AI as Normal Technology

"To view AI as normal is not to understate its impact"

Read the opening paragraphs twice; they set the terms for everything after. The thesis is three things at once — a description of AI today, a prediction about the near future, and a prescription for how to treat it. Critics often attack one and assume they have hit all three. Notice what the frame rejects: technological determinism — the belief that a technology arrives carrying a fixed trajectory society merely absorbs. Here AI is not the agent determining its future. People and institutions are.

"What eventually allowed gains to be realized was redesigning the entire layout of factories around the logic of production lines."

Most timeline arguments measure one clock — how fast methods improve — and treat economic transformation as following automatically. The load-bearing move here is that innovation and diffusion run on their own clocks, and impact arrives only at the far end. The evidence is striking: across the fifty-odd consequential uses of prediction the authors surveyed (criminal risk, insurance, child welfare), deployed systems mostly run decades-old regression. Electrification is the historical spine of the argument, so give it real attention. Motors were available roughly forty years before they showed up in productivity statistics, and owners who bolted them onto steam-era layouts genuinely got little benefit. Ask of any transformation claim: who redesigns the factory, and how long does that take?

"This "capability-reliability gap" shows up over and over."

Carry this distinction out even if you reject everything else in the essay. A system can be capable of a task — able to do it well much of the time — and still be unusable, because deployment needs it to be reliable: right nearly always, and wrong in predictable, recoverable ways. Self-driving cars took two decades of the same improvement loop AlphaZero ran in hours, because every turn had to be survivable in the real world. This is named as the main barrier to useful agents, and it explains a pattern you have probably noticed: impressive demos, sparse production deployments.

"The easier a task is to measure via benchmarks, the less likely it is to represent the kind of complex, contextual work"

The technical heart of the skepticism is a measurement concept: construct validity, whether a test measures what it claims to. A model scoring in the top ten percent on the bar exam tells you about retrieving and applying memorized law; it says little about practicing law, because the parts of lawyering that would matter most if automated — judgment, strategy, filings with no single right answer — resist standardized scoring. Note the general form, sharper than "benchmarks are imperfect": a systematic inverse relation between how measurable a task is and how much it resembles professional work. If that holds, benchmark progress overstates real-world impact by construction, not by accident.

"intelligence is not the property at stake for analyzing AI's impacts. Rather, what is at stake is power"

Part II opens by dissolving a word rather than arguing about it. Treating "intelligence" as a single dimension along which species and machines can be ranked is, the authors hold, not well defined — and not what determines impact anyway. Power is: the ability to modify one's environment. Modern humans are not biologically different from ancestral ones; accumulated technology is what made us powerful. Follow the two-arrow diagram carefully. Capability leads to power, which leads to loss of control; the superintelligence view intervenes at the second arrow through alignment, this essay at the first through the ordinary machinery of deployment. The payoff is a world where a growing share of human work simply is controlling AI — auditing, monitoring, and specifying what the system should do.

"Attempting to make an AI model that cannot be misused is like trying to make a computer that cannot be used for bad things."

Part III's misuse argument is the one most likely to change how you think. Whether a capability is harmful depends on context that lives in the attacker's plan, not in the model: every step of a phishing campaign is individually innocuous. If that is right, model-level refusal training is structurally limited — either too restrictive to be useful or too porous to matter — and the real defenses sit downstream, in email filtering, access controls, procurement screening, and security practices hardened against human attackers for decades. The misalignment section applies the same lens, classing catastrophic misalignment as speculative in a precise sense: uncertainty of the kind research could actually resolve.

"we believe that nonproliferation-based safety measures decrease resilience and thus worsen AI risks in the long run"

Part IV is where the worldviews become policy and actively conflict. Defending against a possible superintelligence favors concentration — fewer actors, central control, export restrictions, no open weights. Defending against normal-technology harms favors the opposite: competition, decentralization, many regulators rather than one. The authors sort interventions by robustness — back what helps whichever future arrives, and be wary of measures that only make sense if you are already certain. Whether or not you accept the conclusion, that is a useful test to carry into any AI policy argument.