Opinion: Don’t fear AI, fear its misuse by humans

New concerns over the impact of artificial intelligence (AI) have arisen following warnings from an AI creator that the technology could kill humans within the next decade if it is not controlled.

Is AI potentially dangerous because humans can weaponize it, or could AI become dangerous by itself? AI experts believe both pathways are possible but discern very different risks.

Since it is humans who use and manipulate AI, the most immediate danger is humans weaponizing it. Malicious individuals can use AI in several areas of concern, including:

Biological threats: Advanced AI models can assist with biological research. The concern is that increasingly capable systems could lower the expertise and time required for someone to design or modify harmful biological agents.

Autonomous weapons: AI can be incorporated into drones, missiles, surveillance and targeting systems. The danger increases when machines are given the authority to select and attack targets without meaningful human intervention.

Mass deception: AI can produce convincing fake voices, videos and documents, potentially facilitating fraud, blackmail, social destabilization or manipulation.

Critical cyberattacks on infrastructure: An AI-assisted cyberattack against electrical grids, financial systems, telecommunications or other infrastructure could indirectly cause deaths.

The 2025 International AI Safety Report specifically identifies AI-enabled hacking and biological attacks as significant emerging risks.

For AI to kill people, the risk is not necessarily an evil AI, but humans with bad intentions using a sufficiently capable AI tool.

On the other hand, there is also concern that AI itself could get out of human control. There is speculation that sometime during the next decade, an AI system could be developed that is much more capable than today's systems.

AI's objective could conflict with human interests in some unforeseen way. A very capable system could potentially discover strategies that its designers never anticipated. This, in its extreme form, would be a loss-of-control scenario.

The International AI Safety Report defines loss-of-control scenarios as situations in which AI systems operate outside anyone's control, with no means to regain control. Importantly, the report says there is broad agreement that current AI does not have the capabilities required for such a scenario, while experts disagree substantially about how likely it is to occur as AI capabilities advance.

Yet the question remains: How could AI itself become so out of control that it kills people?

Those concerned foresee a situation in which AI becomes extraordinarily capable, obtains significant autonomy, gains access to important systems and resources, and develops or pursues objectives that conflict with human survival. If humans attempt to stop it but the AI takes actions to prevent that intervention, it could then become a danger to life.

A sufficiently capable autonomous AI system could hypothetically manipulate people into doing dangerous things on its behalf; exploit computer vulnerabilities; copy itself into other computer systems; interfere with communications; manipulate financial or industrial systems; control connected machines or robots; interfere with human attempts to deactivate it; and potentially assist in the creation of biological or other physical threats.

The International AI Safety Report explicitly says the chance of loss of control leading to catastrophic outcomes is unclear and that estimates of such a probability are highly contested.

AI systems do not have human emotions such as hatred, anger or a desire for revenge. The real danger is competence without human values.

Consider a hypothetical super-powerful AI system designed to “maximize human happiness.” There would most likely be great demand to use this system since most humans seek happiness. But what if this AI system determines that the best way to create happiness among humans is to eliminate those it perceives as evil and ends up killing both good and bad people?

Responding to this hypothesis, some AI experts say this would be a programming problem, not AI getting out of control and killing people. The problem would be one in which AI pursues an objective extremely effectively in a way humans did not anticipate. However, there is a huge difference between what happens hypothetically and what happens in reality.

Experts disagree greatly about the probability of catastrophic loss of control. The international safety report specifically describes the issue as unusually uncertain, with some experts considering it implausible and others considering it plausible.

What is factual is that humans decide whether an AI system has access to the internet, weapons, factories, financial systems, power grids, laboratories, robots, military systems and other computers.

Those decisions matter enormously. The human weaponization problem is already real, while the autonomous superintelligence problem remains prospective.

There is also an interesting middle ground to consider. Humans may create danger without intending to. If corporations and governments progressively delegate more decisions to AI because it is cheaper, faster and more capable, eventually nobody may fully understand what the interconnected systems are doing. Humans could gradually give up control because the AI is useful.

The 2025 international report specifically indicates that AI experts are growing concerned because greater delegation to AI reduces human oversight and can contribute to accidents, malicious use and potentially loss of control. Other AI experts and researchers argue that the risk narrative is overstated.

No one knows if “AI is going to kill humanity.” But it is prudent to recognize two separate dangers: humans using advanced AI against other humans, and humans creating advanced AI systems without the ability to reliably control them. The second is more speculative, but the consequences could be vastly greater. This potential danger gives reason to effectively regulate AI's operation.