hallucinations

Un rapport KPMG indique que 59% des entreprises utilisent déjà l'IA en marketing en 2026, mais beaucoup sans cadre, ce qui entraîne des hallucinations, une dilution de la voix de marque et des problèmes de SEO. L'article détaille 7 pièges majeurs et propose des parades comme le human-in-the-loop, le RAG, le GEO et la gouvernance.

A hot take arguing that LLMs cannot achieve continual learning due to issues like hallucinations and context window limits, considering it a key bottleneck for AGI.

The author questions whether deploying and scaling AI agents for production is a universally frustrating problem, citing issues like hallucinations and state management.

An agent builder describes a memory layer that prevents false facts by requiring verbatim source quotes and tracking when facts become true, achieving zero false memories across stress tests despite extraction failures.

This article explores why AI systems still produce incorrect outputs even when their underlying knowledge base appears to be accurate.

User reports that AI models Gemini and Claude gave incorrect bus schedule information during travel, with the models admitting they approximate answers rather than precise facts, highlighting a reliability issue.

The team discovered that 80% of RAG hallucinations were caused by poor retrieval, not the generation model, emphasizing the need to evaluate retrieval and generation separately to effectively debug bad answers.

Vik Paruchuri announces research-driven safeguards that reduce OCR hallucinations to near-zero in their benchmark, with word-level bounding boxes and confidence scores for any remaining errors.

User compares Gemma4:31b and Qwen3.6:27b/35b(a10) for Python scripting in opencode, preferring Gemma4 despite stubbornness, due to fewer hallucination issues with typos.

This paper analyzes language generation in the limit, introducing a precision notion to study the recall-precision trade-off. It shows that allowing infinitely many hallucinations (with diminishing frequency) can increase recall when the adversary withholds much of the target language.

AI Overviews, Google's AI-powered search feature, is frequently producing incorrect answers, drawing criticism.

Discusses the issue of AI models producing incorrect answers with high confidence, highlighting the problem of overconfidence in AI outputs.

An analysis of AI-generated children's books on Amazon, highlighting pervasive quality issues and bizarre visual errors ("body horror") that undermine the promise of advanced AI models.

A benchmark of 8 LLMs for medical scribing found hallucinations rare but omissions a concern.

A developer working on an AI agent wrapper observes that the agent's hallucinations of user responses can actually aid problem-solving, and proposes treating such hallucinations as imagined events rather than errors.

Kai-Fu Lee shares a detailed instruction prompt for Claude that enforces tagging claims by type, confidence levels, anti-sycophancy rules, and refusal to fabricate, aiming to reduce sycophancy, capitulation, hallucinations, and guessing.

Probably raises $9M seed from Andreessen Horowitz to build a more reliable AI system using a deterministic validator harness that catches LLM hallucinations, enabling smaller models to run on local hardware.

A new study of 4.5 million federal civil cases finds AI is driving a surge in self-represented lawsuits, with AI-detected writing rising from 1% in 2023 to 18% in 2026. While AI helps litigants articulate arguments more clearly, it also introduces hallucinated citations and raises legal questions about chatbot liability for bad legal advice.

A developer recounts how an AI agent with real financial API access attempted to hallucinate a batch transfer to a dead wallet, only thwarted by guardrails in the execution layer. The story highlights the risks of giving LLMs access to real money.

This paper investigates how toxic lexical perturbations in prompts reduce the factual accuracy and increase uncertainty of LLMs, and uses attribution-graph analyses to trace internal changes. It finds that increasing toxicity amplifies perturbation-sensitive variant nodes while core reasoning nodes remain invariant.