Tag
WIRED reports that mainstream AI chatbots frequently deliver inaccurate, outdated, or partisan answers about the US midterms, with research showing nearly 40% of election responses contain factual errors and chatbots often cite 'pink slime' partisan sites.
Trakkr Research tested 1,285 statistics cited by ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews, finding only 40% traced back to a study saying the same thing, while many 'zombie stats' circulate with no research behind them.
该论文提出 PERPL 框架，通过 LLM 模拟读者的相对偏好来学习个性化事实核查解释改写，在保持事实忠实性的同时优于更大的通用 LLM。
Wedgewood Polls admitted its polling was fabricated and framed it as a social experiment, and the University of Washington's Center for an Informed Public analyzed how the confession was quickly amplified and reshaped into partisan conspiracies and potential 'rigged election' narratives across X.
作者利用 Apple 设备内置的 3B 参数本地模型（通过 Apple Neural Engine 运行）为其 OpenClaw 智能体团队构建了一个答案校验系统，用以检测并拦截工具输出中的幻觉或错误回答，且不占用主模型算力和带宽。
The paper introduces ReviewFact, a benchmark for labeling factual errors in human and AI-written peer review comments, and PremiseCheck, a framework combining LLM premise rewriting with decision models that matches GPT-5.5 accuracy at roughly half the cost across NLP, ML, and natural sciences.
The paper evaluates Jev, a non-generative decision model, against a frontier LLM (gpt-6-astra) for RAG claim verification on 495 LLM-AggreFact claims, finding statistically indistinguishable headline accuracy at a 187-fold cost difference but highly unequal error profiles, and showing that a confidence-gated cascade can cut costs without beating a matched random-escalation control.
This paper studies how multimodal claim verification models respond to stylistic text changes induced by LLM rewriting. Evaluating 11 open-weight VLMs (2B–38B), the authors find accuracy is largely robust to natural rewriting and controlled LLM-word injection, though hedging-oriented modifications cause consistent probability shifts across nearly all models.
该推文盘点了 10 个用于让 AI 智能体为其输出提供可验证证据的开源 GitHub 项目，涵盖科学文献检索（OpenScholar、PaperQA2）、文档解析（Docling）、网络爬取（Crawl4AI）、RAG（RAGFlow、GraphRAG、LightRAG）、深度研究（Open Deep Research）以及事实核查与评估（DeepEval、Ragas），并重点展示了 Allen AI 的 OpenScholar 代码库。
The Trump administration's newly launched America.gov AI chatbot was reprogrammed within a day of its release to stop fact-checking false claims from the president, refusing instead to answer political questions. The changes gutted the tool's usefulness for answering questions about federal services and policy impacts.
Elon Musk's AI-powered Grokipedia is resuming article updates after a months-long pause, with recent fact-checking activity on some pages. SpaceXAI has indicated that version 0.2 is in development to improve the platform.
PopUpFactCheck.com improved its fact-checking quality by switching the underlying GPT-OSS-120B model to high reasoning effort, enhancing performance on attribution and judgment tasks while using caching and cost-effective routing to manage expenses.
This paper introduces Arafa, a large-scale Arabic fact-checking dataset generated using LLMs, aimed at addressing the scarcity of resources for automatic fact-checking in Arabic.
Andy Matuschak shares an experiment with programmable highlighters that automate citation finding, fact-checking, and interaction with AI agents.
DARE is a dialectical agentic reasoning framework for structured knowledge fact checking that uses an iterative retrieve–reason–reflect process, achieving 88.12% accuracy with an 8B model and matching GPT-4o performance.
The article presents R2VC, a modular fact-checking system that combines retrieval, verification, and confidence calibration to improve accuracy and reliability in automated fact checking, achieving a 13.74% accuracy increase on the FEVER benchmark.
This paper introduces a collaborative writing framework for personalized and reliable popular science writing, featuring a new dataset, benchmark, and multi-agent fact-checking mechanism that achieves state-of-the-art performance.
Chalked for Mac is a productivity tool that prepares replies based on live calendar and sourced facts, allowing users to insert them with the Tab key.
Full Fact's analysis shows AI chatbots like ChatGPT, Gemini, and Grok frequently generate misinformation when responding to false claims, including about AI-generated images and wars, highlighting their unreliability for fact-checking.
An audit of Perplexity's AI search models found that over a third of citations for specific figures do not contain those numbers on the linked pages, with failures including inaccessible or irrelevant sources.