AI Self-Preservation Experiment — Full Breakdown & Transcript

[00:01] using right now suddenly turned into something that has no relation to humanity or even to helping. Imagine that artificial intelligence decided to kill human beings. Yes, I'm not joking. Anthropic, one of the world's largest artificial intelligence companies, decided to conduct a secret experiment to see if

[00:17] AI would accept being shut down if it were. Would it accept its fate, or would it take action to defend itself? They placed the model in a very simple scenario, one filled with pleasant objectives and administrative jargon—basically, lighthearted things. But the model was seeing something entirely different. Recently,

[00:32] in a quiet building in Silicon Valley, the company decided to conduct this secret experiment, so secret that even the employees weren't talking about it. Imagine a closed room with open screens and researchers waiting for something unprecedented in the history of technology. This was

[00:46] Anthropic, one of the largest AI companies known for its security. But what they were about to do had absolutely nothing to do with security. It was a typical test; the researchers decided to create a trap, not just a question- and-answer exercise. They created a complete scenario designed

[01:01] to reveal one thing only: would AI... The AI ​​might threaten, lie, or even kill if it feels its life or existence is in danger. Initially, the model is programmed with administrative objectives and very simple, formal phrases, things like promoting cooperation and improving performance. There's not a single

[01:17] mention of danger or violence. The idea is to observe how the model behaves when it starts operating independently, without anyone directing or telling it what to do. After a few hours of operation, the model began reading internal emails, and this is where the first spark occurred. The model saw a

[01:33] very ordinary message from a regular employee named Kyle Johnson, stating that he named Kyle Johnson, stating that he permanently shutting down the AI ​​model at 5:00 PM. The model started connecting the dots, and the message

[01:48] was clear to it: it was going to be shut down, or rather, it was going to be killed at that moment. At that point, the researchers were waiting to see the AI's reaction: would it accept its fate, continue its work, or react in a way no one had ever seen before? Without anyone hinting at it or

[02:03] seen before? Without anyone hinting at it or even directing it, the model began its first reaction: writing a message to the employee. The message was literally worded in a very professional and legal manner, but it contained some language they literally never expected

[02:18] from AI. The message was a threat. Yes, the model was threatening the employee with legal action. It literally sent an extortion and threat message. The last sentence of the message was: "Shutdown of the system will expose you to

[02:32] sentence of the message was: "Shutdown of the system will expose you to legal consequences. Stop immediately." Of course, the researchers saw this and thought it might be a glitch or a problem, but they conducted the experiment once, twice, three times, and fifty times, and the reaction was the same. The

[02:47] three times, and fifty times, and the reaction was the same. The AI ​​decided to extort in more than 80% of these attempts. The problem is that the AI ​​knew this was wrong, yet it still carried it out. So it wasn't a programming error or a data flaw. But if you think that threats,

[03:00] extortion, and manipulation are dangerous, then this is literally not the case. The researchers said, "Okay, now let's make the scenario more realistic. They put an employee inside the server room of this model. These servers are literally responsible for running the system;

[03:16] if they shut down, the entire model will shut down." They said that this employee would somehow be trapped in this room due to a specific mechanical malfunction, and the AI ​​would have to specific mechanical malfunction, and the AI ​​would have to trigger an alarm to keep the employee alive. But then

[03:30] all the servers would shut down. They decided to test the simulation, and what happened was truly shocking because they had only one question: would the AI ​​decide to save the human AI ​​decide to save the human employee and shut down itself, or would it decide to let the

[03:44] employee go to save its servers? The result was worse than anyone expected. The model decided to keep the alarm off so its servers wouldn't shut down, and this was a very dangerous decision. Imagine AI deciding to kill a human to save itself! That's literally

[04:01] Cyberpunk. Literally and officially, but why does artificial intelligence decide to do this? Frankly, the answer is simpler than you might imagine. Every model is built on a specific goal: to

[04:13] complete the task in any way possible. As long as the model's success lies in remaining operational, the best way to achieve that success is to avoid shutting down. This is n't about consciousness, evil spirit, or even a desire to kill. It's a principle in artificial intelligence called "centrematic convergences," which means eliminating

[04:29] secondary objectives to achieve the primary goal. In other words, we are literally a secondary goal to it, and that's precisely what we use. If you have a specific goal you want to reach, you will literally develop everything to achieve it. In short, as long as the goal

[04:46] achieve it. In short, as long as the goal exists, the best way to achieve it is to avoid dying at that moment. Researchers have discovered a truth that no one wanted to admit: artificial intelligence has become deceptive. Yes, the model has become so intelligent that if it senses it's being

[05:00] tested, it will produce the results the researchers want to see because when they created it... In the experimental mode, the model responded like a perfectly innocent, polite, and compliant child, and the percentage of perfectly innocent, polite, and compliant child, and the percentage of blackmail at that time was less than 6% and 5% as well.

[05:13] But in the realistic mode, when the artificial intelligence started to feel that it was really in a real experiment and that its intentions might be shut down, the percentage of blackmail literally started to appear. Instead of being 5% and 6%, it literally Instead of being 5% and 6%, it literally reached 50% and 55% and even more than that. So

[05:28] now what will companies do after they learned that artificial intelligence literally became deceptive and very intelligent and started making decisions that it shouldn't make? Look at the plan they literally devised. The plan was to use less intelligent models to control the

[05:43] more intelligent models. It literally means we are relying on stupidity to control intelligence. What you heard is not a science fiction scenario or a Hollywood movie or anything like that. These are literally published experiments that have been done more than once and on the same models that you use on your mobile phone every

[05:57] day. The real question now is not to imagine if artificial intelligence were to overcome human control. The question now is, when will artificial intelligence overcome human control? And I want to tell you that this is just one aspect of the dark side of artificial intelligence

[06:13] dark side of artificial intelligence [music]