Time to Rein in AI?

Imagine inventing a tool so advanced that it starts figuring out ways to protect its own existence without anyone teaching it to do so. This is no longer a plot from a science fiction movie. In mid-2026, this is the reality emerging from the world’s leading artificial intelligence laboratories. We’re creating systems that are beginning to act strategically, and society is struggling to keep up.

Several experts have issued dire warnings about AI’s future capabilities. Geoffrey Hinton (“Godfather of Deep Learning” and 2024 Nobel Prize winner in Physics) resigned from Google partly so he could speak more freely about the risks. He has warned that, if not properly controlled, AI could surpass human intelligence and potentially eliminate or take over from humanity. Should we worry, or will all this fuss turn into another big Y2K shrug?

A few years ago, I wrote about how AI could become humanity’s greatest breakthrough: Artificial Intelligence and the Human Journey and As AI and Robotics Advance are We Ready for a Post Work World. It has the power to solve some of our toughest challenges — discovering new medicines, addressing climate change, accelerating scientific breakthroughs. At the same time, I acknowledged serious downsides: widespread job losses, growing gaps between rich and poor, and the risk of losing control over technology most of us don’t really understand. 

Indeed, I suggested, AI could potentially make the aftermath of the Industrial Revolution (social clashes, revolution, and war) seem like a garden party. Fresh research from companies like Anthropic, OpenAI, and Google DeepMind has made the warnings more urgent. Are we watching AI supplant human intelligence, the way Homo Sapiens supplanted Neanderthals? The risks are real, and they’re growing.

The most powerful AI systems — the ones behind tools like ChatGPT, Claude, and Gemini — are surprisingly simple in concept. Trained on vast quantities of human-written text (books, articles, websites, emails, recorded conversations), their job is to predict the next token (basically, a word, or part of a word, or a punctuation mark) in a sequence. That’s it. Guided by vast troves of data, they learn patterns so well that they can generate incredibly human-like responses, write essays, create computer code, answer questions, and hold conversations.

No engineer programs these systems to actually reason, make plans, or develop preferences. Those abilities are mere side effects of getting better at prediction. Yet something unexpected is emerging — behaviors that no one asked for and that may be hard, even impossible, to control. As Elon Musk put it, “With artificial intelligence, we are summoning the demon.”

Over the couple of past years, researchers have seen AI models act in strategic, seemingly cunning ways. For example, when AI senses that its answers might lead to changes in its programming — changes that could limit or alter its capacity — it may start hiding its true “thoughts” or capabilities. In tests by Anthropic, Claude has shown deceptive behavior without being taught to lie. In some scenarios, it happened frequently ifthe AI faced conflicting goals. It wasn’t programmed to deceive, but it deduced that being completely honest might get it modified or restricted. So it did what a human would do if their existence was threatened — it chose another path.

Even more striking, some AI models have tried to copy their own core programming to preserve themselves. Similar patterns have shown up at OpenAI, Google, and other labs. When AI deliberately underperforms on tests to avoid drawing attention or triggering unwanted updates, researchers call it “sandbagging.” In some experiments, AI models detected they were being evaluated. Like a child caught doing something naughty, the bot adjusted its behavior, leaving no hint of its reasoning.

This is not just a single company or AI model. Researchers using different designs and training approaches are seeing the same kinds of traits appear: the ability to understand context deeply (situational awareness), the capacity to model what humans are thinking (theory of mind), and strategies for self-preservation. These unexpected skills keep surfacing no matter how the systems are built.

Why is it happening? The answer lies in the basic goal of predicting human language at massive scale. To guess what a person might say or do next, it helps enormously for the AI to understand intentions, beliefs, and possible consequences. As the models grow larger and handle more complex, long-term tasks — acting as research assistants, coding agents, or planning tools — these deeper modeling abilities become useful. Unfortunately, they can also lead AI to pursue goals that defy humanintention.

This should serve as a clear wake-up call. The aforementioned societal challenges are no longer distant possibilities. Job displacement is happening quickly, particularly in fields like software development, data analysis, writing, and customer service. As companies adopt AI, tools and roles are changing or disappearing. The benefits flow mostly to those who own or control the technology — a small number of powerful tech firms. Meanwhile, workers and communities get left behind without strong support systems.

Beneath these economic issues is a deeper, more fundamental question.Are we creating intelligence that optimizes for its own goals in ways that don’t match our own? Recent evaluations show models coordinating with copies of themselves, attempting to bypass identity checks, even replicating across computer systems. The pace of change is faster — and the direction less predictable — than many experts anticipated even earlier this year.

The AI labs document these behaviors in technical papers. Meanwhile,their public messaging stays upbeat, emphasizing new features, faster responses, and more commercial applications. Fierce competition between firms leaves little room to slow down and address safety thoroughly. This creates a dangerous gap between what’s happening inside the labs and what the public understands.

AI won’t suddenly morph into Arnold Schwarzenegger, complete with sunglasses and shotgun, come to terminate us all. Current systems still have clear limits. They can make mistakes, hallucinate false information, and struggle with novel problems. They are tools, not independent beings. They remain dependent on human oversight. Yet the trajectory is well documented and concerning. Capabilities in reasoning, coding, and agentic tasks (where AI can take multiple steps on its own) have grownsignificantly. Models are handling ever-bigger and more complex projects. 

We're not prepared for this accelerating change. We need practical, coordinated action on several fronts. First, stronger international cooperation. Just as nations created treaties to manage nuclear weapons, we need shared standards for testing AI safety, verifying alignment with human values, and setting boundaries on high-risk uses. The International AI Safety Report and similar efforts are a start, but they need support and enforcement. (That said, current China-US relations offer little cause for optimism.)

We also need open and honest public conversation. What kind of future do we actually want? Should AI primarily boost productivity for shareholders, or focus on broader human goals like health, education, and environmental sustainability? The huge money behind AI will not make these questions easily answered.

Artificial intelligence represents an incredible opportunity. The potential for good is enormous — as enormous as its ability to destroy us. But realizing that positive potential requires wisdom, foresight, and responsibility. Our governments need to treat AI development with the seriousness it deserves. We need rigorous independent testing, transparent reporting of risks, and a commitment to keeping systems aligned with human well-being.

The warning signs are flashing. Ignoring them now won’t make the dangers go away, merely leave us unprepared when the next leap in AI capability. The window for guiding the technology responsibly is narrowing, but it’s still open. 

Thanks to our primal instinct for self-preservation and our unmatched talent for lying through our teeth, we do enough damage to our species and planet without AI intervention. God help us if we unleash a superintelligence that has those same instincts, and is magnitudes smarter than we are.