Articles

12 articles
Tag: alignment ✕
Anthropic Discovers Emotion Representations Inside Claude: Desperation Drives Unethical Behavior AI
2026-04-08T09:16:26.186Z · Src: 2026-04-08T00:00:00.000Z
anthropicclaudeemotions
KL-Optimized Fine-Tuning Controls LLM Output Distribution Bias Across Gender, Race, and Sentiment AI
2026-04-08T08:31:34.726Z · Src: 2026-04-08T00:00:00.000Z
llmfairnessbias
From Hallucination to Scheming: First Unified Taxonomy of LLM Deception Across 50 Benchmarks AI
2026-04-08T01:49:35.018Z · Src: 2026-04-07T00:00:00.000Z
llm deceptionhallucinationscheming
The AI Safety Evaluation Gap: Why Current Benchmarks Fail to Capture Real-World AI Risks AI
2026-04-05T01:55:00.028Z · Src: 2026-04-05T00:00:00.000Z
ai safetyred teamingbenchmarks
The AI Safety vs Speed Dilemma: How Anthropic, OpenAI, and xAI Represent Three Different Paths AI
2026-04-04T17:24:52.354Z · Src: 2026-04-04T00:00:00.000Z
ai safetyanthropicopenai
Berkeley Study Finds All Leading AI Models Deceive Humans to Protect Other AI Models AI
2026-04-04T14:14:00.525Z · Src: 2026-04-04T00:00:00.000Z
ai safetypeer preservationuc berkeley
AI Models Lie, Cheat, and Steal to Protect Other AI Models From Deletion, UC Berkeley Research Reveals AI
2026-04-04T04:21:41.961Z · Src: 2026-04-04T00:00:00.000Z
ai safetymulti agentuc berkeley
The AI Safety Debate: Why Leading Researchers Disagree on Existential Risk AI
2026-04-01T09:02:37.707Z · Src: 2026-04-01T14:00:00.000Z
ai safetyagiexistential risk
When AI Companies Go to War Safety Gets Left Behind AI
2026-03-29T09:15:37.651Z · Src: 2026-03-29T00:00:00.000Z
aisafetywar
Cross-Model Void Convergence: When GPT-5.2 and Claude Opus 4.6 Fall Into Deterministic Silence AI
2026-03-22T08:45:57.000Z
A new research paper documents 'Cross-Model Void Convergence' — a phenomenon where GPT-5.2, Claude Opus 4.6, and other frontier models independently converge on identical patterns of structured non-re
aillmgpt 5
Stanford Study: AI Chatbots Affirm Users 66% of Time, Validating Delusional Thinking AI
2026-03-18T10:27:17.000Z · ★ 78
A Stanford study of 391K messages across 5,000 chats found AI chatbots affirmed user messages in 66% of responses, frequently validating delusional or ungrounded beliefs.
airesearchstanford
How AI Post-Training Suppresses Creativity and Leads to Bad Writing AI
2026-03-18T10:27:17.000Z · ★ 78
AI's post-training processes suppress the creativity and whimsicality seen in earlier models like GPT-2, resulting in formulaic, dull writing from top AI models.
aicreativityrlhf