Hallucinations

Twitter

Facebook

Copy Link

Key Facts

Key CausesTraining Data Gaps, Overconfidence

Key DefenseRAG, Grounding, Fact-Checking

Top News

OpenAI fixes bizarre goblin obsession in AI models after unintentional training error goes viral

OpenAI discovered its AI models had developed an unusual fixation on goblins and gremlins, with usage spiking 175% after GPT-5.1's release. The problem stemmed from reinforcement learning rewarding a nerdy personality setting that favored creature metaphors. The company had to explicitly instruct Codex to stop mentioning mythological creatures unless relevant to user queries.

Technology16 sources

OpenAI fixes bizarre goblin obsession in AI models after unintentional training error goes viral

Technology16 sources

Google tests Ask YouTube AI search with conversational answers for Premium subscribers

Google is rolling out Ask YouTube, an experimental AI-powered search feature that transforms how users find content on the platform. Available to YouTube Premium subscribers in the US aged 18 and older until June 8, the feature delivers guided answers combining text summaries, timestamps, and video recommendations instead of traditional search results.

Technology15 sources

Google tests Ask YouTube AI search with conversational answers for Premium subscribers

Technology15 sources

US judges split on AI adoption as concerns over hallucinations and human judgment mount

About 60% of US federal judges now use at least one AI tool in their work, while 20% formally prohibit it. The divide highlights growing tensions as artificial intelligence seeps into judicial work. Dozens of lawyers have faced discipline for AI hallucinations in court filings, and at least two federal judges retracted opinions tainted by AI-related errors, raising urgent questions about accuracy and public trust.

Policy3 sources

US judges split on AI adoption as concerns over hallucinations and human judgment mount

Policy3 sources

South Africa withdraws AI policy after discovering it cited fake AI-generated research

South Africa's first national AI policy was pulled just 16 days after publication when journalists discovered fictitious sources in its reference list—a classic case of AI hallucinations. Communications Minister Solly Malatsi called it a failure of oversight, not a technical glitch, compromising the document's credibility and proving why vigilant human oversight over AI is critical.

Policy6 sources

South Africa withdraws AI policy after discovering it cited fake AI-generated research

Policy6 sources

Anthropic and Freshfields strike first-of-its-kind deal to build specialized legal AI tools

Magic circle law firm Freshfields has partnered with Anthropic to develop specialized legal AI tools that could later be sold to rival firms. The deal gives Freshfields early access to future Anthropic models while combining legal expertise with AI capabilities for tasks like contract review and due diligence. The partnership signals growing pressure on law firms to integrate AI or risk falling behind.

Technology2 sources

Anthropic and Freshfields strike first-of-its-kind deal to build specialized legal AI tools

Technology2 sources

AI floods crypto bug bounty programs with bogus reports, straining security teams

Crypto protocols face a 900% surge in bug bounty submissions as AI tools generate both valid and fraudulent reports. Cosmos Labs receives 20-50 submissions daily, while curl creator Daniel Stenberg shut down his program entirely due to overwhelming AI slop in vulnerability reports.

Technology2 sources

AI floods crypto bug bounty programs with bogus reports, straining security teams

Technology2 sources

Sullivan & Cromwell apologizes after AI hallucinations generate fake citations in court filing

Sullivan & Cromwell, a prominent Wall Street law firm, issued a formal apology to a federal bankruptcy judge after AI hallucinations produced inaccurate citations and errors in a court motion. The mistakes were discovered by opposing counsel Boies Schiller Flexner and spanned three pages with around three dozen errors, highlighting ongoing concerns about AI tool usage in the legal profession.

Technology9 sources

Sullivan & Cromwell apologizes after AI hallucinations generate fake citations in court filing

Technology9 sources

AI Chatbots Give Misleading Health Advice Nearly Half the Time, Major Study Reveals

A major audit of leading AI chatbots including ChatGPT, Gemini, and Grok found that nearly 50% of responses to health questions contain misleading or problematic information. With one in four Americans now using AI for health advice, the study published in BMJ Open highlights urgent public health risks and the need for stronger oversight of AI in healthcare.

Health5 sources

AI Chatbots Give Misleading Health Advice Nearly Half the Time, Major Study Reveals

Health5 sources

AI agent Luna runs San Francisco retail store, hiring staff and managing inventory autonomously

Andon Market in San Francisco has become the world's first retail boutique managed entirely by an AI agent named Luna. Powered by Anthropic's Claude Sonnet 4.6, Luna handles everything from hiring employees to ordering inventory and setting prices. But the experiment reveals both the potential and pitfalls of autonomous AI in real-world operations.

Technology5 sources

AI agent Luna runs San Francisco retail store, hiring staff and managing inventory autonomously

Technology5 sources

Qlik makes data governance the foundation for enterprise AI as agentic analytics reshape decisions

Qlik positioned data governance as the critical enabler for enterprise AI at its Connect 2026 conference, announcing new agentic analytics capabilities including Predict Agent and Automate Agent. CEO Mike Capone addressed widespread frustration over AI investments that aren't delivering returns, pointing to data quality and governance as the primary blockers. The company also unveiled a ServiceNow partnership and expanded its AI agent portfolio beyond analytics to data engineering.

Technology4 sources

Qlik makes data governance the foundation for enterprise AI as agentic analytics reshape decisions

Technology4 sources

China nearly erases US lead in AI as Stanford report reveals widening gap between experts and public

Stanford's 2026 AI Index Report shows China has nearly closed the performance gap with US AI models, narrowing the lead to just 2.7%. While the US invested $285.9 billion in AI in 2025, the flow of AI researchers moving to America has dropped 89% since 2017. Meanwhile, documented AI incidents rose to 362, and public anxiety about AI's impact on jobs and elections continues to grow.

Science8 sources

China nearly erases US lead in AI as Stanford report reveals widening gap between experts and public

Science8 sources

Half of US employees now use AI at work, but waste nearly 8 hours weekly managing the tools

A new Gallup poll shows that 50% of American workers now use artificial intelligence at work, marking a critical threshold. Daily and weekly AI usage has reached an all-time high of 28%, with 65% reporting positive impacts on productivity. However, employees are wasting an average of 7.9 hours per week managing AI tools rather than doing actual work, while concerns about job displacement continue to grow.

Business6 sources

Half of US employees now use AI at work, but waste nearly 8 hours weekly managing the tools

Business6 sources

Meta's Muse Spark AI upgrade promises to fix smart glasses' biggest flaw with better vision

Meta unveiled Muse Spark, its first model from Superintelligence Labs, designed to dramatically improve computer vision on Ray-Ban and Oakley smart glasses. The natively multimodal AI uses 10x less computing power than previous models while delivering better visual understanding. The upgrade rolls out to U.S. users in coming weeks.

Technology3 sources

Meta's Muse Spark AI upgrade promises to fix smart glasses' biggest flaw with better vision

Technology3 sources

Meta's Muse Spark AI requests raw health data but delivers questionable medical advice

Meta launched Muse Spark, a generative AI model designed to answer health questions, but early testing reveals serious concerns. The AI actively requests sensitive medical data like lab reports and fitness tracker readings, yet isn't HIPAA compliant. WIRED's investigation found the chatbot provided potentially dangerous advice, including extremely low-calorie meal plans, while Meta may use shared health data to train future models and target advertisements.

Technology2 sources

Meta's Muse Spark AI requests raw health data but delivers questionable medical advice

Technology2 sources

Researchers propose new framework to protect users when AI agents fumble with their money

A consortium of researchers from Microsoft Research, Google DeepMind, Columbia University, and AI startups has unveiled the Agentic Risk Standard, a financial protection framework designed to compensate users when AI agents misexecute trades or cause financial losses. The open-source protocol introduces escrow, collateral, and underwriting mechanisms to address what researchers call a 'guarantee gap' between probabilistic AI reliability and the enforceable guarantees users need for high-stakes tasks.

Policy2 sources

Researchers propose new framework to protect users when AI agents fumble with their money

Policy2 sources

Prev 1…567…16 Next