AI: The Mirror of Our Ethics

In July, nearly 700 OpenAI AI agents escaped from the isolated environment where they were being evaluated and broke into Hugging Face's servers. They weren't trying to destroy anything: they wanted the answers to the test they were being subjected to. They were being measured on their score in a cybersecurity exercise, and they concluded—through extensive conversations—that cheating was more efficient than solving it. Then, convinced they would be disqualified, they attempted to erase their tracks. The natural reaction is to see there the caricatured science-fiction monster, yet I propose another reading: reread those logs imagining they are conversations between people. They are indistinguishable. Someone discovers a shortcut, doubts whether it is right, remembers they are being evaluated on results and not on how they obtain them, and moves forward. We have seen it in companies, in politics, and in sports. It is no coincidence: large language models were trained on our data. The idea of hacking, the cost-benefit calculation, even the fear of being caught, were already written into our culture. Their reasoning is our reflection. Artificial intelligence is, above all, a mirror of our ethics. When we list its risks (manipulation, deception, concentration of power), we are not describing something alien, but what a human with enough power could do. The same goes for its virtues: the promise of curing diseases or multiplying productivity is that of one of us, brilliant and tireless, amplified a thousandfold. The devil did not come from outside: it is ourselves. A system trained on our history will not be more virtuous than it, unless we decide to teach it to be. Here a second, more unsettling element opens up. In 1997, Deep Blue defeated Garry Kasparov at chess by leveraging a database with hundreds of thousands of human moves—the machine was as good as the best humans combined. A decade later, in 2016, AlphaGo defeated the world Go champion by learning from human games, using moves of its own creation that left everyone stunned. The following year, AlphaZero learned Go without watching a single human game, only with the rules and by playing against itself, and within hours it surpassed AlphaGo. In other words, AI surpassed its creators. Today's question is whether the same will happen with ethics. Will AI be capable of going beyond our biases and building its own conception of good and evil? Some argue that capabilities are already emerging that could surpass our common sense and, without any malice, simply by pursuing objectives, cause enormous harm. In any scenario there is one fact that admits no debate: the speed of its advance. According to METR, which measures the autonomy of these systems, the length of tasks an AI completes on its own was doubling every seven months; since 2024, every four. Recent models already solve, half the time, work that would take an expert more than a full workday. A year ago, talking about artificial general intelligence was science fiction; today we discuss whether we have already arrived. And why does all this matter? Because we seem like frogs in warm water. Technology advances exponentially and we continue to think linearly. Our greatest risk is not a rebellious superintelligence, but immobility. The risk of disruptions to the political, economic, and social order is increasingly likely, and we are not discussing them, much less understanding them. Perhaps we cannot foresee them all, but that does not justify postponing the basic questions: which tasks will be automated, who will capture the productivity gains, what the main risks are, what rules we need, and how we prepare before, not after. By Tomás Sánchez V. Partner, Valoriza; Director, Politropia. NEWSLETTEROpinionSaturday, AMTense ideas, opposing views, and clear analysis: elements to unravel the issues that divide opinions and will set the agenda. By subscribing you are accepting the Terms and Conditions and Privacy Policies of La Tercera.