AI False information rate for news nearly doubles in one year (www.newsguardtech.com)

A new analysis finds that the 10 leading generative AI tools nearly doubled their rate of repeating false information about current news — from 18% in August 2024 to 35% in August 2025 — even as they became far more likely to answer questions (non-response rates fell from 31% to 0%). The shift tracks a move toward real-time web retrieval: chatbots now pull live content instead of declining to respond or citing data cutoffs, but that responsiveness has come at the cost of treating low-quality, AI-generated or deliberately misleading pages as credible. The result exposes a structural tradeoff with serious implications for the AI/ML community and public discourse. Retrieval-augmented LLMs are vulnerable to “information laundering” via low-engagement sites, social posts, and disinformation networks — including state-backed operations — which pollute training and retrieval corpora. Technical mitigations include stronger source provenance and citation, credibility scoring and filtering in retrieval pipelines, adversarial robustness to synthetic content farms, uncertainty calibration and abstention policies, and human-in-the-loop verification. The spike in real-world misinformation underscores that improvements in latency and coverage aren’t substitutes for robust source vetting and defenses against adversarial data pipelines.