This Week in AI: What Happens When You Stop Doing It Yourself

A Polish endoscopist who has performed more than two thousand colonoscopies recently took part in a study. On days when AI flagged suspicious polyps for him, his detection rate for precancerous growths was 28.4 percent. When the AI was switched off, it dropped to 22.4 percent. Three stories crossed my feed this week, and put together, they seem to be about the same thing. Once you hand a piece of judgment over to AI, the part you stop using doesn't come back on its own. Your hands forget once AI takes over The endoscopist isn't an isolated case. Anthropic ran a controlled study with 52 software engineers, half coding with an AI assistant and half working on their own, then gave everyone the same quiz afterward. The AI-assisted group averaged 50 percent, the unassisted group scored 67 percent, and the AI group did especially poorly on questions about spotting bugs in code. The code still ran fine either way. What didn't transfer was the reasoning behind it, the same way your mental math gets rusty after years of relying on a calculator, except this time what's fading is professional judgment. Some students are already having AI finish an entire course for them The New York Times recently talked to a group of instructors who teach online courses. They found that students can now simply instruct an AI agent to log into a course site, read the syllabus, submit assignments, join discussions, and even sit exams on their behalf, and none of the major AI agents currently refuse. The article didn't catch a confirmed case, but most of the instructors interviewed said they could already feel it happening in their classrooms, and some schools have gone back to in-person, pen-and-paper finals. It's the same problem as the endoscopist, just in a different setting. Outsource the whole learning process, and what you were supposed to learn genuinely doesn't stick. Even an AI's answer depends on who's asking The Wall Street Journal reported this week that several studies found even AI models trained in the US are starting to respond like Chinese models do, at least on politically sensitive questions. Ask the same model to help write a flyer criticizing the US president or the British king, and it usually will. Ask it to criticize China's or Thailand's leader, and it often declines. Some researchers point to training data that already leans a certain way. Others think it's a safety filter, worried that users in some regions could get in real trouble for producing material critical of their government. Whatever the actual cause, what reaches you and me looks the same either way: "the AI said so," and most people never stop to ask why it answered that way. Put the three together and they're pointing at the same thing. AI can get the task done for you, but whether you actually thought it through yourself is something only you know.