Will Artificial Intelligence Do More Harm Than Good? Weighing AI’s Social Effects

Artificial intelligence can help interpret medical images, convert speech into text, translate between languages, and handle repetitive office tasks. Those possibilities matter: a tool that makes information easier to access or gives workers more time for complex decisions can improve daily life. But AI can also produce convincing falsehoods, repeat biases in its data, or make errors at a speed and scale that people struggle to detect. The central question is not whether AI is simply good or bad; it is whether society can direct its uses toward public benefit without allowing avoidable risks to become routine. Advanced AI is more likely to benefit society than to cause serious harm—but only if high-stakes systems are tested, monitored, and kept under meaningful human responsibility. Without those safeguards, the same speed and reach that make AI useful can magnify its failures.

The strongest case for AI is not that machines can replace human judgment, but that they can extend people’s abilities. A medical tool might flag an image for a clinician to examine, helping direct attention to a possible concern; it should not be treated as a diagnosis on its own. Speech recognition and text-to-speech tools can help people communicate or access material in formats that suit them. In workplaces, software that drafts routine summaries or sorts large collections of information can leave employees more time for tasks requiring empathy, accountability, and contextual judgment. In each example, the benefit depends on how the tool is used: AI assists a responsible person rather than quietly becoming the decision-maker. This distinction supports the thesis because it identifies a route to real gains without assuming that a model’s output is automatically correct.

However, faster work is not automatically better work, and the gains may be uneven. If an employer uses automation to reduce staffing rather than to support employees, workers may lose income and bargaining power even as the organization becomes more efficient. New roles may emerge, but that possibility does not erase the immediate costs for people whose tasks are displaced or who lack time and resources to retrain. A fair transition therefore requires choices beyond the software itself: employers and governments can provide training, involve affected workers in decisions, and assess whether productivity gains are shared. This complication does not disprove AI’s potential; it shows why benefits should be judged across society, not just by whether a task takes fewer minutes.

The risks become more serious when AI influences decisions about health, employment, public services, or essential infrastructure. A system trained on incomplete or skewed examples could perform less reliably for some groups, while users who do not understand its limits might treat a confident answer as fact. Generative tools can also make fabricated text, images, or audio easier to produce and circulate, weakening people’s ability to judge what is genuine. A single error in a low-stakes draft may be easy to correct; the same error embedded in a consequential decision can deny someone an opportunity or delay needed help. The danger is therefore not intelligence in the abstract. It is the combination of flawed outputs, consequential uses, and institutions that fail to check them.

The National Institute of Standards and Technology offers one practical response to this problem. NIST developed its AI Risk Management Framework through collaboration between its Information Technology Laboratory and private- and public-sector participants. The framework is voluntary and intended to incorporate trustworthiness considerations throughout an AI system’s lifecycle. That approach matters because risk cannot be managed only after a product is released: developers and users need to examine a system’s purpose, possible failures, and effects as it is designed, deployed, and evaluated. NIST released its first framework in 2023 after a consensus-driven process that included public input, and later published a profile focused on risks distinctive to generative AI. These steps do not prove that AI is safe or guarantee compliance. They do show how risk management can become a continuing practice rather than a promise made at launch.

A serious objection is that voluntary standards cannot stop every company from cutting corners, and some harms may be difficult to reverse once an AI-generated falsehood spreads or an automated decision affects someone’s life. This objection is persuasive: guidance alone is not enough for every high-risk use. My argument therefore depends on safeguards with consequences. Organizations should test systems before deployment, check performance across relevant groups, disclose important limitations, and provide a clear way for people to appeal consequential decisions. Independent audits and enforceable rules may be needed where failure could threaten rights or safety. Human review must also be real, not a person mechanically approving every recommendation. These protections cannot eliminate error, but they make it more likely that errors will be found, challenged, and corrected.

There is a useful caution in President Dwight D. Eisenhower’s 1961 farewell address. Warning about the influence of the military establishment and arms industry, he urged citizens to remain alert so that powerful institutions would not erode democratic processes or displace independent judgment. AI is not the same as the defense complex he discussed, and the comparison should not be stretched into a claim that the technologies or institutions are identical. The relevant lesson is narrower: powerful systems require informed public scrutiny, especially when their benefits and risks are distributed unevenly. Applied to AI, that means citizens need understandable information, researchers need room to examine impacts, and decision-makers must be answerable for how systems are used. AI is more likely to do good than harm when innovation is matched by oversight and shared responsibility. Without them, its advantages could become concentrated while its mistakes are borne by others.