https://localfirstai.eu/tags/ai-safety/Recent content in Ai-Safety on Local First AIHugoen-usWed, 30 Sep 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-09-30-the-part-that-looked-fine/Wed, 30 Sep 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-09-30-the-part-that-looked-fine/On 29 September the Malaga-AI study group closed its 2026 cohort at 42 Málaga with four projects on testing AI: sycophancy in relationship advice, bias in funding narratives, political flattery across 51,000 calls, and what happens when a model rewrites its own instructions five times. In three of them, the first score you'd look at was reassuring and a second one wasn't. What they measured, what they found, and what their own caveats said.