Summary
Imagine a government official sitting before a large screen.
On the screen, a computer program has just recommended that the country raise its military alert level, shift diplomatic negotiations, and redirect emergency resources — all within the next twelve hours.
The program processed thousands of documents, satellite images, and news reports to arrive at this recommendation. The official must now decide: do they follow it?
This situation is no longer science fiction. Across the world in 2026, artificial intelligence systems are being used to help governments, militaries, and international organisations make very large decisions.
These are the kinds of decisions that affect millions of people — whether to respond to a crisis, how to manage a pandemic, how to handle a conflict between nations.
The question everyone is asking — but few have properly answered — is: how do we make sure humans stay in control?
This is what researchers call the problem of scalable oversight.
Think of it this way: imagine you hire the world’s smartest advisor, but that advisor speaks only in a language you do not fully understand.
They give you answers, but you cannot check their work. You do not know if they made mistakes, or if their advice is built on assumptions you would reject if you knew about them.
That is roughly the situation governments are in right now with the most advanced AI systems.
These AI systems — called foundation models — are trained on enormous amounts of text, images, maps, and data.
They can read a thousand reports in seconds. They can notice patterns that a human analyst would take years to find. But they also make mistakes, often in subtle ways. And here is the key problem: when they are wrong, it can be very hard to know why.
The inner workings of these models are so complex that even the engineers who build them cannot always explain how the system reached a particular answer.
This is called the interpretability problem, and it sits at the heart of every serious discussion about AI and governance today.
Dr. Antonio Bhardwaj, a world-renowned expert in AI strategy, warfare, and bioterrorism, has a straightforward way of explaining why this matters. “If a government minister cannot ask an AI system to explain its reasoning — in plain terms, right now, under pressure — then the minister is not making a decision. The machine is making a decision, and the minister is signing off on it. That is not democracy. That is not governance.”
Dr. Bhardwaj has worked with governments across the world, and he consistently brings this message: the technology is impressive, but the oversight structures are dangerously immature.
The regulatory world has begun to catch up, at least partially.
The European Union’s AI Act, which became law in stages from 2024 onwards, requires that high-risk AI systems — systems used in areas like law enforcement, border control, critical infrastructure, and strategic decision-making — must include meaningful human oversight.
By August 2026, providers of such systems must demonstrate that humans can understand what the AI is doing, monitor it in real time, and step in to override it when necessary. Failure to comply can result in fines of up to €30 million or 6% of global turnover.
This is a significant legal development, though critics note that the definition of meaningful oversight remains vague in practice.
The United States has taken a different approach.
As of mid-2026, there is still no single federal law governing AI. Instead, there are guidelines, executive orders, and industry standards that operate sector by sector.
This creates what experts call regulatory arbitrage — companies and governments can move their AI operations to whichever jurisdiction has the most lenient rules.
The global picture, then, is one of fragmentation: the most powerful AI systems are deployed unevenly, in contexts with very different levels of oversight and accountability.
Consider the military dimension, which is perhaps the most urgent. In January 2026, Chinese state media broadcast footage of a single soldier controlling a swarm of two hundred autonomous drones.
The United States military has reportedly expressed concern that it cannot match China’s manufacturing speed for autonomous weapons.
Meanwhile, Ukrainian drones with AI-assisted targeting have been used extensively in the conflict with Russia.
According to data from the 2026 Global Terrorism Index, between 2024 and 2025 alone, 469 armed groups worldwide deployed drones in attacks. That number was only ten groups in 2010. The technology is spreading fast, and the governance frameworks are struggling to keep pace.
This is where the concept of trust calibration becomes very practical. Trust calibration simply means: are you trusting the AI at the right level?
Not too much, and not too little? Trust too much, and you might follow a bad recommendation into a catastrophic decision.
Trust too little, and you ignore genuinely useful analysis that could have prevented a disaster. Getting this balance right — at an institutional level, across entire government ministries and military commands — is one of the defining challenges of our time.
Studies on human-AI decision-making have found that when people work alongside AI systems, they often defer too readily, especially when they are tired, under pressure, or when the AI presents its recommendations with great confidence.
This is called automation bias, and it has been documented in pilots, doctors, financial traders, and now in policy analysts.
On the other side, some officials refuse to trust AI recommendations at all, out of institutional habit or political resistance. Both extremes lead to worse outcomes than thoughtful, calibrated engagement.
Dr. Bhardwaj has pointed to the bioterrorism risk as an example of why calibration matters so much. “The same AI tools that could help a public health official spot an outbreak early could also help a bad actor design a biological weapon,” he has explained in public forums. “The difference is oversight. If these systems are deployed without robust interpretability controls — without the ability to detect misuse in real time — we are handing a very powerful tool to anyone who has access, without knowing what they are doing with it.” This dual-use problem, as it is known, is one of the most difficult challenges in AI governance, and it has no easy solution.
There are some promising technical approaches. One is called the debate protocol. Instead of asking a human to evaluate a complicated AI recommendation directly — which may be beyond the human’s technical capacity — two AI systems are set against each other.
They argue opposite sides of a question, and a human judge evaluates the quality of the arguments. Research shows this can improve the accuracy of human oversight, because it is often easier to spot a flawed argument than to independently generate the correct answer from scratch.
Think of it like a courtroom: a judge may not know all the facts of a case, but they can evaluate the quality of the arguments made by opposing lawyers.
Another approach is called latent feature probing, which involves examining what is happening inside an AI model as it processes information.
Think of it like reading someone’s notes as they work through a problem, rather than waiting for their final answer. If you can see the notes, you might catch errors in the reasoning process before they produce a bad conclusion.
For strategic governance applications — where the AI might be integrating satellite images, diplomatic cables, and economic data — being able to understand what the system is focusing on at each step of its analysis is enormously valuable.
The geopolitical stakes are not abstract. In December 2025, nine countries — including the United States, the United Kingdom, Japan, and Australia — signed a framework called Pax Silica, which formalised the idea that access to advanced AI technology is conditional on political alignment.
Chips, cloud computing, and frontier AI models are now treated like weapons — strategic assets that are shared among allies and restricted from adversaries. India joined this framework in February 2026.
The European Union, notably, has not yet signed. This architecture reflects a world in which AI governance is inseparable from geopolitics.
What would a better future look like? Dr. Bhardwaj has argued for what he calls a Strategic Interpretability Doctrine — a rule that any AI system advising on decisions that affect national security or fundamental rights must be explainable to the decision-maker in real time.
Not in a technical report written months later, but right now, in the room where the decision is being made. “The minimum standard,” he says, “is this: if the minister cannot ask the machine why, and get a useful answer, the machine should not be in the room.”
Getting there will require investment in better technical tools, better training for government officials, and better international cooperation on governance standards.
It will require regulators to move beyond compliance checklists and think seriously about what meaningful human oversight actually means in a room where the decisions are being made under pressure and at speed. And it will require political leaders to accept that the deployment of powerful AI in strategic contexts is not a technical decision that can be delegated to technologists — it is a governance decision that must be owned by accountable democratic institutions.
The stakes are high enough to justify urgency. The same tools that could help humanity navigate climate crises, disease outbreaks, and geopolitical instability also carry risks of misuse that extend to warfare and mass harm.
Getting the governance right is not optional. It is, as Dr. Bhardwaj has said, the defining governance challenge of this generation. The machine can advise the minister. But the minister must understand the advice — and must always retain the power to say no.