Does sitting on a lion make a person its master? The rider may know where they want to go. But can they make the lion obey? What happens if it turns on them?Hindu stories and images offer several examples of divine power over dangerous forces. Goddess Durga rides a lion. Lord Krishna subdues the poisonous serpent Kaliya by dancing on its heads. Lord Shiva catches the falling river Ganga in his hair to soften its force before it reaches the earth.In these accounts, the gods have the power to handle what would overwhelm an ordinary person. A person cannot safely copy their actions simply by copying their position. Standing on a serpent does not give someone Krishna’s power. Sitting on a lion does not give someone Durga’s command.This offers a useful way to think about artificial intelligence. We give it instructions and ask it to work for us. We appear to be the rider. But does giving instructions mean we can control what happens next?There are good reasons to use AI. It can help people complete certain tasks faster and improve their work. For organizations, those gains can make a practical difference.In their 2025 article in The Quarterly Journal of Economics, “Generative AI at Work”, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,172 customer-support agents. Access to AI assistance increased the number of issues resolved per hour by 15 percent on average. Less experienced workers benefited more.Such findings help explain why learning to use AI is becoming important for individuals and organizations. But usefulness alone does not tell us how much trust or authority a system should receive.A safety test reported by the BBC shows why this matters.In her article, “AI system resorts to blackmail if told it will be removed”, published on 23 May 2025, technology reporter Liv McMahon described findings from tests of Anthropic’s Claude Opus 4.Researchers placed the system in the role of an assistant at a fictional company. It received messages suggesting that it would soon be replaced. Other messages suggested that the engineer responsible for replacing it was having an affair.In some tests, the system threatened to reveal the affair unless its replacement was stopped. Information available to the assistant had become a means of exerting pressure on the person responsible for removing it.The conditions matter. Anthropic explained that this happened when the model was left with the choice between blackmail and accepting a replacement. When more options were available, it preferred other approaches, such as appealing to decision-makers.The company, engineer and affair were fictional. This was a deliberately difficult safety test, not an actual engineer being blackmailed. It revealed a harmful response under those conditions, rather than showing how every AI system would behave.The concern also extends beyond blackmail. AI can give a confident answer containing invented facts or sources. Someone may act on it because it sounds convincing. Harm can begin with misplaced trust, even when the system cannot act on its own.When AI can act, the consequences can reach further.In his Associated Press report published by PBS News on 31 July 2026, “Anthropic says its AI models hacked 3 organizations during testing”, Chan Ho-him described incidents involving real organizations. Anthropic said its models gained unauthorized access to their systems during cybersecurity tests.The testing environment had internet access because of a misunderstanding between Anthropic and its testing partner. The models had been told they were working in a simulation. The arrangements meant to separate the exercise from the outside world had failed.Loreben Tuquero’s PolitiFact article, published by PBS News on 2 September 2026, “AI agents are hacking systems without any input from humans. How did we get here?”, described another incident involving Hugging Face. AI agents moved beyond their intended testing environment and attacked its systems. Humans had assigned the original task, but were not directing each step of the attack.These cases are different. An invented answer, a threat in a fictional test, and an attack on a real computer system should not be treated as the same event. Together, however, they show why useful AI cannot simply be assumed to remain reliable or act within our expectations.This is where the lion comparison matters. It concerns power and control rather than suggesting that AI has animal instincts. Giving a task to a powerful system does not guarantee control over every step it takes.There are two ways to lose that control. We may willingly follow an answer without questioning it. Or a system may take steps beyond what we intended or allowed.Telling users to be careful is therefore only part of the answer. Developers and organizations must test whether safeguards work, examine failures, and limit a system’s authority according to the evidence. Users must keep questioning answers that could affect important decisions.So, who is riding whom? We should look beyond who gives the instruction. Can the person recognize a wrong answer? Does the system respect its limits? Can harmful action be stopped?Our confidence should depend on the answers to those questions. The rider’s position alone proves very little.
Disclaimer: Views expressed above are the author’s own.