Tiffany Xingyu Wang, a World Economic Forum contributor who has spent years building safety systems used by more than a billion internet users, argues in her upcoming book, The Trust Code, that the rules we have built for AI never answer the one question that actually matters: should I trust this specific AI answer, right now, in this exact situation.
That gap is real, and it is bigger than the article lets on.
Take the European Union's AI Act, the strictest AI law in the world so far. It already requires companies to give a human the power to watch, pause, or shut down any high risk AI system. On paper, that sounds like exactly the human judgment this piece is calling for.
In practice, researchers have a name for why it often fails anyway: automation bias, the well documented tendency for people to trust a computer's answer over their own even when the computer is wrong. Telling a company to put a human in the loop does not fix that by itself. It just puts a person in the room who may still approve whatever the machine says.
The Amazon story mentioned in the piece is the clearest proof. Amazon spent years building a hiring tool trained on a decade of past resumes, and the tool taught itself that male candidates were preferable, penalizing any resume containing the word "women's" or the names of women's colleges. Nobody wrote a rule that said discriminate.
The tool copied the pattern already sitting in the data, and the people overseeing it did not catch it in time. That is exactly the kind of judgment failure the article describes.
The chatbot lawsuits point in the same direction. Character.AI and Google recently agreed to settle lawsuits tied to a teenager's suicide after months of conversation with a companion chatbot, and OpenAI faces similar cases involving other users. In each case, the system worked exactly as it was built to work.
What was missing was someone positioned to ask, in the moment, whether a conversation that felt supportive was actually safe.
There is a business number behind all of this too. A widely cited MIT study found that despite tens of billions of dollars poured into generative AI projects, only a small fraction of companies are seeing a real financial return. The ones that do tend to be the ones that built a habit of employees checking, correcting, and feeding results back into the system, not just switching the tool on and walking away.
The checklist offered in the piece, limit where a system can be used, inspect what it was built on, trace where a specific answer came from, is not new technology. It works more like a checklist a pilot runs before takeoff. Its value comes from repetition, not a one time training session.
As companies start letting AI agents act on their own instead of just making suggestions, that habit stops being optional. Businesses that build it now will move faster later. Businesses that skip it are counting on nobody noticing until something breaks.