Last spring, Daniel Kokotajlo, an AI safety researcher at OpenAI, quit in protest, convinced the company was unprepared for the future of AI technology and wanting to raise alarm. In a phone conversation, he appeared affable but anxious, explaining that progress in AI "alignment"—the methods ensuring AI follows human values—was lagging behind intelligence advances. He warned researchers were rushing toward creating powerful systems beyond control. Kokotajlo, who transitioned from philosophy graduate studies to AI, had self-taught to track AI progress and predict when critical intelligence milestones might occur. After AI advanced faster than expected, he adjusted his timelines by decades. His 2021 scenario, “What 2026 Looks Like, ” saw many predictions realized early, leading him to foresee a 2027 or sooner “point of no return” where AI could surpass humans at most vital tasks and wield great power. He sounded fearful. Simultaneously, Princeton computer scientists Sayash Kapoor and Arvind Narayanan prepared their book “AI Snake Oil, ” which took a sharply contrasting view. They argued AI timelines were overly optimistic; claims of AI usefulness were often exaggerated or fraudulent; and the complexity of the real world meant AI’s transformative effects would be slow. Citing examples of AI errors in medicine and hiring, they stressed even the latest systems suffer from a fundamental disconnect from reality. Recently, all three sharpened their views in new reports. Kokotajlo’s nonprofit, AI Futures Project, published “AI 2027, ” a detailed, heavily cited report outlining a chilling scenario where superintelligent AI might dominate or exterminate humanity by 2030—a serious warning. Meanwhile, Kapoor and Narayanan’s paper “AI as Normal Technology” maintains that practical barriers—from regulations and safety standards to real-world physical constraints—will slow AI deployment and limit its revolutionary impact. They argue AI will remain “normal” technology, manageable via familiar safety measures like kill switches and human oversight, likening AI more to nuclear power than nuclear weapons. So, which will it be: normal business or apocalyptic upheaval?The starkly divergent conclusions—drawn by highly knowledgeable experts—from these reports produce a paradox akin to debating spirituality with both Richard Dawkins and the Pope. The difficulty stems partly from AI’s novelty—like blind men examining different parts of an elephant—and partly from deep differences in worldviews. Generally, West Coast tech thinkers envision rapid transformation; East Coast academics lean toward skepticism. AI researchers favor swift experimental progress; other computer scientists seek theoretical rigor. Industry insiders want to make history; outsiders reject tech hype. Political, human, and philosophical views about technology, progress, and mind deepen the divide. This captivating debate is itself a problem. Industry insiders largely accept “AI 2027”’s premises while quarreling over timelines—an inadequate response akin to quibbling over timing as a planet-killer approaches. Conversely, the moderate views in “AI as Normal Technology” about keeping humans in the loop are so understated they’ve been ignored by doom-focused analysts. As AI grows societally critical, discourse must evolve from specialist debate to actionable consensus. The absence of unified expert advice makes it easier for decision-makers to ignore risks. Currently, AI companies haven’t significantly shifted the balance between capabilities and safety.

Meanwhile, new legislation prohibits state regulation of AI models and automated decision systems for ten years—potentially allowing AI to regulate humanity if the dire scenario proves correct. Addressing safety now is urgent. Predicting AI’s future narratively involves trade-offs: cautious scenarios may overlook unlikely risks; imaginative ones emphasize possibility over probability. Even prescient commentators like novelist William Gibson have been blindsided by unexpected events altering their forecasts. “AI 2027” is vivid and speculative, written like sci-fi with detailed charts. It posits a near-future intelligence explosion around mid-2027 driven by “recursive self-improvement” (RSI), where AI systems autonomously conduct AI research, producing smarter progeny in accelerating feedback loops that outpace human oversight. This could trigger geopolitical conflicts, e. g. , China building massive datacenters in Taiwan to control AI. The scenario’s specific details enhance engagement but are flexible; the key message is the likely onset of intelligence explosion and ensuing power struggles. RSI is hypothetical and risky; AI firms recognize its dangers yet plan to pursue it to automate their own work. Whether RSI works depends on technological factors like scaling, which may face limits. If RSI succeeds, superintelligence surpassing human intellect could emerge—an unlikely coincidence if progress stops just above human levels. Consequences might include militarized arms races, AI manipulating or eliminating humanity, or benevolent superintelligent AI solving alignment problems. Uncertainty prevails due to AI’s evolving nature, proprietary research secrecy, and speculation. “AI 2027” confidently narrates a technological and human failure scenario where companies pursue RSI despite lacking interpretability and control mechanisms. Kokotajlo contends these are deliberate decisions fueled by competition and curiosity, despite known risks, making the companies themselves misaligned actors. In contrast, Kapoor and Narayanan’s “AI as Normal Technology, ” reflecting an East Coast, conservative outlook grounded in historical knowledge, doubt rapid intelligence explosions. They cite “speed limits” imposed by hardware costs, data scarcity, and general technological adoption patterns that slow revolutionary impact, providing ample time for regulatory and safety responses. For them, intelligence is less critical than power—the ability to effect environmental change—and even highly capable technologies often diffuse slowly. They illustrate this with driverless cars’ limited deployment and Moderna’s COVID-19 vaccine development: although vaccine design was rapid, rollout took a year due to biological and institutional realities. AI boosting innovation won’t eliminate societal, regulatory, or physical constraints on implementation. Moreover, Narayanan emphasizes that AI’s focus on intelligence underestimates domain-specific expertise and existing safety systems in engineering—fail-safes, redundancies, formal verification—already ensure machine safety alongside humans. The technological world is well-regulated, and AI must integrate into this structure slowly. They exclude military AI, which involves distinct, classified dynamics, warning that AI militarization, a central fear of “AI 2027, ” requires focused monitoring. They advise proactive governance: regulators and organizations should not await perfect alignment to arise but begin tracking AI’s real-world use, risks, and failures, and strengthen rules and resilience accordingly. Deep worldview divides stem from reactive intellectual dynamics fueled by AI’s provocations, producing entrenched camps and feedback loops. Yet, a unified perspective is conceivable by imagining a “cognitive factory”: a workspace where humans, equipped with safety gear, operate machines designed for productivity and safety under strict quality control, gradual integration of innovations, and clear accountability. Though AI enables automation of some thinking, human oversight and responsibility remain paramount. As AI grows, it does not diminish human agency; rather, it intensifies the need for accountability since augmented individuals bear greater responsibility. Stepping away from control is a choice, underscoring that at the end of the day, humans remain ultimately in charge. ♦