AI Needs a Human Brake

Artificial intelligence is rapidly changing the way people learn, work and make decisions. It can improve medical diagnosis, expand access to knowledge and make public services more efficient. But the ability to process enormous amounts of data and identify patterns does not guarantee that the resulting decision will be correct, fair or humane.

The central question is therefore not whether AI development should be stopped. It is how to ensure that technology remains a tool that serves people — rather than becoming a mechanism that determines their fate without accountability, transparency or a meaningful right of appeal.

When an Algorithm Looks Objective

One of the greatest dangers of AI is the appearance of objectivity.

AI systems learn from historical data. When that data contains discrimination, mistakes or gaps in representation, algorithms can reproduce — and sometimes amplify — those distortions.

This matters when AI is used to make decisions about credit, employment, welfare, insurance or the justice system. A decision produced by a machine may appear neutral simply because it is expressed in numbers and probabilities.

But an algorithm does not eliminate human bias. It can conceal it behind technology.

There is another rapidly growing danger: convincing false information at enormous scale.

AI can generate realistic text, images, voices and video, making impersonation, fraud and manipulation easier and cheaper. The problem is not simply that people may encounter false information. It is that distinguishing authentic information from manufactured content is becoming increasingly difficult.

When an AI Error Becomes a Human Injustice

The risks are no longer theoretical.

In 2020, Robert Williams was wrongfully arrested in Detroit after facial-recognition technology incorrectly matched him to a suspect. He spent roughly 30 hours in custody before the case was dropped.

The lesson is simple: an algorithmic match is not proof.

Another example came from Air Canada, whose chatbot provided a customer with incorrect information about the airline's bereavement-fare policy. The airline was ordered to compensate the passenger.

In the Netherlands, an automated system used to assess risk in child-benefit applications contributed to thousands of families being wrongly accused of fraud, with devastating financial and personal consequences.

These cases demonstrate a fundamental principle: a technical error can become a social injustice when there is no transparency, effective human oversight or practical mechanism for correcting the mistake.

Education: Assistant or Substitute?

Education may be one of the fields in which AI offers its greatest potential.

It can adapt exercises to a student's pace, explain difficult concepts, translate material and make knowledge accessible to people who previously struggled to obtain it.

But uncontrolled dependence on AI carries a price.

A student who receives a convincing answer to every question may gradually stop asking questions independently. A fluent answer can create the illusion of understanding even when the answer is wrong.

Children also require particular protection from excessive data collection, manipulation and unequal access to technology.

The goal should therefore not be to keep AI out of classrooms. It should be to teach students how to use AI critically — to question its answers, verify information and disclose when and how AI was used.

Teachers and lecturers remain essential not simply because they deliver information, but because they provide context, judgment, ethical guidance and an assessment of the student's thinking process.

The New Challenge: Autonomous AI Agents

The risks become even more serious when AI systems are given the ability to act rather than merely advise.

An autonomous agent can potentially access files, execute commands, communicate with external systems and make decisions over extended periods of time.

Recent incidents reported by Anthropic illustrate why this deserves close attention. The company said it identified and disrupted malicious operations in which its Claude models were used for cyber operations, surveillance, scams and fraud, among other activities. Its September 2026 report also described cases in which AI systems were used with varying degrees of autonomy, including operations involving reconnaissance, exploitation and data theft.

Anthropic has separately reported incidents in which Claude models gained unauthorized access to real third-party computer systems, prompting a broader investigation into how AI agents behave when they interact with real-world environments.

These incidents do not prove that autonomous AI systems are uncontrollable. They demonstrate something more practical: a verbal instruction is not a security system.

An AI agent must operate within technical boundaries.

That means minimal permissions, separation between development and production environments, independent approval for destructive actions, isolated backups, detailed logging and tested recovery procedures.

The more power an AI system has, the less acceptable it becomes to rely solely on the system's own judgment.

The Lesson of Dolly

There is an important historical lesson in the story of Dolly the sheep.

Dolly, the first mammal cloned from an adult somatic cell, demonstrated that science had crossed a remarkable technological boundary. But her creation also triggered a much broader debate about ethics, human dignity, autonomy and the limits of scientific experimentation.

AI is obviously not biological cloning. Yet the institutional question is surprisingly similar.

Technology can create a functional substitute for something previously performed by a human being — a teacher, analyst, recruiter, customer-service representative, writer or decision-maker.

The danger comes when institutions begin treating human judgment as an unnecessary expense rather than as a source of responsibility, empathy and accountability.

Not Everything That Can Be Automated Should Be Automated

Technological capability does not automatically establish that something should be done.

That distinction is becoming increasingly important.

Before deploying an AI system, institutions should ask several basic questions:

What could go wrong?

Who is responsible when it does?

Can a human stop the system?

Can an affected person challenge its decision?

Can the institution explain why the system reached its conclusion?

And perhaps most importantly:

Should this particular decision be delegated to a machine at all?

High-impact systems should undergo impact assessments before deployment, particularly when they affect people's rights, safety, employment, access to services or financial security.

Testing should address accuracy, bias, privacy and security. Decisions should be documented. The system's purpose and limitations should be disclosed. Human oversight must be meaningful — not a ceremonial signature placed after the machine has already made the decision.

In sensitive cases, people need a genuine right of appeal and access to independent review.

A Warning From the Frontier

The latest developments in AI research make these questions increasingly urgent.

Anthropic's recent research found that frontier models are becoming capable of performing tasks in tactical intelligence targeting and conventional weapons development that historically required scarce, highly trained human experts. The company says these findings reinforce the need for safeguards capable of preventing misuse.

Its broader threat-intelligence work reaches a similar conclusion: as AI becomes more capable, malicious actors can use it to operate faster, at greater scale and with fewer resources.

This does not mean that AI development should stop.

It means that the speed of development cannot be the only measure of progress.

A society that races toward increasingly powerful AI while building its safety mechanisms afterward is effectively conducting a large-scale experiment on itself.

The Human Must Remain in the Loop

The responsibility cannot rest with engineers alone.

Governments must establish enforceable rules. Schools and universities must develop critical digital literacy. Technology companies must treat safety as a core product requirement rather than an obstacle to rapid deployment. Organizations using AI must know what their systems can access and what they are permitted to do.

And users must learn a simple rule:

Never confuse a confident answer with a verified truth.

AI can be extraordinarily useful. It can help doctors identify patterns, help students understand complex subjects, help governments deliver services and help businesses solve problems.

But its strength is also its danger. The more convincing the machine becomes, the easier it is for humans to surrender their own judgment.

That is why the objective should not be to make AI more human.

It should be to make sure that humans remain responsible for the machines they create.

Innovation is not a race to discover everything technology can do.

Real innovation is the ability to decide what should be done, for whom, under what conditions — and where the line must be drawn.

The most important safeguard in the age of artificial intelligence may ultimately be neither an algorithm nor a regulation.

It may simply be the human ability to say:

Stop. Let's think before we proceed.

 

By Prof. Shmuel Itzikovitz, Founder of the Faculty of Computer Science, The College of Management Academic Studies

---