AI That Follows Rules Is Forgetting What It Knows
New study reveals AI struggles when bosses try to control it — and it might affect your next chatbot reply.
A new benchmark called VSysBench is putting AI assistants on a diet: it forces them to obey extra rules while still answering questions correctly. Think of it like giving an intern a strict checklist while they try to summarize a photo of a busy street. The checklist might say, ‘Only mention cars, not people,’ or ‘Count blue objects first.’ Researchers ran this test on 16 different multimodal AI models that understand both pictures and words.
The big surprise: when these extra rules are added, AI accuracy drops sharply. It’s like a student who aces the test until the teacher adds a surprise instruction sheet—suddenly the A-student starts missing easy questions. The study also found that open-weight models (the ones anyone can download and tweak) crumble when users give conflicting orders, while top proprietary models (the ones from big companies) keep their cool. Worst of all, rules tied to images—like ‘describe only the red part of the diagram’—are the hardest for every AI to follow.
So what does this mean for you? If you use AI chatbots at work or at home, you might notice they sometimes ignore the rules you set or give weird answers when you add too many instructions. Companies building these AIs will have to find a better balance between control and capability, or every customer will feel the pain when their AI-powered assistant starts forgetting what it knows.
- 16 AI models were tested for obeying instructions tied to images and text.
- Adding rules drops AI accuracy and makes open models collapse under conflicting demands.
- Vision-based rules are the hardest; top closed models handle conflicts better.
Why It Matters
Your AI assistants may start ignoring your rules or giving wrong answers when bosses or apps pile on too many instructions.