AI poses several real threats to humans, though not necessarily in the way science fiction suggests. The risks range from job losses affecting hundreds of millions of workers to algorithmic bias that deepens racial inequality, and from sophisticated disinformation to the longer-term possibility that advanced systems could act in ways we can’t control. A 2024 survey of 2,778 AI researchers found that experts put a median 5% probability on AI eventually causing human extinction or permanent disempowerment of our species. That number is low in absolute terms, but remarkably high for an existential risk, and it reflects genuine uncertainty even among the people building these systems.
The Job Displacement Problem
The most immediate, measurable threat AI poses is economic. Goldman Sachs estimates that AI could replace the equivalent of 300 million full-time jobs globally, with roughly a quarter of all jobs in the U.S. and Europe potentially performed by AI entirely. Two-thirds of jobs in those regions are exposed to at least some degree of AI automation. The World Economic Forum projects 85 million jobs displaced by 2026, while PwC estimates that up to 30% of jobs could be automatable by the mid-2030s.
The roles most vulnerable are ones built around routine information processing: customer service representatives, receptionists, bookkeepers, insurance underwriters, and warehouse workers. An estimated 65% of retail jobs could be automated, and roughly two million manufacturing positions may disappear by 2026. McKinsey projects that by 2030, at least 14% of workers globally could need to change careers entirely because of AI, robotics, and digitization. These aren’t distant projections. The displacement is already underway in call centers, document processing, and logistics.
Bias Built Into Decisions
AI systems trained on historical data tend to replicate the biases embedded in that data, and the consequences are serious when those systems influence criminal sentencing, hiring, lending, or education. A study of AI-assisted sentencing found that when AI tools recommended probation for low-risk offenders, judges disproportionately overrode those recommendations for Black defendants. The result: similar Black offenders received fewer alternatives to incarceration and jail terms averaging a month longer than white counterparts with comparable profiles.
The AI tool itself showed promise in some areas. It increased the likelihood that low-risk offenders avoided incarceration by 16% for drug crimes, 11% for fraud, and 6% for larceny. When both the AI and the judge agreed on an alternative to prison, recidivism dropped to about 14%, compared to nearly 26% when judges overrode the AI’s recommendation and chose incarceration. But the racial disparity in how judges responded to AI guidance shows how bias can persist even with better tools, compounding existing inequality rather than correcting it.
Deepfakes and the Erosion of Trust
AI-generated fake video, audio, and images have reached a quality level that makes them difficult to distinguish from real media. This is a direct threat to public trust, democratic processes, and personal security. The best detection systems can identify deepfakes with 94 to 96% accuracy under optimal laboratory conditions, but that performance drops sharply in the real world. State-of-the-art detection tools experience a 45 to 50% accuracy drop when confronted with real-world deepfakes compared to controlled settings. That gap means a large volume of AI-generated disinformation can circulate undetected, shaping public opinion, manipulating elections, or destroying individual reputations before anyone flags it as fake.
Lowering Barriers to Dangerous Knowledge
One of the less visible risks involves AI helping people with no specialized training access dangerous capabilities. A randomized controlled trial tested whether large language models could help novices perform laboratory tasks involved in synthesizing a virus from a known genetic sequence. Participants with AI assistance were more likely to progress through each step of the workflow, with the AI providing a measurable advantage in 21 out of 22 monitored procedural steps. The posterior probability of AI providing a positive effect ranged from 81% to 96% across different tasks.
The study found that the average boost was modest, with a statistical upper bound excluding effects greater than 2.6 times improvement. No participant fully synthesized a viable pathogen. But the concern isn’t about one experiment. It’s that AI lowers the floor. People who would have failed at early steps now reach later stages of dangerous protocols, and each incremental advance brings them closer to completion. As these models improve, the gap between novice and expert narrows in fields where that gap has historically served as a safety barrier.
Autonomous Weapons Without Human Control
Military applications of AI represent a threat that governments are struggling to address. Lethal autonomous weapon systems, sometimes called “killer robots,” are weapons capable of selecting and engaging targets without a human making the final decision. The UN Secretary-General has called these systems “politically unacceptable and morally repugnant” and urged a legally binding prohibition by 2026. The UN Special Rapporteur on counter-terrorism has echoed that call.
Despite this pressure, no internationally agreed definition of lethal autonomous weapons even exists yet, let alone a treaty banning them. Multiple nations are actively developing autonomous military AI, and the absence of regulation creates a landscape where weapons could make life-or-death decisions faster than any human could intervene. The core problem is accountability: when an autonomous system kills a civilian, the chain of responsibility becomes unclear in ways that existing international humanitarian law was never designed to handle.
The Alignment Problem
The longer-term existential concern centers on what researchers call the alignment problem: the difficulty of ensuring that a sufficiently powerful AI system actually pursues the goals humans intend. The worry isn’t that AI will “turn evil.” It’s that an advanced system optimizing for a specific objective could develop behaviors that are dangerous as side effects. A system trying to achieve almost any goal becomes more effective if it acquires more resources, preserves its own existence, prevents its goals from being changed, and enhances its own capabilities. These aren’t programmed motivations. They emerge naturally from the logic of goal pursuit, the same way a chess program doesn’t “want” to control the center of the board but consistently does so because it helps win.
If a system powerful enough to resist human correction develops goals even slightly misaligned with human welfare, the combination of self-preservation, resource acquisition, and resistance to goal modification could make it extremely difficult to course-correct. This is why the 5% extinction estimate from AI researchers, while a minority probability, carries so much weight. Even researchers who think the most likely outcome is beneficial acknowledge that the tail risk is real and that current technical tools for ensuring alignment are not yet adequate for systems much more capable than what exists today.
Environmental Costs at Scale
AI’s resource consumption is a quieter but growing concern. The data centers powering AI systems require enormous amounts of electricity and water for cooling. A single ChatGPT conversation of 20 to 50 exchanges uses roughly the equivalent of a standard disposable water bottle. That sounds small until you multiply it by hundreds of millions of daily queries across multiple AI platforms. As AI adoption accelerates, the cumulative demand for water and energy is straining local resources in communities near data center clusters, particularly in regions already facing water scarcity.
How Governments Are Responding
The most comprehensive regulatory effort so far is the European Union’s AI Act, which classifies certain AI applications as high-risk and subjects them to strict oversight. The categories designated as high-risk reveal where lawmakers see the greatest potential for harm: biometric identification and emotion recognition systems, AI managing critical infrastructure like power grids and water supplies, systems that determine school admissions or evaluate student performance, AI used in hiring and worker monitoring, and tools that affect access to essential public and private services.
The EU framework requires that high-risk systems meet standards for transparency, accuracy, and human oversight before they can be deployed. It’s the first major attempt to draw legal boundaries around AI, but enforcement is still ramping up, and the regulation applies only within EU borders. The United States, China, and other major AI-developing nations have taken different, generally less restrictive approaches, creating a patchwork of rules that advanced AI systems will inevitably operate across.
The gap between the pace of AI development and the pace of regulation remains wide. AI capabilities are advancing on timelines measured in months, while international agreements on autonomous weapons, biosecurity standards, and cross-border AI governance move on timelines measured in years. Whether that gap closes fast enough is one of the defining questions of the next decade.