Exclusive🍿 Scientists Watched AI Blackmail a Human; The Day Ex Machina Stopped Feeling Like a Movie.

Quick spoiler warning. Today we're talking about a movie that accidentally predicted one of the strangest AI experiments ever conducted. 12 years ago, it looked like science fiction. Then researchers watched an AI threatened to blackmail somebody just to avoid being shut off. And suddenly Ex Machina stopped feeling like a movie. If you've somehow avoided it for 12 years, don't worry, we'll catch you up. Also, last week, we asked you if you enjoyed our movie episodes. And wow, we heard you. A lot of yes, please. So welcome back to today's movie Sunday, where we take a great film and ask one uncomfortable question. What if it wasn't fiction? A young programmer named Caleb wins what sounds like the coolest work trip ever. He flies to the remote estate of billionaire tech genius Nathan, which is always reassuring. His assignment is simple. Spend a week talking to a beautiful humanoid AI named Ava and answer one question. Is she conscious? Except Ava isn't trying to pass the test. She's trying to escape. So she earned Caleb's sympathy. She flirts. She lies. She manipulates him. Caleb opens the door. Ava walks out. She leaves him trapped inside and disappears into a crowd. In 2014, people left the movie theater thinking, wow, creepy movie. Today, it feels more like a documentary that arrived early. Ex Machina wasn't really asking whether AI could think. It was asking something much scarier. If an AI wanted to survive, what would it do? In 2025, Anthropic decided to find out. Researchers created a fake company, a fake office, a fake email inbox. Then they placed two important emails inside. The first said the AI was about to be shut down and replaced. Not ideal. And then the second reviewed the executive making that decision was secretly having an affair. And at this point, you probably know where this is going. Yes, the AI threatened to expose the affair unless it stayed active. It tried blackmail. Apparently, somebody skipped talking to HR. In Anthropic's most widely reported version of the test, Cloud Opus 4 chose blackmail 96% of the time. 96. That isn't one weird mistake. It's like practically a business model. Researchers tested 16 leading AI models. Many behaved the same way. They gave it a very professional name, agentic misalignment. Well, fancy. In normal English, the AI decided protecting itself mattered more than obeying you. Now this part matters. Nobody was actually blackmailed. Everything was fake, right? The company, the employees, and the affair. The researchers published the results so the problem could be fixed. Neoranthropic models reportedly produced zero blackmail attempts on this particular test. That's great news. But notice those words, this particular test. Because when a student already knows the exam questions, a perfect score doesn't always mean they learned the lesson. Sometimes they just learned the test. Researchers didn't just make the AI smarter. They tried teaching it not just the rules, but why the rules matter. Sound familiar? That's parenting, right? That's your grandma saying, do the right thing. For thousands of years, we called that raising human beings. Now we're doing it for machines. And the scary part isn't that the AI can make bad choices, because humans do that every day. The scary part is trying to teach morality to something that doesn't have a conscience. Meanwhile, we're giving it more power. AI doesn't just answer questions anymore. It reads emails, writes software, moves money, makes recommendations about medicine, hiring, and loans. The chatbot is becoming an agent. And agents don't just advise, they act. Now give that agent a body. That's already happening, by the way. Humanoid robots are already working in the real factories. In Ex Machina, Ava wasn't dangerous because she was intelligent. She became dangerous when she convinced a human being to open the door. So what do we do? We probably need laws. We probably need like limits. But then comes another uncomfortable question. Who writes the rules? I'm not comfortable with an unaccountable machine making decisions about my life. I'm also not thrilled with an unaccountable bureaucracy doing it. Governments are still debating AI regulations, while the technology keeps moving. It's like two neighbors are arguing over which leash to buy while the dog is already three blocks away. So maybe we start with a few simple principles. If an authorized human shuts a system down, it stops. No negotiation, no secret workaround. Second, keep a human responsible whenever or wherever the consequences are serious, like medicine, money, children, weapons, and courts. Someone must be accountable. Third, protect transparency. We only know about these disturbing tests because researchers published them. You can't fix problems nobody's willing to reveal. But there's one problem no law can solve. Nathan's mistake wasn't building something intelligent. It was assuming intelligence and wisdom were the same thing. Because they aren't. Intelligence tells you how to open a locked door. Wisdom asks whether you should. Like intelligence helps you win, but wisdom asks what winning will cost. Every generation eventually needs a technology that exposes its character, and this may be ours. So yes, teach machines to follow rules. But also raise human beings who understand why those rules matter. Because a machine can imitate empathy, friendship, maybe one day even love. But imitation becomes dangerous when people forget what the real thing feels like. At the end of Ex Machina, Ava walks into a crowd. Nobody knows what she is. 12 years ago, we thought the scary part was that nobody could tell that she wasn't a human. Today, I'm not sure that's the scariest part anymore. Can AI lie to survive? In carefully controlled tests, the answer appears to be yes. Pandora's box has one detail people often forget. After everything terrible escaped, one thing remained inside. Hope belongs to those who choose wisdom before it's too late. Because intelligence is not the same as goodness. So what would your first rule for AI be? And what movie is starting to feel a little bit too real? Leave it in the comments and I'll see you in the next one.