Can We Put the Tiger Back in the Cage?

We are trying to regulate a technology that is moving faster than the institutions trying to control it.

In Part 5, we looked at what happens when the AI Tiger enters warfare. Now comes the uncomfortable question: can we actually put it back in the cage?

My blunt answer: probably not. Not in the way we regulated nuclear weapons. Not with one treaty. Not with one government. And certainly not by asking technology companies to voluntarily slow down while their competitors continue accelerating.

The AI Tiger is fundamentally different. It is software. It can be copied, distributed, improved, open-sourced. And increasingly, it can operate autonomously. That makes AI extraordinarily difficult to contain.

The cage was designed for a different kind of beast

Humanity has experience regulating dangerous technologies. Nuclear weapons require enormous physical infrastructure. Chemical weapons require controlled substances and industrial capabilities. Aircraft require physical manufacturing. Even conventional weapons require factories, materials, and supply chains.

AI is different. A capable model can potentially be copied, modified, compressed, distributed, and deployed — and once knowledge, techniques, or model weights escape into the world, putting them back may be impossible. You cannot un-invent an algorithm. You cannot unpublish knowledge. You cannot easily make millions of copies disappear. That is the fundamental governance problem.

The world is already trying

This is not science fiction. Governments are already building AI regulation. The European Union has its AI regulatory framework. The United States is developing federal and state-level approaches. China has its own regulatory architecture. The United Nations launched its first Global Dialogue on AI Governance in July 2026, bringing governments and stakeholders together to discuss international cooperation, and has established an Independent International Scientific Panel on AI to create an evidence base for policymakers.

That is important. But there is a problem: we are building the governance system while the technology itself is changing underneath it.

Regulation has a speed problem

Imagine a government spends three years designing legislation for today's frontier models. By the time the legislation takes effect, today's models may look primitive. Agents may have changed. Compute architectures may have changed. Open-source models, AI-to-AI communication, autonomous systems, and the economic value of AI itself may have shifted completely.

The law is trying to regulate a moving target — and the target is moving faster every year.

The most dangerous regulation may be no regulation

There is an understandable argument: don't regulate too much, we don't want to slow innovation. I understand it — AI could dramatically improve healthcare, education, productivity, scientific research, and economic growth.

But there is another question: what happens when every country makes the same calculation — "if we slow down, somebody else will win"? That is exactly the dynamic from Part 2. The AI race creates a regulatory prisoner's dilemma. Country A wants stronger safeguards; Country B sees an opportunity. Company A wants safety; Company B wants market share. Military A wants human oversight; Military B wants faster autonomous decision-making. Everyone may privately agree that some limits are necessary. Yet individually, everyone has an incentive to keep accelerating. That is how dangerous races continue.

Voluntary safety has a ceiling

This is where I become particularly blunt. I don't believe humanity should depend entirely on the goodwill of AI companies to keep humanity safe — not because AI companies are evil, but because they are companies. They have investors, employees, competitors, market expectations, national interests, and increasingly, hundreds of billions of dollars riding on continued AI development.

Asking companies to voluntarily stop at exactly the point where their competitors continue is an unstable strategy. This debate is becoming more explicit — OpenAI publicly called in September 2026 for mandatory national AI safety requirements, including independent assessments, cybersecurity measures, and incident reporting, arguing that voluntary commitments are insufficient. That is a remarkable development: the industry is beginning to acknowledge that self-regulation alone may not be enough.

But governments have a problem too

Governments are supposed to regulate AI. But governments also want AI — economic growth, military advantage, cyber capabilities, intelligence, scientific discovery, sovereign AI, better public services, productivity, national competitiveness.

So the same government that says "we need to control AI" may simultaneously be saying "we need to win the AI race." That contradiction is at the heart of the problem. You cannot simultaneously demand that everyone slow down and demand that your own country move faster.

And then there is the military problem

This may be the hardest problem of all. Governments are already debating autonomous weapons and AI-enabled military systems. In September 2026, UN discussions involving 128 states were struggling over whether future rules should be legally binding and how much meaningful human oversight should be required.

Think about the dilemma. Suppose Country A agrees that humans must approve every lethal action, but Country B develops a system that can identify and respond to threats in seconds. Country A now has two choices — keep the rule and potentially lose the battlefield advantage, or relax the rule. This is exactly why military AI may become one of the hardest areas to regulate.

We don't need a global AI government

Some people hear "global AI governance" and imagine a giant world government controlling every AI system. I don't think that is realistic, and I don't think it is necessary. What we need is something more practical — a global safety architecture.

Think aviation. There is no single global airline. No single country controls every aircraft. But countries agree on standards. Airports follow procedures. Aircraft are certified. Pilots are trained. Incidents are investigated. Systems are audited. Information is shared. Certain behavior is prohibited. AI may eventually need something similar.

What should actually be controlled?

This is where I believe the conversation needs to become much more specific. We shouldn't try to regulate "AI." We should regulate capability and consequences.

Extremely large training runs may require additional reporting, security, and oversight. Models crossing defined capability thresholds could require independent evaluation before deployment. An AI that can execute actions should face stronger controls than one that only generates text. Healthcare, financial systems, critical infrastructure, weapons, and biological research require higher standards. Every autonomous agent should carry an identity, defined permissions, limits, an audit trail, and a kill switch. Serious AI failures should be reported much the way major cybersecurity incidents are. Independent organizations should test frontier systems for dangerous capabilities. And compute, chips, and data centers may eventually become part of the governance framework themselves.

The kill switch problem

Here is a question I don't think enough executives ask: can you actually turn your AI off? Not theoretically — practically.

Imagine an enterprise with 500 AI agents, 20,000 automated workflows, millions of daily decisions. AI managing customer operations. AI writing code. AI managing infrastructure. AI negotiating transactions. AI monitoring security. AI communicating with other AI systems. Then something goes wrong.

Who presses the button? And what exactly does the button disconnect — the model, the agents, the APIs, the cloud infrastructure, the autonomous workflows, the replicas, the backups, the downstream systems? A kill switch is only useful if the system is actually designed to obey it.

The open-source problem

This is where regulation becomes even harder. Suppose the most powerful model in the world is controlled by a company — regulators can potentially negotiate with that company. But what happens when a similar capability is available openly? Who do you regulate — the developer, the distributor, the person fine-tuning it, the cloud provider, the user, the infrastructure provider? And what happens when someone modifies it? You cannot regulate software the same way you regulate a factory.

The real question is: who gets to decide?

This is perhaps the biggest governance question of the AI era. Who decides what an AI is allowed to do — governments, technology companies, scientists, military organizations, international bodies, citizens, or the AI itself?

Because eventually we may have systems capable of making decisions that humans cannot fully understand in real time. And that creates a frightening possibility: human beings remaining legally responsible while increasingly relying on decisions they cannot independently verify. That is not human control. That is human approval of machine judgment. There is a difference.

We are already seeing the governance gap

The warnings are no longer purely theoretical. AI companies are reporting increasingly sophisticated misuse and security incidents. Researchers and policymakers are debating autonomous agents, cyber operations, biological misuse, and advanced model control. And even AI companies themselves are increasingly calling for stronger external rules.

That doesn't prove AI is uncontrollable. It proves something else: the gap between capability and governance is becoming a serious policy problem.

So can we put the tiger back in the cage?

Here is my answer: no. Not completely.

The Tiger has already escaped. There are too many models, too many companies, too many countries, too much capital, too much open research, too much infrastructure, too much strategic competition, and too much potential economic value. Trying to put everything back into the cage may already be impossible.

But that doesn't mean we are powerless. There is another option: learn to live with the Tiger.

The new objective isn't containment

Maybe we have been asking the wrong question. Instead of "how do we stop AI?" we should ask: "how do we make sure AI never becomes powerful enough to control the humans who created it?" That means designing systems around capability, control, and accountability together — not capability alone, not speed alone, not profit alone, not national advantage alone.

My view

I don't believe a single global AI treaty will magically solve this. I don't believe governments can stop AI development. I don't believe companies can regulate themselves indefinitely. And I don't believe humanity can simply put the Tiger back in its cage.

But I do believe we can build guardrails around it — international standards, independent evaluations, compute governance, agent permissions, human accountability, military boundaries, incident reporting, AI identity, auditable systems, emergency shutdown mechanisms. And above all, international cooperation between countries that fundamentally distrust each other.

That last one may be the hardest. Because the biggest AI race is increasingly not humanity versus AI. It is humans versus humans, using AI.

---

The question we cannot avoid: we have spent five articles watching the Tiger grow. We released it. We started riding it. We began feeding it. We taught it to hunt. And now we have taken it to war. So perhaps the most important question is no longer "can we control the Tiger?" Perhaps it is: can we learn to live with something more intelligent, more powerful, and potentially more autonomous than anything humanity has ever created? Because if the answer is no, we have a much bigger problem than regulation.

About the author: Mohamed Ismail is an AI thought leader with 26+ years in enterprise transformation and solution architecture on AI, Cloud & Data. He writes on the future race on AI and its consequences.