training and evaluation procedures reward guessing over acknowledging uncertainty, and proposes that evaluation metrics should prioritize honesty about limitations over raw accuracy.