LIM WOONG In the early innings of a technological shift, the public narrative is almost always fear and “pushback.” We fret over video games that swallow adolescents’ interest in books, internet pornography that "trains" the user’s brain into compulsion, and teenagers trapped in the dopamine loops of social media. Layered atop this is the geopolitical sludge of propaganda, surveillance and industrial-scale disinformation, often orchestrated by state actors like China and Russia.Soon enough, the counter-narrative arrives. In panel discussions and at dinner parties, we hear the canned lines that tools are neutral and that only the user determines whether they become instruments of good or ill. A car can commute or it can kill; a kitchen knife can prepare a meal or it can puncture a lung. The proposed remedies are familiar: discretion, regulation, shared governance and “tech literacy” that teaches critical thinking and appropriate uses of technology.This framing feels failsafe, but in an era of AI everywhere, it is far from complete. When an AI system is designed and deployed, we must ask whether its use predictably produces damage on a scale that overwhelms its legitimate benefits. Answering this requires more than saying people should know better. It requires ethical sorting — dividing AI into applications that are relatively good by design, inherently bad, and “it depends.”To speak of “relatively good” artificial intelligence is not to claim moral purity. The goal is systems designed to help and protect us — and when they fail, leave a clear trail showing what went wrong. Image recognition can flag an unattended bag in a terminal or help read X-rays where doctors are scarce. When emergency hotlines are understaffed, an AI agent that asks the right questions can backstop human fatigue without replacing judgment. Customer service chatbots can reduce friction for routine tasks, so long as there is a direct path to a human. In clinical settings, AI expert systems can act as high-speed checklists, catching mistakes with dangerous drug combinations and warning labels. Add real-time captioning and satellite mapping for disaster relief. The unifying thread is not perfection, but net-helpful defaults and failures that can be audited and corrected.To label an AI system “inherently bad” is a strong indictment, reserved for technologies whose primary value is deception, exploitation, or the erosion of trust. Voice cloning may begin as a prank, but it collapses the credibility of our simplest authentication method: recognizing a human voice. This becomes a scalable attack on the elderly and on institutions that still rely on phone-based trust. Likewise, image generators that fabricate synthetic material can turn disputes into evidentiary chaos; even when a forgery is debunked, the victim pays the immediate cost. Deepfake pornography industrializes sexual humiliation. Timed to elections or crises, video deepfakes create civic damage: citizens stop believing real footage while clinging to fakes that confirm their biases. Add spearphishing copilots, stalkerware, disinformation farms, and biometric “threat scoring” for surveillance.Between good and bad lies a massive grey area: Tools for translation, search, writing, and coding. The threat here isn’t the technology. It’s the rush to cash in. Companies push these systems to the public long before society builds the guardrails — standards, accountability, and rules. In education, for example, generative AI that produces polished essays for assessments hollows out the cognitive struggle (argument structure, synthesis, revision) that schools aim to teach, offering counterfeit merit in its place. Parents pay, companies profit, students lose the chance to build foundational skills and agency, and schools burn time policing outputs instead of teaching what matters.This is where a utilitarian strain of ethics, most notably associated with Peter Singer, both philosopher and activist, becomes practical. Singer’s core idea is not that emotions are irrelevant, but that moral seriousness requires counting interests impartially. If an action predictably reduces suffering or increases well-being for many without imposing devastating costs on a vulnerable few, it is worth endorsing. Applied to AI, this forces a sober accounting of public welfare: How many people are affected, and how intensely? Is harm an edge case or the default of the most profitable use? Who gets convenience, and who gets the fraud? Can the damage to our shared ability to agree on what is true be repaired? We must also weigh opportunity cost: What medical devices or essential home or industrial appliances could we have built with the same talent and computation instead poured into “deepfake” technologies? Culture complicates the accounting, but Singer’s baseline holds across borders: Avoidable suffering is a moral failure. We cannot ignore societal damage under the banner of progress.The task before us is neither panic nor blanket acceptance, but moral triage. If an AI system reliably increases suffering or corrodes baseline trust, we must stop treating it as neutral. We must name it as a design choice that fails the test of goodness and, without melodrama but with firmness, refuse to fund and scale it.- - -Lim WoongLim Woong is a professor at the Graduate School of Education at Yonsei University in Seoul. The views expressed here are the writer’s own. — Ed.