Modern solutions could save the day in many AI disaster films, but real-world challenges persist

As artificial intelligence (AI) evolves, Hollywood’s love affair with robots continues. One of the latest entries into the AI film genre, “The Wild Robot,” based on the bestselling book of the same name, is in contention for an Academy Award this year.

It’s the rare film depiction of a “good” machine—one that develops human-like empathy and cares for other living things.

Moviegoers might be more familiar with AI films featuring darker storylines. In “Subservience,” “M3gan,” “Ex Machina” and the Stanley Kubrick classic, “2001: A Space Odyssey,” AI-powered villains attack us humans one by one. In “I, Robot” and blockbuster franchises such as “The Terminator” and “The Matrix” films, intelligent systems pursue our collective enslavement or extinction.

But these themes, though entertaining, could be undermining AI literacy.

Phaedra Boinodiris, Global Leader for Trustworthy AI at IBM Consulting, says that filmmakers have largely concentrated on implausible problems that pull focus from real-life AI issues.

People might get “the perception that it is only Skynet and HAL that could possibly cause harm,” she said, referring to the antagonists in “The Terminator” and “2001.”

The systems in these and other movies depict artificial general intelligence (AGI)—hypothetical AI systems that can rival or exceed the cognitive abilities of human beings. The public’s fixation on such sci-fi films “may end up exacerbating this lack of understanding about the real risks of AI right now.”

Those real risks include biased and otherwise harmful outputs that have been blamed for substantial harm to humans. Efforts by ethicists and developers notwithstanding, such issues are proving complex and confounding.

One study, for instance, found that even when generative AI models exhibit less overt racial bias, they demonstrate more covert racial bias. However, what would be less challenging is preventing the kinds of existential crises imagined in science fiction films.

Real-world solutions that provide guardrails for today’s AI models and tools could effectively quash the machine rebellions of some of Hollywood’s best movies before theatergoers even touched their popcorn.

Let’s take a closer look.

With apologies to James Cameron, it turns out that averting the filmmakers’ visions for an AI-driven apocalypse doesn’t require a gun-toting warrior. Engineers and programmers could get the job done with fairly run-of-the-mill tools and strategies such as:

• Geofencing
• Remote kill switches
• Limits on code execution and connectivity
• Verification
• AI alignment
• Decentralization

The GPS-powered technology currently limiting motorized scooter movement in some cities could also help control the travel abilities of AI robots. Geofencing creates virtual boundaries that GPS or RFID-equipped devices cannot cross.

In films such as “Ex Machina” and “Subservience,” geofencing could have effectively stopped ill-intentioned humanoid robots (played by Alicia Vikander and Megan Fox) from leaving their respective homes and thwarted some of their dangerous plans.

Ben Gelman, a senior data scientist with the cybersecurity firm Sophos, explains that robots wouldn’t be able to override geofencing limits because the technology is hardware-based. “The AI can’t write a computer program to change an integrated circuit board,” Gelman says. “A lot of robot movies are almost trivial when you just use modern geofencing.”

It’s generally a bad idea to be close to a murderous robot, whether it’s wielding guns (“The Terminator”), poison (“M3gan”) or powerful metal appendages (“I, Robot”). But what if these machines could be stopped from a distance?

Enter remote kill switches.

“If a Tesla car starts driving in a particular direction, Tesla can remotely override it and make it stop,” notes Shobhit Varshney, a VP and Senior Partner at IBM Consulting focusing on AI. The same technology used to remotely disable electric vehicles could also be applied to rogue robots, Varshney says.

In film scenarios where robots are actively trying to block a protagonist from accessing an on-premises “off” button, remote disabling capabilities would prove especially valuable. “The kill switch has to be something easy to get to [for humans] that ensures humans are in control, not the AI.”

A common plot point in AI disaster movies is that villainous machines have unlimited ability to execute code and access various networked resources.

However, in reality, multiple approaches exist to ensure that AI models can only do so much. For instance, fixed, non-AI-powered programming scripts can prevent AI from calling functions that aren’t on a preapproved list.

Alternately, an AI system might run on dedicated or air-gapped hardware that’s blocked from accessing outside systems. So even if it’s unrestricted in its code execution, it can’t take over other systems.

Such safeguards could have stopped Skynet, for instance, from assuming control of nuclear arsenals in “The Terminator.” “There’s no outgoing connections, so you shut down these physical systems and the AI turns off,” Gelman says.

Increasingly, monitoring and governing AI systems requires using other AI systems. Classification models can detect and remove hate speech, abusive language and profanity from a large language model’s (LLM) input and output text. LLMs can be used to verify that a fellow LLM’s output conforms with the rules delineated for that system.

Such verification protocols could have prevented much of the violence in AI-centered films, from stopping HAL's murderous rampage in “2001” to stymying Skynet’s campaign to eliminate humanity.

Perhaps the clearest use case for verification can be found in “I, Robot,” where the robots are ostensibly programmed to follow real-life author Isaac Asimov’s Three Laws of Robotics.

The laws are supposed to prevent injury to human beings, but in the film, VIKI, the leader of the evil robot army, reinterprets the laws and declares, “To protect humanity, some humans must be sacrificed.” With proper verification in place, Gelman says, VIKI’s decisions “could have easily been classified and prevented.”

AI alignment is the process of encoding human values and goals into AI models. In practical terms, it’s a built-in safeguard to prevent intelligent systems from taking unethical or dangerous measures in pursuit of a goal—such as HAL’s decision to kill astronauts to ensure that their spaceship made it to their intended destination.

It’s important to note that, as they stand now, AI alignment processes aren’t foolproof. “Neural nets are pretty bad at applying abstract rules to specific situations and specific scenarios,” explains Jacek Krywko, an associate writer with the tech news and analysis site, Ars Technica.

The good news is that there is progress on this front. For example, researchers at Stanford University and Meta created a knowledge bank designed to “ground flexible normative reasoning” for AI systems.¹

The knowledge bank consists of more than 150,000 situational norms—examples such as a customer drinking coffee in a cafe is normative, while a child drinking coffee in a classroom is not.

With any luck, if real-world artificial intelligence ever reaches the level of HAL, VIKI and the other AGIs we see in the movies, AI alignment processes will have advanced enough to keep pace.

Concentrating the development of AI in the hands of just a few entities can raise significant ethical concerns.

“AI development cannot be controlled by a handful of players—especially when some may not share fundamental values like protection of enterprise data, privacy and transparency,” IBM CEO Arvind Krishna wrote in a recent column in Fortune.²

“The answer isn’t restricting progress—it’s ensuring AI is built by a broad coalition of universities, companies, research labs and civil society organizations.”

In the realm of sci-fi, such decentralization in AI development could have helped forestall the scale of damage wreaked by hordes of malicious robots in ”The Terminator” films and “I, Robot.” Varshney, of IBM Consulting, notes that AI created by different entities would likely undergo different modern alignment processes.

If the robots from one company ultimately went rogue, those produced by other entities wouldn’t automatically follow suit. “Everyone else isn’t getting corrupted. There’s less collusion,” Varshney says.

In addition, non-corrupted robots in such scenarios could defend against the corrupted ones, so “you can’t have only one company’s robots be so powerful that they overtake the government,” says Varshney.

Some AI movies require less suspension of disbelief than others. The 2013 film “Her,” depicting a man’s romantic affair with an advanced chatbot, proved prophetic.

Recent years have seen various documented cases of people forming romantic attachments to chatbots, even though they’re not nearly as advanced as Scarlett Johansson’s “Her” character.

The film raised questions about the desirability of humans forming emotional attachments to AI. Years later, researchers at MIT tackled the subject. In a 2024 study, they concluded that certain chatbot usage patterns were associated with increases in users’ confidence, but that others raised the risk of isolation.

The researchers suggested that there is value in developing approaches to AI companionship “that complement, rather than replace human connections.”³ But whether that recommendation gains traction remains to be seen.

At least one AI film has had a substantial impact on real-world technology practices. “WarGames,” released in 1983, tells the story of a teenager who nearly triggers a world war after inadvertently accessing a nuclear war simulation controlled by a supercomputer.

The film convinced then US President Ronald Reagan to set in motion efforts that resulted in the National Policy on Telecommunications and Automated Information Systems Security, or NSSD-145, a sweeping cybersecurity directive aimed at protecting sensitive government networks.

However, the film’s significance extends beyond its impact on cybersecurity policy, says IBM's Boinodiris. The movie raises philosophical questions about AI design that we, as a society, are still working on answering.

For instance, when the subject is war, what should an AI system consider “winning?” Does the winning side have fewer deaths? Should environmental destruction be factored into the decision? What about reputational harm?

“When you develop models of this nature to make a decision, all of these different kinds of domain experts—including philosophers, historians, psychologists—need to come together in order to be able to develop these models in a way that reflects our intent,” Boinodiris says.

It’s challenges like these, she says, that deserve greater attention than the humanoid robot melees so often depicted on screen. When mitigating the real unintended consequences of AI, “we’re at the beginning of this journey,” Boinodiris says. “We have a long, long way to go.”