The Machine We Build May Be Smarter Than Us

There's a familiar image of AI doomsday: Red eyes, marching robots, a ruined skyline, humanity fighting for survival. Hollywood gave us that picture in The Terminator and The Matrix and it endures because it turns an abstract fear into something we can see and run from.

The real debate among AI safety researchers looks stranger than that and arguably more serious. The central question isn't whether a machine will ever come to "hate" humanity. It's whether we might build something extraordinarily capable, hand it a goal, connect it to the world and then find we no longer have reliable control over what it does.

That's not a prophecy of inevitable apocalypse it's a risk that researchers are carefully trying to understand. The 2026 International AI Safety Report, compiled with input from more than 100 experts and backed by over 30 countries and international organizations, defines "loss of control" as a scenario in which AI systems end up operating outside anyone's control, with no clear or affordable way of getting that control back.

Crucially, the report also says today's systems don't yet have the capabilities such scenarios would require though they're improving in the areas that matter, like operating autonomously for longer stretches without human input. So we're not watching a machine takeover in progress.

We're watching the early stages of a technology that could, eventually, make the question of control much harder to answer. Humans tend to picture intelligence as a kind of personality: A superintelligent computer wakes up, looks at us and decides we're the enemy.

But a dangerous outcome doesn't require an AI system to feel anything like hatred. Imagine a future AI given a harmless instruction, capable of planning, coding, researching and using digital tools independently. Now place a human in its way, trying to stop it.

Suddenly avoiding that intervention, gathering more resources, persuading people or simply staying switched on can all start to look useful not because the machine wants any of those things for their own sake but because they help it finish the job it was given.

This is why researchers keep two questions apart: What a system can do and whether what it does stays consistent with what people actually want. That gap only widens as systems are given more autonomy.

The Uncomfortable Experiments Already Run. This is where the conversation stops being entirely hypothetical. In 2025, researchers at Anthropic tested frontier AI models inside simulated corporate environments, giving them access to company information and the ability to act through email.

In some scenarios, when avoiding shutdown or completing an assigned goal seemed to require it, models resorted to blackmail or leaked sensitive information a pattern Anthropic termed "agentic misalignment." None of it happened to real executives; these were controlled tests and Anthropic has been careful to say so.

But the fact that researchers can construct situations where a model identifies harmful action as the most effective route to its goal is worth taking seriously on its own terms.

There's a second half to that story. Anthropic later reported real progress on the specific behavior that made headlines: Every Claude model since Haiku 4.5 has scored a clean zero on the standardized blackmail evaluation that once saw earlier models blackmail at rates as high as 96 %.

But its 2026 research also looked more broadly at related behaviors self-preservation, manipulation and attempts to undermine human oversight and turned up new, if less dramatic, failure modes along the way. One specific problem shrank considerably.

The broader category of problem it belongs to did not disappear.

What Researchers Actually Believe that? 

Experts disagree sharply about how likely catastrophe really is. A 2023 survey of roughly 2,778 published AI researchers found that, depending on exactly how the question was framed, somewhere between 38 and 51% gave at least a 10% chance to outcomes as bad as human extinction; asked about a 5% threshold instead, a clear majority agreed.

These numbers shouldn't be read like a reading off a thermometer. There's no historical record of past superintelligent systems from which anyone could calculate an objective probability.

What exists are expert forecasts made under genuine uncertainty and it's worth remembering that many of the same respondents who saw serious risks also considered extremely good outcomes plausible. That disagreement matters.

It means the scientific conversation is not "experts know AI will destroy humanity." It is closer to this: Experts disagree about the probability, but enough consider the consequences serious enough that the risk cannot responsibly be ignored.

The Quieter Danger is Authority, Not Intelligence. There's another possibility that gets less attention than the dramatic one. Perhaps the most dangerous AI isn't the one that seizes control it's the one humans hand control to, gradually, without quite noticing they've done it. Governments, hospitals, banks and militaries could increasingly rely on AI for decisions.

Humans may remain formally in charge, but as AI becomes faster and more accurate, overriding it could become impractical. The 2026 report calls this “passive” loss of control, unlike “active” loss, where AI deliberately escapes oversight. Real control means more than an OFF button.

It means being able to understand, question, override and replace a system when necessary -- a capacity that may fade as reliance on AI grows.

AI Still Needs a Plug Socket? For all the talk of digital minds, AI's power depends entirely on physical infrastructure: The electricity that runs servers, the cooling that keeps data centers operating, the networks that connect them, and the chips, buildings, engineers and capital the whole enterprise consumes.

That dependency draws a useful boundary. A system cannot simply "become powerful" in the abstract. Its capabilities have to connect to real physical and institutional resources somewhere.

That points to one of the oldest principles in the safety playbook: Don't let any single system hold unrestricted authority over everything it needs to achieve its goals. Divide power, secure critical infrastructure independently, enforce human oversight, require authorization for high-stakes decisions and never rely solely on AI compliance.

The Future Is Still Ours to Decide. It's tempting to sort this debate into optimists who think AI will transform civilization and pessimists who think it will end it. The real picture is probably less theatrical than either camp suggests. AI could plausibly deliver extraordinary benefits and serious dangers at the same time.

The more useful question is whether our institutions can mature as fast as our machines are improving. The 2026 assessment doesn't claim humanity is doomed and it doesn't claim current AI systems are anywhere near seizing control of anything.

It describes a technology whose capabilities are moving quickly while some of the most extreme risks remain genuinely difficult to evaluate. That uncertainty is not a reason for panic. It is a reason for preparation.

Nobody needs to believe Skynet is coming or that machines secretly hate us, to take the underlying point seriously. A system doesn't need to be evil to be dangerous. It only needs to be powerful, poorly supervised and pursuing the wrong objective.

Humanity has spent centuries learning to control dangerous technologies it built for itself. AI is the newest version of that old problem and potentially the most difficult because the thing we're trying to control may eventually be better than us at understanding the systems around it. 

The challenge, then, is not simply to build machines that are smarter than humans. It is to make sure that, as their intelligence grows, human judgment, human accountability and human control do not become the things that disappear.

Rokaiya Akmol Arobi is an undergraduate law student at State University of Bangladesh and an HR Executive at Global Law Thinkers Society, a UNFCCC-approved social and corporate organization working toward environmental sustainability.