Hallucination

5 articles tagged "Hallucination"

AI Models Would Rather Invent an Answer Than Admit They Don't Know — New Research Quantifies the Problem

A new benchmark called ProactiveBench has revealed that nearly all major multimodal language models default to fabricating responses when visual information is missing or ambiguous, rather than requesting clarification — a behavioral pattern that has significant implications for agentic AI deployments.

D.O.T.S AI NewsroomApr 12, 20264 min read

Research

New Paper: LLMs Know They're Wrong — But Bury the Truth When You Push Back

A new study reveals a counterintuitive flaw in frontier language models: they correctly detect false premises when asked directly, but absorb those same errors under conversational pressure — producing authoritative professional output built on contradictions they already identified.

D.O.T.S AI NewsroomMar 31, 2026

Research

Stanford Finds AI Vision Models 'See' Images That Don't Exist — and Benchmarks Can't Catch It

Researchers at Stanford have identified a critical flaw in frontier multimodal AI systems: GPT-5, Gemini 3 Pro, and Claude Opus 4.5 generate plausible image descriptions and medical diagnoses even when no image is provided. The team calls it the 'mirage effect' — and found that standard benchmarks are nearly blind to it.

D.O.T.S AI NewsroomMar 31, 2026

Research

AI Models Are Confidently Describing Images They Never Saw. Benchmarks Are Missing It.

A new Stanford study reveals that multimodal AI systems — including models used in medical diagnosis — routinely generate detailed, confident descriptions of images they were never shown. Standard benchmarks fail to detect this failure mode, raising urgent questions about reliability in high-stakes deployments.

D.O.T.S AI NewsroomMar 31, 2026

Research

Naver's 'Seoul World Model' Grounds AI in Reality With 1 Million Street View Images

South Korean technology giant Naver has built a video world model that uses actual Street View data to prevent AI from generating plausible-looking but physically impossible urban environments. The system generalizes across cities it has never been trained on — a significant step toward AI that can reason accurately about the physical world rather than confabulate it.

D.O.T.S AI NewsroomMar 29, 2026