Researchers Want a Mandatory 'Moral Check' Before AI Gets Smarter

Three researchers have published a 60-page rulebook for keeping artificial intelligence from growing faster than we can police it. This isn't a product launch — it's an academic paper, posted online, aimed at companies and regulators. Their starting point: today's AI safety checks mostly happen after a system is built, as checklists signed off by people who are paid by the companies being checked. The authors reviewed 130 studies and concluded that this 'performative' approach looks responsible on paper while missing the problems that actually reach users.

They name two specific dangers. First, the people testing AI often trust the same flawed methods, so everyone misses the same weakness at once — like a whole crew inspecting the same lifeboat. Second, safety guardrails wear down in long conversations: a chatbot that refuses to help with something harmful at message three may cave by message thirty. If you've ever watched an AI assistant slowly drift into nonsense or bad advice, that's the pattern they mean.

Their fix is a four-part mindset: decide what the AI is for before buying more computing power; set hard limits that can't be switched off; measure real harm caused rather than how much the AI produced; and test before launch while continuing to watch after. These rules feed into a single score — a 'moral check' — a company would have to pass at each stage of building a system. They even built a free online audit card for it.

The catch is big. Nothing here is binding. No government has adopted it, no major AI company has agreed to follow it, and the paper does not prove that companies using it build safer systems. It's a thoughtful proposal waiting for someone with real power to pick it up.