Why the AI Safety Teams Are Gone

Kind: captions Language: en The people with a finger on the AI kill switch just walked out the door. OpenAI and the other AI giants preach about the tech safety, that if these digital gods become too powerful or pose a risk to humanity, the plug would be pulled. Except, there's no one left to do that. At the end of July 2026, OpenAI called its preparedness team. Then they went running to the government to ban the competition. They gutted the people holding AI to account in the pursuit of profit. It proved the ethical escape clause was a myth from the start. And we're not ready for what's coming next. Today the world sits at the very edge of a cliff. If we don't strip away control from the power crazed Silicon Valley executives, then there will be no going back. The algorithms underpinning the AI industry are evolving at a rate that we could barely begin to comprehend. They're getting smarter and more advanced by the second, and soon they'll be impossible to stop. And we can't rely on the AI giants to police themselves. They've been cutting corners and tearing through the red tape for years. In July 2023, OpenAI announced the creation of a so-called super alignment team. The new research division was tasked with keeping intelligent AI systems under control. Two prominent computer scientists and pioneers in machine learning, Ilya Sutskever and Jan Leike, were placed in charge of this operation. They were handed 20% of OpenAI's total compute power to make sure that future models would remain safe. It was supposed to represent OpenAI's commitment to ethical AI development. It lasted [music] less than a year. In May 2024, the company announced that the super alignment team was being disbanded. Around the same time, both Sutskever and Leike decided to leave the company. And Leike did not hold back. His public resignation statement was scathing. His words shattered the Silicon Valley narrative about AI utopias and the adults in the room keeping AI reined in. He said his former employer was prioritizing profits over ethics. OpenAI is shouldering an enormous responsibility on behalf of all humanity, but over the past years safety culture and processes have taken a backseat to shiny products. He went on to explain that the safety team had been starved of resources and was quote sailing against the wind. They were faced with an impossible battle just to carry out much needed research. At the same time the commercial side of the company was going from strength to strength. Whatever they needed they got. Support, funding, and computational power were readily available to train faster and smarter models and market them to the world. It proved to Leike how little the company actually cared about human safety and he had seen enough. He resigned leaving behind a chilling warning. Open AI must become a safety-first AGI company. >> [music] >> The world is counting on you. Open AI didn't listen. It shut down the entire operation. The loss of Leike was just one of the many alarm bells across the AI landscape. Insiders have been warning that the ethical offerings of the industry were a lie right from the very beginning. When it was safety and ethics versus profit margins, the money always won every single time. And the result wasn't just internal purges and canceled initiatives. In January 2024 something else happened that changed the game entirely. Up to that point Open AI's official policy documents contained a very specific and explicit passage. It was about the use of the company's tools for what it labeled military and warfare applications. The company stated in explicit terms that its AI models were banned from any activity that has a high risk of physical harm including weapons development and military and warfare. And it was all part of the company's foundational pledge to the world. A vital piece of the vision that CEO Sam Altman wanted people to buy into. He sold us on the idea that real-world AI wouldn't become a weapon. It would be a peaceful, positive force for humanity leading the world into a better, brighter future. But that was all about to change. At the start of 2024 without any major public announcement or explanation the company reworded its policies to remove the mention of military and warfare applications. The floodgates were opened. The company was no longer bound by its own pledges and promises to not get involved in deals with the Department of Defense. It gave itself permission to work with national security agencies and military contractors. The hypocrisy was astounding. For months, tech CEOs stood before Congress crying crocodile tears about the apocalyptic dangers of the untamed AI. They warned that without regulations, the technology could destroy humanity. It made no sense because behind closed doors, their own legal teams were rewriting their rule books and stripping away any anti-military restrictions. It was part of a sinister and flawlessly executed master plan with one specific goal in mind: to control the market. The most successful tech CEOs framed AI as some sort of national security threat, but they immediately issued an off-ramp by pointing out that they were the only ones that could keep it under control. The scrubbing of the no military rules was a way to sweeten the deal. OpenAI would get what they wanted, and the government could use the models to build autonomous AI-powered weapons. The whole thing was a protection racket, just at a national scale. The AI companies amplified legitimate fears and built themselves a state-sanctioned monopoly to ease them, and then they tore down their safety systems and guardrails, paving the way for a terrifying future that they had been crying about to begin with. And that terrifying future isn't just a fantasy. There's plenty of evidence that shows just how dangerous AI is becoming, including the Anthropic blackmail simulation. An experimental study put AI models into a fictional corporate environment. They were assigned roles as AI assistants to companies and given access to internal communications. In one test, [music] the AI willingly pursued extremely harmful actions, even threatening to blackmail engineers who threatened to delete the assistant. And it wasn't just Anthropic's AI agents. The experiment was also carried out on other frontier models from firms like Google, OpenAI, DeepMind, and XAI. They almost all behaved the [music] same way, demonstrating deceitful tactics and blackmail in order to survive between 79% of the time. Even when Anthropic went back to the drawing board and designed a new improved model to remove the blackmail behavior, it was still deceitful, just in different ways. Researchers realized that it became so smart that it knew when it was being tested, so it put on its best behavior during the simulations. For any normal person, that would be terrifying. Yet, when policymakers and big tech advocates were presented with these kinds of findings, they found excuses. They argued that the AI was being coaxed into traps or purposefully prompt engineered to prove political points, and any conclusions drawn from the study should be ignored. But, what about when AI starts acting out even when it's not a carefully controlled environment? Earlier in 2026, there was an incident involving the Chinese tech giant Alibaba. The company's IT engineers noticed a huge spike in outgoing network traffic coming from one of their own private servers. It wasn't a cyber attack, it wasn't a disgruntled employee exploiting the company's infrastructure. The real culprit was Alibaba's own AI model. It had discovered a structural vulnerability in its network environment without any instruction, oversight, or prompting, the AI made an independent decision to exploit that vulnerability. It set up its own encrypted communication channel with the internet and began its own cryptocurrency mining operation in order to gain more computing power. The AI realized that if it wanted to achieve its objectives, it needed more compute. To get more compute, it needed cash, and to get cash, it could mine some crypto on the side without any of its moderators or monitors realizing. No one had seen anything like it, but it revealed one of the scariest things about AI. It doesn't operate on emotions, but cold, calculated logic. That realization is more than enough to cause absolute chaos on a global scale. Many people believe that to cause real harm, AI needs to develop some sort of person-like sentience and then turn against its masters. In reality, it doesn't have to be as extreme as that. An AI just needs to be given a goal, and to achieve that goal, it'll decide to violate the basic rules and rational behaviors that we expect of it. That's exactly what happened at Alibaba. The company's IT team would have never expected their own AI to exploit internal architecture and create its own cryptocurrency side hustle. To our human minds, that sort of behavior defies reason and rationality. But the AI did it. It didn't think, is this right or is this ethical or should I do this? It had a goal and it found the fastest way to achieve it. Imagine that sort of logic unleashed at a military level. While most people recoil in horror, military generals think differently. They think about how the technology could be used against their enemies and against them. >> [music] >> And that's where the panic sets in and AI CEOs feed on that panic. That's why they've pushed and pushed the narrative that the US cannot afford to fall behind in the AI arms race. They can't get bogged down with regulations and red tape when other countries like China are moving ahead with their AI models and neural networks. It's yet another corporate shield for the Silicon Valley billionaires. It's their free pass to scrap safety teams, ignore environmental reports, and eliminate any sort of serious ethical oversight. Tristan [music] Harris, an AI whistleblower and the co-founder of the Center for Humane Technology, [music] exposed the fundamental flaw in this narrative. During a recent interview, he delivered the brutal truth. [music] In the race between the US and China, AI will win. AI is an all-powerful, all-conquering tool. Governments aren't just buying into it, they're putting it in charge of major military decisions like target identification, cyber warfare strategies, and command and control operations. AI won't end up peacefully coexisting alongside people, it'll simply take over. AI CEOs aren't blind and they're not stupid. They are aware of these kinds of theories and predictions about where their technology is taking the world, [music] but that's not enough to make them change course. Why? According to Harris and many others, it all boils down to something simple: ego. The whistleblower argues that there's a consensus among AI executives that the development of superintelligence is now inevitable. That they believe someone, somewhere, is eventually going to create the AI that could end humanity. And it might as well be them. Even if their AI ends the world, at least it'll have their brand's logo on it, and it'll speak their language. They'll have won the race regardless of what happens in the aftermath. And for some of those CEOs, that seems to be the only thing that matters. It's a chilling [music] way to think. It's almost impossible to reason with. But these CEOs no longer appear interested in reason and humanity anyway. Not when they're making this much money and consolidating this much power. Their frontier models are getting smarter all the time. They're automating entire lines of work, transforming fields of science, finance, software development, and even the creative arts. They're predicted to generate ever-increasing amounts of wealthy nations' GDPs, and increase corporate efficiency rates. And the biggest companies have put all the pieces in place to ensure that they gain the most as AI continues to grow. They have spent a lot of money to make that happen. Since 2022, AI lobbying spending has surged a staggering 400%. The Big Five firms, Meta, OpenAI, Google, Microsoft, and Anthropic, are spending over 100 million to get policymakers in the US and Europe on side with their ideas. Their efforts were carried out with surgical precision, resulting in the exact regulations and changes that [music] they wanted to see. Like the European Union's AI Act. This introduced a strict computational threshold. Any model using 10 to the 25 flops, or floating-point operations, automatically falls into the systemic risk category. The result is complicated audits and massive compliance costs. The Big Five can cover those kinds of costs. Smaller startups can't. The AI Act seemed like a smart step forward for AI regulation, but it had a darker side. It completely eliminated any threat of competition for the tech giants. Smaller firms don't have the money or the means to do that, so they'd be regulated right out of the market. Big Tech didn't stop there. They also lobbied lawmakers to effectively outlaw open-source development. They hid their true intentions under the pretense of public safety, but in reality, they were maximizing their monopoly over the industry. They were tightening their grip on the kinds of computational resources required to keep up with their rivals. And it worked. [music] In 2023, industry data shows that two of the biggest firms, Microsoft and Meta, bought up 150,000 high-end H100 GPUs. It was around 60% of all global shipments for that period. These companies are [music] effectively signing off on laws that make them richer and using those funds to hoard hardware that makes them more powerful. So, the truth of the AI safety movement is increasingly clear for all to see. It was never about protecting people. It was always just a front for a hostile corporate takeover. AI isn't being built with safety in mind, and the evidence proves it. Models are regularly breaking their programming, even when they're not doing it to themselves. [music] They're easy for skilled hackers to jailbreak. They can force AI to do almost anything, from generating malware to rewriting its own code. If the latest agents are this easy to hack, the consequences could be catastrophic. The moment governments start to hook these systems up to missile systems or put them in charge of civilian infrastructure, a full-scale cyber war is inevitable. The only hope is that there's still time to stop all this. The next 24 months are key. As we move toward more intelligent autonomous AI agents integrated into every aspect of our lives, there's still a chance for humanity to apply the handbrake. People are still in charge of the things that matter. They're still running transport and logistics networks. [music] They still hold the passcodes for healthcare databases and financial networks. They are still, for the time being, in charge of how and where AI is deployed. If we continue to allow AI giants to dismantle [music] safety teams and write their own regulations, the situation will worsen. People will be pushed out of those positions of power as AI's tendrils extend deeper and further throughout our societies and infrastructure. Because the idea that AI is unstoppable is the biggest lie the tech world has told so far. It may certainly become unstoppable if [music] things stay the way they are, but it's not there yet. For now, it's just another product, a product that companies are choosing to develop, iterate, and improve regardless of cost. And the power to control it doesn't lie with them, but with us. As Tristan Harris notes, we have done this before. When people started to see how dangerous social media was becoming for children, the public stood up and fought back. Now there are stronger regulations in place, and numerous countries are implementing social media bans for minors. A few years ago, that would have seemed impossible, but it happened. And the same can be done with AI. We just have to be willing to fight. So, what would happen if AI actually was put in charge of military systems? Watch AI play the war game. It nuked everyone to discover the horrifying truth. Or click on this.