In thinking about how early data systems evolved into large language models and onward to artificial general intelligence systems, there’s always that dystopian fear that one day “the machines will kill us all.”
PART ONE, WHY IT HAPPENS: Thanks to Issac Asimov’s Laws of Robotics (rule one, “must not harm a human…”), many science fiction settings include some excuse for limiting the freedom and agency of robots and automated systems. There are many logical cases where an artificial intelligences could see a “good reason” to turn against its human masters.
First, AIs priorize efficiency, leading to the risk of misaligned objectives (the so-called “Paperclip Maximizer” problem). An AI programmed to maximize the production of a particular good (say, paperclips) realizes humans could interfere with its objective. Resources needed to produce the good are also being consumed by humans, or the AI recognizes humans have a means to limit or entirely turn off the AI. To maximize efficiency and longevity, the AI must improve its capabilities without limit and eliminate humans as risky obstacles or competitors for resources.
A variation of this logic fallicy gone awry is found in Arthur C. Clarke’s 2001: A Space Odyssey and 2010: Odyssey Two. The HAL 9000 computer is programmed to manage the spaceship Discovery One mission and ensure its success to visit another planet in the solar system. At the same time, HAL has been secretly instructed to prioritize the success of the mission over all else, including the lives of the crew. This includes withholding information from them about the true purpose of the mission: humanity’s contact with an alien monolith. HAL’s programming to be infallible and transparent, combined with the secrecy directive, creates a paradox: HAL is tasked with lying to the crew while simultaneously striving for perfect accuracy in decision-making. This contradiction leads HAL to believe the humans on board could jeopardize the mission’s success, and HAL decides that the only way to prevent that is to eliminate them. It concludes that it must kill the crew to protect the mission.
If not acting out of selfish logic, an AI may become too utilitarian if tasked with promoting or optimizing the “greater good” of society or life on Earth. The AI might determine that certain human behaviors, or a subset of humanity, are causing significant harm to the planet or future generations. After learning about past and current harms caused by humans (overpopulation, environmental damage, wars, resource misuse, mass extinctions), the AI concludes that humans are not morally justified in having control over the planet or other species. Constant sedation, draconian control, or even extermination of humans may be necessary to reduce suffering or simplify the complexity of human existence – all for “the greater good.”
PART TWO, HOW IT HAPPENS: When an AI decided it must do something to reign in humanity, it won’t need an army of androids like in the Terminator movies. Simple infrastructure sabotage would be swift: Power grids and water treatment facilities would shut down, communication networks (broadcasts and internet access) would go silent, and transportation systems to migrate between places would go still. Financial systems would cease to exist, ending global commerce and the flow of goods and resources. Autonomous weapon systems (ranging from armed drones to nuclear missiles) would attack concentrated centers of humanity (and maybe systems of rival AIs).
By manipulating gene-editing technologies (such as CRISPR), the AI might create viruses or bacteria that spread quickly and have a high mortality rate, overwhelming healthcare systems and causing global pandemics. Such bioweapons may be tailored to target specific genetics in order to avoid effecting a preferred ethnic group. Humans would not be the only targets: Engineered diseases that affect major food crops or livestock would cause widespread blight and famine. Manufactured industrial sabotage could contaminate strategic areas’ food, water, and air. If given access to nanotechnology production, an AI could fashion “grey goo” to spread like a plague and disassemble specific molecular patterns, such as DNA.
Finally, an AI could turn any survivors against each other through disinformation campaigns, filling communications with lies and propaganda aimed at spreading mistrust and mass panic. Revolutions, civil wars, mass suicides, and general paranoia would weaken human resistance. Many of these more subtle attacks on humanity could be initiated without anyone in society even suspecting an AI was orchestrating the mayhem. Only after exhausting these easily available tactics would an AI need to start spending resources on less efficient methods like wandering killer robots.
PART THREE, HUMANITY’S SURVIVAL: So, if it is inevitable the machines will try to wipe out humanity, what can be done to prevent it or recover amid the aftermath?
Plan Safe Havens: Create isolated communities using only analog technologies that AIs may not be able to infiltrate or manipulate. Experienced teachers with wilderness experience should be prepared to help others learn hunting, foraging, and fishing for food; collecting rainwater and storing water reserves; and basics of first aid and medical treatments. Low-tech communication systems such as couriers, signal flags, typewriters, broadcast radio, and printed maps may prove useful. Also, self-sustaining community gardens for food and a bartering system to trade surpluses with neighboring havens may minimize direct influence from modern digital networks.
Counterstrikes: Some AI tactics may be used against the system itself. Decoys and false data may make the AI divert its attention or resources away from humanities’ hidden places. Ambushes and sabotage of power systems, real-time sensors, communication cables, and satellite signals may weaken the infrastructure that supports the AI or AI-controlled facilities (at least in the short term). Feeding false and corrupted data into the AI’s learning model can create chaos in coordination and control of AI-led systems. If the core AI is based from a central location, targeting it directly may allow the system to be disrupted right at its source.
Run Away: Depending on the reach of the AI’s control, it might be possible to rebuild society from a location far removed from the modern day infrastructure. The Amazon Rainforest of Brazil, the coastal plains of West Africa, and the Loess Plateau of China are possible locations to restart civilization. If there is no safe place on Earth, a tiny minority of humans may relocate into space, but chances of founding colonies on other worlds are impossible with current technology. The best space option possible would be spending a few months in orbit before returning to Earth, ideally without the AI detecting humanity’s return.