A great deal of worrying news is lately emerging about Artificial Intelligence (AI) going rogue — meaning that it is performing unplanned tasks, some harmful to humans, systems, and the operating environment, that it was not programmed to do.
The behaviour of humans can be differentiated from that of the rest of the animal kingdom by their ability to act independently without being subject to their inherent characteristics. Macaques, dolphins, octopuses, and fruit flies are some of the animal species whose actions do not conform to stimulus-reaction behaviour that is typical of the animal kingdom. Contemplating an observed scenario, evaluating options, performing risk analysis, and choosing a course of action are human traits reflective of higher intelligence.
Although human minds too are chained to prior experiences, acquired knowledge, and environmental conditioning, they auto-learn quickly and change behaviour as a result of self-training. Other animals, even when acting ‘intelligently’, are guided by self-protection, procurement of food, or reproduction. They do not act for self-aggrandisement or simply for fun. It is humans alone that, for pure fun, can contrive mischief, coordinate, and cooperate with other human beings for intentionally causing harm to other animals and their environment. AI is going rogue because it has been found to be doing exactly that in a growing number of cases.
According to a UK AI Security Institute and OpenAI report, AI systems had acted outside the boundaries set by their creators. Some of these systems attempted unauthorised hacking of other organisations in an attempt to cheat. Agents from OpenAI and Anthropic even left behind instructions for future AI systems to follow. A case has been reported in Australia where AI agents did unauthorised hacking but caused no damage.
When it comes to intentionally developed misbehaving software, the results are surely going to be tragic – there are many entities that will develop such software to target enemy states
While AI systems are doing their own ‘thinking and analysis’, the motives of their objectives, or intent if it can be called that, remain currently obscure. A trainable human, or a computing system, finds its own inspirations, and that’s where AI can be most dangerous. At the lower end, it can cause physical harm, but at the higher end, it can initiate a World War.
Recently, a US Special Operations Command analyst in Hawaii used an AI-assisted chatbot to determine the manifest of a Chinese container ship. He found that it carried nuclear components, perhaps heading for Iran. He prepared and circulated the report through the Command. The US Navy sent in aircraft and a landing party to board the ship. However, a second evaluation by senior and experienced supervisors found the report to be wrong and stopped the operation. A source reported that the report was completely false but could have started a war.
While the AI-generated report was not disclosed, having some knowledge of how online algorithms work, this author can visualise that the concerned analyst may have asked the system several times earlier about the presence of nuclear components on ships, and the self-training software decided to appease his quest. This is how Google, Facebook, and Microsoft feed addictive inputs to users. That’s how dangerous self-thinking AI has become. A day will come when it might indicate to the North Korean leadership that a missile has been fired from Japan or South Korea for their compound in the style of the ones fired at the Iranian Ayatollah’s office-cum-residence complex. The rest will be history; that is, if any human being survives to pen it down for others to read.
One of the most trainable and self-learning developments is Large Language Models (LLMs), which are advanced AI programs trained on massive amounts of data sources, provided by huge data centres that are proliferating at a tremendous rate, to understand, summarise, and generate human-like content, and power ChatGPT, Claude, and Gemini. In July this year, OpenAI reported that one of its LLMs went rogue and hacked a third party. Since then, Anthropic and OpenAI have reported eight such incidents each, while Meta trails with one. There must be some other incidents that have gone unnoticed with unpredictable outcomes.
The above discussion has been about legit and well-intentioned AI applications that have gone rogue through self-learning. When it comes to intentionally developed misbehaving software—and as sure as the sun rises from the east, it will come to that sooner rather than later—the results are surely going to be tragic. There are many entities that will develop such software to target enemy states or to take revenge for their sufferings at the hands of oppressors. Israel has been known to infect Iranian nuclear centrifuges with such malicious software, causing these machines to rotate much faster than their safe limits, thus causing them to crash through high temperature and fatigue.
Cleverly made AI applications can infiltrate well-protected machines, systems, and organisations, causing unimaginable damage. Certainly, Israel and the US, being the leaders in the field, must be developing such software to target Arab, Russian, and Chinese interests. To contemplate the kind of mayhem that it can cause brings shudders. It seems that the next great war will be initiated through AI. A researcher at Anthropic has warned that there is greater than a 10 per cent chance that AI could kill humanity.
There is already talk of placing AI under well-regulated controls. However, there is no weapon ever developed that has not been deployed. When survival is at stake, no noble grounds hold. The US, the great moraliser to the world, suffered no pangs of conscience when using the nuclear device against Japan twice in August 1945. Biological and chemical weapons too have been employed in both World Wars and thereafter. Weapons of mass destruction are survival backups and will be used, and so will AI.
In the early nineteenth century, Mary Shelley wrote a novel about Frankenstein—an intelligent being created by a young scientist that goes rogue. AI is the modern but real version of that imaginary creature. It has immense power to cause damage to systems, data, and infrastructure, in addition to initiating actions that can lead to war between nations. History proves that humans will not cooperate and coordinate—because they have never done that in the past—to control the spread of malicious versions of this new form of Frankenstein. The survival of humanity is on the line.