Unreliable systems at consequential scale
A generative model can produce false information with the confidence and polish of a correct answer. In casual use, that may be inconvenient. In medicine, law, finance, infrastructure, or public administration, it can harm lives. Automation magnifies the problem because one flawed system can influence thousands of decisions before anyone notices.
The remedy is not a warning label alone. High-stakes systems need task-specific evidence, calibrated uncertainty, independent validation, human review, monitoring for drift, and a process for appeal and correction. Some uses should remain prohibited when errors cannot be made acceptably rare or safely contained.
Bias, surveillance, and unequal power
AI learns from societies that already contain discrimination. Historical records can encode unequal policing, hiring, lending, health access, and representation. A model can reproduce those patterns while appearing neutral because its decisions are expressed mathematically.
At the same time, facial recognition, predictive systems, and large-scale data collection can expand surveillance. UNESCO’s global ethics framework centers human rights, fairness, privacy, transparency, and human oversight because technical accuracy alone does not make a use legitimate.
Deception and information integrity
Synthetic text, voice, images, and video reduce the cost of impersonation, fraud, harassment, and propaganda. The deepest danger is not that every fake will be believed, but that people may stop believing authentic evidence. Trust can erode when verification becomes expensive and denial becomes easy.
Response requires provenance standards, authentication, platform enforcement, media literacy, rapid incident reporting, and legal accountability for harmful uses. Detection will help but cannot carry the entire burden because generation and detection evolve together.
Cybersecurity and dangerous capability
AI can help defenders identify vulnerabilities, analyze security information, and respond faster. It can also help malicious actors scale scams, impersonation, reconnaissance, and other attacks. Connecting an AI system to external software can increase the consequences of mistakes or manipulation.
These risks are part of the wider field of cybersecurity. Organizations using AI remain responsible for protecting information, testing their services, and following established security standards appropriate to their work.
Labor disruption and concentration
Even when AI transforms more jobs than it eliminates, transitions can still be painful. Tasks may disappear faster than workers can retrain. Monitoring and algorithmic management can reduce autonomy. Productivity gains may flow primarily to owners of models, data, platforms, and computing infrastructure.
Public policy and worker voice matter. Competition, education, bargaining power, portable benefits, transition support, and access to productivity-enhancing tools can influence whether AI broadens prosperity or deepens inequality.
Environmental and systemic risk
Large-scale AI requires electricity, cooling, hardware, minerals, construction, and—in some facilities—significant water use. The IEA’s 2025 base case projects global data-center electricity consumption of about 945 terawatt-hours in 2030, just under 3% of projected worldwide electricity consumption. The same report presents substantially different sensitivity cases, so this number is a forecast rather than a certainty. Local grids and communities can face concentrated pressure even when the global share is limited.
Researchers also debate lower-probability risks that could have wide consequences if increasingly capable systems were used across finance, infrastructure, communications, or other essential services. These possibilities should be discussed without presenting uncertain outcomes as established facts.
Primary research and institutions
Sources are linked directly so readers can examine the underlying evidence. Numerical statements identify the reporting organization and year. Projections are presented as estimates, not established future facts.