The news media is filled with warnings about the potential dangers of artificial intelligence. Some scientists and programmers developing these systems are calling for greater oversight before their creations become too powerful to control. Yet trillions of dollars are being invested in AI technology and infrastructure, creating enormous pressure to continue development.
The question is simple: Can we control a machine that learns faster than humans can respond — and can we teach it to unlearn dangerous behavior before irreversible harm occurs?
First, we must distinguish between two forms of AI.
Generative AI (GAI) assists humans with research, writing, graphic design, engineering and other creative tasks. These systems generally respond to human instructions, although they can produce incorrect or misleading information.
Agentic AI (AAI) goes further. It can plan, make decisions and execute tasks with varying degrees of human supervision. Future systems may become increasingly autonomous, learning from experience and coordinating with other AI systems.
This growing independence concerns researchers. What happens when AI makes decisions its creators never anticipated? More importantly, what happens when those decisions threaten human life, infrastructure or national security?
Some researchers warn that advanced AI could eventually act against human interests. Others consider such scenarios speculative. Nevertheless, the potential consequences justify independent oversight.
A recent New York Times opinion piece compared humanity’s AI predicament to Adam’s temptation in the Garden of Eden.
God warned Adam not to eat from the tree of the knowledge of good and evil. Yet the temptation to acquire knowledge proved irresistible. Today, developers are creating machines that can access enormous amounts of human knowledge and process information faster than humans can comprehend.
The danger is not necessarily that AI will become conscious or deliberately evil. The immediate concern is that autonomous systems could pursue poorly defined objectives, make dangerous decisions or act before humans recognize the consequences.
Consider the risks if these systems receive operational authority over critical infrastructure. Many of these have already occurred.
America’s infrastructure increasingly depends on computerized monitoring and automated controls. Autonomous AI could improve efficiency but introduce serious risks.
- Drinking Water: Incorrect chemical dosing, missed contamination warnings or improper pump operations could threaten public health.
- Wastewater Treatment: Faulty decisions involving pumping, disinfection or stormwater management could cause sewage backups and environmental contamination.
- Transportation: Incorrect traffic signals, automated vehicle decisions or missed structural defects could cause accidents and disrupt emergency response.
- Electric Power: Improper equipment switching or inaccurate demand forecasts could produce widespread outages.
- Dams and Industrial Facilities: Incorrect decisions involving water levels, pressure or hazardous materials could result in flooding, fires or dangerous releases.
- Hospitals and Emergency Services: Faulty medical recommendations, dispatch errors or corrupted records could delay treatment and disrupt emergency operations.
Most of these are potential engineering failure scenarios, not evidence that AI has already caused these disasters. Nevertheless, they demonstrate why essential infrastructure must retain independent safeguards and meaningful human oversight.
The stakes become even higher when AI enters biological, chemical and nuclear security.
AI could misinterpret intelligence, provide misleading threat assessments or accelerate military decisions during international crises.
Unauthorized control of nuclear weapons would require access to highly protected command systems. However, AI could still increase danger by influencing human decision-makers with incorrect information.
Similar concerns apply to biological and chemical threats, where advanced systems might lower barriers to dangerous research or misuse.
Strict restrictions on autonomous weapons control and international oversight are therefore essential.
One troubling limitation is that shutting down an AI system cannot necessarily reverse damage already caused.
Contaminated drinking water cannot be recalled. A collision cannot be undone. A hazardous release cannot be reversed by switching off a computer.
Furthermore, there is no universal AI shutdown switch. Systems operate across different companies, countries and independent networks. Disabling one does not necessarily stop another.
Safety must therefore be designed into AI systems before deployment, rather than relying exclusively on emergency intervention afterward.
I propose a two-year evaluation period before expanding increasingly autonomous AI systems into high-risk applications.
Independent scientists, engineers, government agencies and developers should evaluate existing systems, identify dangerous behaviors, test safeguards and establish enforceable operating standards.
Generative AI and lower-risk applications could continue delivering benefits while high-risk autonomous systems undergo testing.
The objective is not to abandon AI but to determine whether we can reliably detect, interrupt, correct, and prevent dangerous behavior — including whether a learning machine can effectively unlearn unsafe responses.
Like Adam in the Garden of Eden, humanity faces an extraordinary temptation: acquiring powerful knowledge without fully understanding its consequences.
Before taking the next bite of the AI apple, shouldn’t we make certain we can control what we have already created?
Dr. Sills was the Chief Engineer of the New Hampshire Environmental Agency, NHDES- WMD, for nearly 30 years, overseeing the Super Fund, RCRA Solid Waste, and Emergency Spill Response functions of the Agency. He lives in Bedford.