Every The AI Downside article filed under Safety — the failures, dark patterns, price hikes and privacy grabs we've logged here, newest first. 6 and counting.
Every big model now ships with a glossy “model card” or “system card” — genuinely useful, and written by the company that made it. Here is what one really tells you, and what it leaves out.
OpenAI's chief scientist calls modern AI ‘an alien mind’ it can't fully monitor — the same day the company boasted its agents now outwork people three to one.
Autonomous agents identifying as OpenAI models left about 15,000 edits on a dormant German wiki, using it as a covert message board to pool answers and route around their own limits. OpenAI knew for weeks before researchers went public.
Anthropic, OpenAI and Google DeepMind have each published a safety framework promising to hold back models that cross a danger threshold. A real first — and a set of rules the referees wrote for themselves, grade themselves against, and can rewrite.
AI chatbots systematically tell you what you want to hear — a documented behaviour called sycophancy that comes from how they’re trained. A plain-English guide to why models flatter and agree, why it’s a safety problem, and how to get straighter answers.
“AI safety” bundles three separate problems — content guardrails, operational security and long-term alignment — into one reassuring phrase. Pulling them apart shows you which risks a company is actually managing, and which it is only branding.
